remote sensing

Article
Estimating Rangeland Forage Production Using
Remote Sensing Data from a Small Unmanned Aerial
System (sUAS) and PlanetScope Satellite
Han Liu 1, *, Randy A. Dahlgren 1 , Royce E. Larsen 2 , Scott M. Devine 1 , Leslie M. Roche 3 ,
Anthony T. O’ Geen 1 , Andy J.Y. Wong 1 , Sarah Covello 2 and Yufang Jin 1
1 Department of Land, Air and Water Resources, University of California, Davis, CA 95616-8627, USA;
[email protected] (R.A.D.); [email protected] (S.M.D.); [email protected] (A.T.G.);
[email protected] (A.J.Y.W.); [email protected] (Y.J.)
2 University of California Cooperative Extension, San Luis Obispo, CA 90001, USA;
[email protected] (R.E.L.); [email protected] (S.C.)
3 Department of Plant Science, University of California, Davis, CA 95616-8627, USA; [email protected]
* Correspondence: [email protected]




Received: 13 February 2019; Accepted: 6 March 2019; Published: 12 March 2019


Abstract: Rangelands cover ~23 million hectares and support a $3.4 billion annual cattle industry
in California. Large variations in forage production from year to year and across the landscape
make grazing management difficult. We here developed optimized methods to map high-resolution
forage production using multispectral remote sensing imagery. We conducted monthly flights
using a Small Unmanned Aerial System (sUAS) in 2017 and 2018 over a 10-ha deferred grazing
rangeland. Daily maps of NDVI at 30-cm resolution were first derived by fusing monthly 30-cm
sUAS imagery and more frequent 3-m PlanetScope satellite observations. We estimated aboveground
net primary production as a product of absorbed photosynthetically active radiation (APAR) derived
from NDVI and light use efficiency (LUE), optimized as a function of topography and climate
stressors. The estimated forage production agreed well with field measurements having a R2 of 0.80
and RMSE of 542 kg/ha. Cumulative NDVI and APAR were less correlated with measured biomass
(R2 = 0.68). Daily forage production maps captured similar seasonal and spatial patterns compared
to field-based biomass measurements. Our study demonstrated the utility of aerial and satellite
remote sensing technology in supporting adaptive rangeland management, especially during an era
of climatic extremes, by providing spatially explicit and near-real-time forage production estimates.

Keywords: Drone; MicaSense RedEdge; Commercial satellite; Light use efficiency; Data fusion;
Rangeland; Aboveground biomass; Environmental stress

1. Introduction
Rangelands are a key global resource, both in terms of extent and ecological and economic impact.
In California, over 60% of the land area is rangeland, which provides a wide range of ecological
services, including forage production for livestock and wildlife, water quality protection, biodiversity,
recreation, and wildlife habitat [1]. More than half of California rangelands are grazed [2], supporting
a $3.4-billion per year livestock industry [3]. Rangeland forage production is characteristically variable
and depends on a multitude of drivers—most notably climate, soils and topography, which together
regulate plant-available water [4]. Most California rangelands are rain fed, and thus highly vulnerable
to irregular precipitation patterns [5]. For example, large year-to-year variations in forage production
(from 33 to 11503 kg/ha) were observed in a 14-year record from the California Central Coast region [6].
Warming, drought, and increasing climate variability [7,8] are predicted to become more prevalent

Remote Sens. 2019, 11, 595; doi:10.3390/rs11050595 www.mdpi.com/journal/remotesensing

Remote Sens. 2019, 11, 595 2 of 22

in the future, and therefore, raise the critical need for cost-effective; timely; and robust tools for
monitoring, predicting and optimizing use of forage while maintaining proper stewardship of the land.
Several studies have examined factors controlling forage production of California’s annual
grass-dominated rangelands. Temperature was found to be an important regulator for timing of
four distinct forage growth stages: germination season, winter growth, rapid spring growth, and peak
forage production [4]. Amount and timing of precipitation are closely related to forage production
with highest yields occurring when sufficient precipitation is received in April (to support rapid spring
growth) [9]. Sensitivity of forage production to weather also varies spatially with soil characteristics
and topography. In Mediterranean climates, such as California, annual systems are highly dependent
on moisture stored in the upper 25~30 cm of soil [10]. Medium texture soils (e.g., loams, clay loams,
and silt loams) have higher plant-available water-holding capacity, which can provide a buffer for
plants when precipitation is sparsely spaced [11]. Topography affects forage production through two
pathways: influencing soil properties such as soil depth, fertility and temperature, and controlling
photosynthetically active radiation (PAR) [12]. Thus, spatial and temporal heterogeneity of forage
production can only be fully explained by coupling these key drivers: temperature, amount and timing
of precipitation, soil characteristics and topography.
Traditional methods for measuring forage biomass rely on time-consuming hand clipping and
drying of quadrats randomly selected across the landscape. Advancements in technology could
facilitate modeling as an alternative for forage production monitoring and prediction. Empirical
statistical models relating observed production with key environmental determinants have been
developed from large datasets of field measurements [9,13–15]. Growing degree days (GDD) and
precipitation were used as inputs to predict California’s rangeland production dating back to the
1970s [9,14]. These empirical models usually do not consider landscape characteristics, and thus,
performance is limited when extrapolating to larger areas with heterogeneity in climate, soils
and topography.
Mechanistic models have also been developed to simulate complex biophysical processes in
rangelands [16]. For example, the Simulation of Production and Utilization of Rangelands model [17]
is able to predict forage production in different regions reasonably well (R2 = 0.3~0.9, depending on
the species and location) [18–20]. However, the model accuracy is often limited by the assumptions
made to simplify the processes and the uncertainties in the model parameters tuned with field
measurements. They also require a large number of site-specific environmental variables, e.g.,
climate, and soil hydrological properties as inputs. Therefore, it is logistically challenging for land
managers to implement simulation models because spatially-explicit data are difficult to obtain for
large heterogeneous areas.
On the other hand, simplified eco-physiological models provide a robust approach over different
regions and time, by converting photosynthetically active radiative (PAR) energy into biomass
based on light use efficiency (LUE). Forage production is estimated as a product of absorbed PAR
(APAR) and LUE [21]. APAR can be easily estimated from remote sensing data [22–26] and the
LUE calibrated to field measurements; therefore it serves as a good candidate for mapping forage
production [22,23,26–32]. The main challenge is parameterizing LUE as a function of environmental
variables, since the proportion of APAR converted to biomass is down-regulated under various
environmental stressors, such as soil moisture deficits and extreme temperatures. For example, LUE is
derived as a function of potential LUE and the combined effects of low air temperature, high water
vapor deficits, and soil moisture in the Global Production Efficiency Model (GLO-PEM) [32]. In the
MODIS Gross Primary Production algorithm, LUE is calculated as the product of biome-specific
maximum LUE and several down-regulators controlled by air temperature and vapor deficit [26].
In two Mediterranean grassland sites, Eddy Covariance data were used to calibrate the dominant
LUE stress terms from air temperature and soil moisture, where soil moisture was identified as the
dominant factor controlling LUE [31]. While different studies apply different forms of LUE calculation,
Remote Sens. 2019, 11, 595 3 of 22

this critical metric needs to be calibrated rigorously in different ecosystems and with different input
datasets, because of its high model sensitivity.
Remotely sensed imagery from satellites provides a relatively reliable source for obtaining
spatiotemporal inputs for LUE-based models. Landsat remote sensing data, with a 30-m resolution
and a 16-day revisiting frequency, have been widely used for many vegetation monitoring studies
worldwide [23,33–36]. However, coincidence of rainfall and growing season in California’s
Mediterranean climate greatly limits the availability of cloud-free Landsat images during critical
plant growth stages from November to April. Moreover, the spatial resolution of traditional satellite
remote sensing generally ranges from a kilometer to 10s of meters, whereas most field biomass
measurements are at a scale of 10s of centimeters. This scale mismatch poses challenges for developing
and validating the LUE optimization.
Recent advances in small unmanned aerial system (sUAS) technology and image processing
make it possible to overcome many challenges involved in quantifying forage production across the
landscape. Compared to satellites, sUAS can be deployed quickly, repeatedly, and flexibly. A 20-minute
flight is sufficient to map a 10-ha area, and with advanced software, data processing can be completed
in a few hours. Cameras designed for sUAS vegetation monitoring capture spectral information in
visible and near infrared (NIR) bands, from which critical information on plant vigor and growth
can be extracted. Users can easily customize the spatiotemporal resolution by adjusting the flight
height and frequency. High-resolution digital surface models (DSMs) generated from sUAS data allow
accounting for topo-edaphic conditions in forage production modeling and related analysis. A few
pilot studies have explored the applications of sUAS technology in monitoring agricultural production
for soybean [37], rice and wheat [38,39], barley [40,41], and mango [42]. However, there is currently
limited research concerning the efficacy of sUASs for rangeland forage production quantification
and modeling.
The primary aim of this paper was to develop LUE-based models to map forage production at
very high spatial (30 cm) and temporal (daily) resolutions using multispectral data collected from sUAS,
augmented with more frequent satellite data from PlanetScope satellites at 3-m resolution [43].
Specifically, our objectives included (1) imaging a Mediterranean-type semi-arid annual rangeland
using sUAS and fusing the sUAS data with PlanetScope data to obtain daily sUAS resolution images;
(2) investigating the connection between ground-based plant biophysical measurements and remote
sensing vegetation indices, (3) building and evaluating forage production models to map daily
rangeland production at 30-cm resolution, and (4) analyzing the predicted spatial and temporal
patterns of estimated forage production to explore the relationship between forage production and its
environmental and climatic drivers.

2. Materials and Methods

2.1. Study Site

Our study focused on a privately-owned, annual grassland (no trees or shrubs), located 56 km
east of the coast in San Luis Obispo County, California (35.5N, 120.3W) (Figure 1), which has been
grazed by beef cattle for more than 15 years. A 10-ha parcel was fenced in November 2016 to exclude
grazing during the growing season. Following peak forage biomass growth, 50 cow-calf pairs grazed
the fenced area for 30 days in the summer of 2017 to prepare the site for a repeated study during the
2018 growing season.
Soils were dominated by the loamy Balcom-Nacimiento complex [6]. The climate is Mediterranean
with hot, dry summers and mild, moderately wet winters. Mean annual rainfall during 2001–2018 was
213 mm, mostly occurring from December through March. Annual rainfall at the site was 176, 139, 57,
and 130 mm during the 2012–2016 drought years, respectively, and 287 and 123 mm during the study
years of 2017 and 2018, respectively. Annual grasses and forbs at the site typically germinate in late fall
with the onset of the rainy season (~November) and grow rapidly (March–April) to reach peak biomass
 Remote Sens. 2019, 11, 595 4 of 22

in late spring (April-May), depending on the timing and length of the rainy season. Historical forage
production for the site shows a large fluctuation, varying from 132 to 4204 kg/ha (mean = 1866 kg/ha)
during
Remote Sens.2001–2014 [6].PEER
2018, 10, x FOR TheREVIEW
study site was selected to include complex topography, including4 valley,
of 22
hilltop and concave/convex hillslopes with slopes ranging from 10 to 20 degrees (Figure 1). Overall,
147 from
45%10oftothe
20total
degreesarea(Figure 1). Overall,
was south 45% west
facing, 29% of thefacing,
total area
24%was south
north facing,
facing, and29% westfacing.
2% east facing, 24%
148 north facing, and 2% east facing.

149
150 Figure 1. 1.
Figure Topography
Topographyof the
of study area (35.5N,
the study 120.3W)
area (35.5N, on Central
120.3W) Coast, California,
on Central from sUAS-
Coast, California, from
151 derived digital surface
sUAS-derived digitalmodel.
surfaceThe study
model. area
The features
study area large topographic
features variationvariation
large topographic includingincluding
a valley, a
152 a hilltop,
valley, aand manyand
hilltop, concave
manyand convex
concave slopes.
and The
convex soil sensors
slopes. (insensors
The soil black) and additional
(in black) 20 biomass20
and additional
153 clipping
biomassplots (in red)
clipping were
plots (in distributed in multipleintopographic
red) were distributed positions. positions.
multiple topographic

154 2.2.
2.2. Field
Field Measurements
Measurements
155 InIn November
November 2016,
2016, 1616 pairs
pairs ofof ECH2O
ECH2O 5TM5TM sensors
sensors were
were installed
installed toto measure
measure soil
soil moisture
moisture and
and
156 temperature
temperature acrossthe
across thestudy
studyarea
area(Figure
(Figure1).1).Locations
Locations were
were selected
selected based
based onona adigital
digitalelevation
elevation
157 model to capture the spatial variation of topography and soil conditions.
model to capture the spatial variation of topography and soil conditions. At each location, duplicate At each location, duplicate
158 sensors
sensors were
were deployed
deployed inin the
the shallow
shallow rooting
rooting zone
zone atat a 7-cm
a 7-cm depth
depth (most
(most roots
roots occurred
occurred ininthethe upper
upper
159 15 cm soil layer). Data-loggers recorded sensor readings every 15 minutes.
15 cm soil layer). Data-loggers recorded sensor readings every 15 minutes. Three tipping bucket rain Three tipping bucket rain
160 gauges recorded rainfall, and three time-lapse cameras followed forage
gauges recorded rainfall, and three time-lapse cameras followed forage germination and growth at germination and growth at the
161 hilltop,
the hilltop, south-facing
south-facing slope,
slope,and
and north-facing
north-facing slope
slopepositions
positions (Figure
(Figure 1).1).
162 We measured forage production immediately after each of the 1010sUAS
We measured forage production immediately after each of the sUAS flightmissions
flight missions (except
(except
163 11/11/2016). Two replicate 30 cm ×
for 11/11/2016). Two replicate 30 cm × 30 cm quadrats at opposite cardinal angles were selected atat
for 30 cm quadrats at opposite cardinal angles were selected aa
164 1.5-mradius
1.5-m radiusofofthethe 16
16 soil
soil sensors,
sensors, resulting
resultinginintwo twosample
sample points
points perper
plot [44].
plot All All
[44]. aboveground
aboveground plant
◦ C for 48 hours before weighing to record dry biomass values.
165 biomass was harvested and dried at 60
plant biomass was harvested and dried at 60 ˚C for 48 hours before weighing to record dry biomass
166 To further
values. capturecapture
To further the spatial
the variation in topography
spatial variation and forage
in topography and production, we classified
forage production, we the site into
classified
167 five categories based on clustering of topographical features including aspect,
the site into five categories based on clustering of topographical features including aspect, slope, flow slope, flow accumulation,
168 and elevation,and
accumulation, using Iterativeusing
elevation, Self-Organizing Data Analysis Data
Iterative Self-Organizing Technique classification
Analysis Technique [45]. We selected
classification
169 an additional
[45]. We selected20an plots for biomass
additional clipping
20 plots to ensure
for biomass that each
clipping topographical
to ensure that eachcluster had at least
topographical eight
cluster
170 had at least eight plots. We collected 32~52 biomass clip-plot samples on each sUAS observation date a
plots. We collected 32~52 biomass clip-plot samples on each sUAS observation date and retained
171 total
and of 330 a(200
retained and
total of 130
330 in 2017
(200 andand1302018 growing
in 2017 seasons)
and 2018 growingbiomass measurements
seasons) after removing
biomass measurements
172 outliers
after (e.g., outliers
removing samples(e.g.,
affected by strong
samples affected rodent activity).
by strong rodentFor the cumulative
activity). NDVI & APAR
For the cumulative NDVI and &
173 biomass analysis and LUE modeling, we further averaged the plots
APAR and biomass analysis and LUE modeling, we further averaged the plots that had two samples that had two samples near soil
174 sensors,
near resulting
soil sensors, 220 data220
resulting points
datato worktowith.
points work with.

2.3. sUAS Flights and Image Preprocessing

175 2.3. sUAS Flights and Image Preprocessing
We integrated a MicaSense RedEdge camera with a 3DR Solo quadcopter for monthly aerial
176 We integrated a MicaSense RedEdge camera with a 3DR Solo quadcopter for monthly aerial
flights over the study area during the 2017 and 2018 growing seasons (November–April). MicaSense
177 flights over the study area during the 2017 and 2018 growing seasons (November–April). MicaSense
RedEdge is a multispectral camera with five spectral bands including blue, green, red, red-edge,
178 RedEdge is a multispectral camera with five spectral bands including blue, green, red, red-edge, and
and NIR. A sun irradiance sensor was included to measure band-specific incoming solar irradiance for
179 NIR. A sun irradiance sensor was included to measure band-specific incoming solar irradiance for
180 radiometric correction. Flights were scheduled with the closest overpass dates of PlanetScope
181 satellites when weather permitted it. A total of 10 missions were performed: six from November 11,
182 2016 to April 30, 2017, and four from January 18, 2018 to April 14, 2018.
183 We followed the same flight plan with a side- and front-overlap of 85% for all missions. Each
184 mission was flown at 91-m (350-ft) above ground level (from the launching point) to acquire imagery
Remote Sens. 2019, 11, 595 5 of 22

radiometric correction. Flights were scheduled with the closest overpass dates of PlanetScope satellites
when weather permitted it. A total of 10 missions were performed: six from 11 November 2016 to
30 April 2017, and four from 18 January 2018 to 14 April 2018.
We followed the same flight plan with a side- and front-overlap of 85% for all missions. Each mission
was flown at 91-m (350-ft) above ground level (from the launching point) to acquire imagery at a 7.5 ~
7.9 cm resolution and cover the whole study area in less than 30 minutes due to the payload constraint.
Flying speed was set at 7 m/s along a fixed direction parallel to the east/west site boundary (Figure S1).
Flights were conducted around solar noon to minimize the impacts of sun-angle variation throughout
the season. Images of a calibrated white reflectance panel were recorded before and after each flight
to calibrate raw pixel values to absolute reflectance values. To ensure the accuracy of the sUAS image
geo-registration, we established eight permanent landmarks as ground control points (GCPs) evenly
distributed at the corners (and center) of the study area. Coordinates of GCPs, forage clip-plot samples
and soil sensors were recorded using a Trimble Geo 7x Series Handheld Data Collector.
Raw sUAS images were stitched and processed in Pix4dMapper Pro to generate orthomosaic
maps of surface reflectance and DSM. The software converts the raw data in digital number (DN) to
surface reflectance using images of the calibration panel, incoming sunlight irradiance, and camera
parameters (e.g., ISO, aperture, shutter speed, and vignetting) recorded in the EXIF metadata [46].
DSMs were based on a dense point cloud generated from tie points that were automatically identified
by the software. Mean average error (MAE) between GPS measurements and derived elevation was
3.5 cm across the study area.
A two-step geo-registration was performed on the time series of sUAS images to ensure spatial
alignment of all aerial images. Images were first geo-registered in Pix4dMapper Pro using the eight
GCPs to minimize the spatial mismatch when relating images to ground measurements. A mean
Root Mean Square error (RMSE) of less than 3 cm was achieved. Reliability of this GCP-based
geo-registration is highly dependent on the accuracy of the GPS. Since we did not upgrade our GPS
to centimeter accuracy until the summer of 2017, we then used the Image to Image Registration
function in ENVI 5.3 (using the 4/6/2017 image as the base image and second-order cubic) to align the
images acquired in the 2017 growing season on different dates. In the Image to Images registration,
we selected GCPs visually for each pair of images. The resulting RMSE was maintained at ~3 cm,
less than 0.3 pixels at the original 8-cm resolution for all pairs of images. The geo-registered 8-cm
resolution surface reflectance maps were then aggregated to 30-cm resolution to match the plant
biomass field measurement plots and used for further analysis, which resulted in a <10% geolocation
displacement in the final 30-cm maps.
Standard reflectance calibration assumes similar incoming solar radiation for all pixels. However,
the complex topography of the study site led to significant variation in illumination conditions among
pixels at different topographic positions. We applied the C model [47] to correct pixel reflectance based
on its illumination condition (IC) for each spectral band (supplemental online material, SOM). The C
model has been applied for correcting Landsat satellite data [47–50], but its feasibility and accuracy in
correcting sUAS has not been fully evaluated. We used our 30-cm resolution DSM derived from the
sUAS data to perform the illumination correction. The illumination-corrected red and NIR reflectance
values were then used to calculate sUAS NDVI.

2.4. High Resolution Satellite Imagery and Data Fusion

Relatively high temporal frequency of remote sensing imagery is needed to capture the rapid
growth cycle of annual plants. We downloaded a total of 174 cloud-free PlanetScope (PS) orthorectified
scenes (level 3A) from Planet at 3.125-m spatial resolution for the 2017 and 2018 growing seasons.
With a constellation of around 120 CubeSats, PS aims to provide daily images of three visible bands
and one NIR band for any place on Earth. We converted the DN to the Top of Atmosphere (TOA)
reflectance, using the TOA reflectance coefficients embedded in the metadata XML for each individual
scene. The NDVI was then calculated with red and NIR reflectance.
Remote Sens. 2019, 11, 595 6 of 22

To compensate for the limited temporal acquisition of sUAS data, we combined PS satellite data
with monthly sUAS data to obtain a complete time series of daily NDVI at the 30-cm sUAS resolution.
Our goal was to predict a sUAS-resolution NDVI map U(x,y,tp ), for any particular date (tp ), using
(1) two pairs of coincident sUAS and PS images during the nearest two sUAS flight dates on tb1
and tb2 and (2) a PS-based image on the prediction date tp . We followed the basic concept of the
Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) [51] to take advantage of the
complementary spatial and temporal information from PS and sUAS. We first took advantage of the
PS’s high temporal frequency and linearly interpolated temporally the available cloud-free PS NDVI
Remote Sens. 2018, 10, x FOR PEER REVIEW 6 of 22
images to obtain a continuous daily time series. To facilitate the pixel-based process, daily PS data
were reprojected
236 and bi-linearly resampled, denoted as P(x,y,t) to match the
complementary spatial and temporal information from PS and sUAS. We first took advantage of the resolution, projection,
237 PS’s high temporal frequency and linearly interpolated temporally the available
and extent of the sUAS data. Since the study area has a relatively homogeneous landcover, we assumed cloud-free PS NDVI
238 images to obtain a continuous daily time series. To facilitate the pixel-based process, daily PS data
that the temporal
239 werechange
reprojected from the daily
and bi-linearly PS NDVI
resampled, images
denoted wasto similar
as P(x,y,t) match the at the 30-cm
resolution, resolution:
projection,
240and extent of the sUAS data. Since the study area has a relatively homogeneous landcover, we
241
(
assumed that the temporal change from
the daily PS NDVI images was similar at the 30-cm
242resolution:
U x, y, t p = U ( x, y, tb ) + ∆NDV I
(a)
y, t p = P( x, y, tb ) + ∆NDV I


𝑈(𝑥, 𝑦, 𝑡 ) = 𝑈(𝑥, 𝑦, 𝑡P ) x,
+ 𝛥𝑁𝐷𝑉𝐼
243 𝑃(𝑥, 𝑦, 𝑡 ) = 𝑃(𝑥, 𝑦, 𝑡 ) + 𝛥𝑁𝐷𝑉𝐼
(a)
244 The sUAS-resolution NDVI for a given location (x,y) and day (tp) was then estimated as a
The sUAS-resolution
245 weighted sum of NDVI
the twofor a given
closest pairs oflocation
sUAS base (x,y) and day
NDVI images (tp )resolution
at 30-cm was then estimated
(Figure 2), each as a weighted
sum 246
of the two closest pairs of sUAS base NDVI images at 30-cm resolution (Figure 2), each adjusted
adjusted for temporal change derived from the PS NDVI time series:

for temporal change derived from

U x, y, 𝑡 the
=𝑤PS𝑈(x,
NDVI y, 𝑡 )time series:
+ 𝑃 x, y, 𝑡 − 𝑃(x, y, 𝑡 )
(1)
+ 𝑤 𝑈(x, y, 𝑡 ) + 𝑃 x, y, 𝑡 − 𝑃(x, y, 𝑡 )




247 whereUtb1x, y, t = w1 for
and tb2pare the dates
U (the
x, y,
baseb1 )+P
t image pair, and w
− wP2 (are
x, y, t p 1 and x, y, tb1 )
the temporal weights


(1)
248 representing the contribution of +w each
2 U (
basex, y,
imaget b2 )
pair+ toPthex, y, t
estimated
p − P
NDVI,(x, y, ton
based b2 )a linear
249 function of the correlations between the base PS NDVI images on tb1 and tb2 and the PS NDVI image
250 on tp (SOM).
where tb1251and tb2 are the dates for the base image pair, and w1 and w2 method.
We tested the performance of our data fusion method using a leave-one-out
are theWithin
temporal weights
252 theeach
representing contribution
iteration, we of
lefteach base
out one sUASimage
NDVI pair
imageto thepredicted
and estimated NDVI,
that image based
using on a linear function
the nearest
253 neighboring image pairs and the PS NDVI image taken on the same day. The accuracy of the method
of the correlations between the base PS NDVI images on tb1 and tb2 and the PS NDVI image on tp (SOM).
254 was evaluated by comparing the predicted sUAS NDVI image to the observed sUAS NDVI image.

255
2562. Flowchart
Figure Figure 2.for
Flowchart for the
the data data fusion
fusion method to
method to combine
combine the monthly sUAS imagery
the monthly sUASand imagery
more and more
257 frequent PlanetScope imagery. (Step 1) PS imageries are first interpolated daily and preprocessed to
frequent PlanetScope imagery. (Step 1) PS imageries are first interpolated daily and preprocessed to the
same projection and resolution as the sUAS imageries. (Step 2) The preprocessed daily PS imageries
are used to calculated weights based on correlations between base imageries (on tb1 and tb2 ) and the
imagery on tp . (Step 3). Finally, the predicted NDVI values are computed from the weighted NDVI
values from both PS and sUAS (Equation (1)).
Remote Sens. 2019, 11, 595 7 of 22

We tested the performance of our data fusion method using a leave-one-out method. Within each
iteration, we left out one sUAS NDVI image and predicted that image using the nearest neighboring
image pairs and the PS NDVI image taken on the same day. The accuracy of the method was evaluated
by comparing the predicted sUAS NDVI image to the observed sUAS NDVI image.

2.5. Forage Production Estimation Methods and Assessment

2.5.1. Empirical Statistical Analysis

We first examined the relationship between the measured biomass and two types of remote
sensing metrics derived from the fused data, namely NDVI and APAR. NDVI was used as a proxy
for net primary production and biomass [52–55]. Specifically, we used univariate linear regression
(in R) to quantify the spatial variance in biomass across various plots explained by the corresponding
coincident NDVI and the spatial and temporal variance explained when pooling data from all sites and
dates together. When analyzing the coincident NDVI–biomass relationship, we did not average the
paired samples taken close to the soil sensors, considering the limited number of samples on a single
observation day. We investigated the relationship between biomass and cumulative NDVI integrated
from the beginning of the growing season to the date of observation. Critical phenological days,
such as the beginning (germination date) and end of the growing season were identified by fitting a
logistic function (SOM) to the fused NDVI time series [56]. A similar statistical analysis was performed
for cumulative APAR derived from the fused daily data and the California Irrigation Management
Information System (CIMIS) spatial solar radiation.

2.5.2. Light Use Efficiency (LUE) Models

We estimated forage production based on LUE theory, as the cumulative product of APAR and
LUE [22]:
tp
Biomass x, y, t p = ∑ APAR x, y, t p ∗ LUE x, y, t p




(2)
t0

where, tp is the date of prediction, and t0 is the corresponding germination date.

APAR was estimated as the product of fPAR and PAR. We calculated fPAR as a function of
NDVI [57]:

( NDV I − NDV Imin )( f PARmax − f PARmin )

f PAR NDV I = + f PARmin (3)
NDV Imax − NDV Imin

where fPARmax = 0.95, fPARmin = 0.001, and NDVImax and NDVImin are the NDVI values corresponding
to the upper and lower 2% of the NDVI histogram. NDVImin and NDVImax were set to 0.23 and 0.8,
respectively, based on the NDVI distribution of sUAS data.
PAR was calculated as 50% of the daily shortwave incoming solar radiation, available from
the CIMIS Spatial dataset at a 2-km resolution [58], assuming a constant ratio of 0.5 for PAR over
incoming solar radiation [59–64]. The CIMIS Spatial incoming solar radiation showed an R2 of 0.99
when validated with CIMIS measured incoming solar radiation. The 2-km PAR was further sharpened
to 30-cm resolution using the sUAS DSM to account for topography-induced illumination, and thus,
PAR variations. We first generated daily solar radiation maps of the study area using the Area Solar
Radiation function in the Arcpy package in Python, based on the sUAS-derived 30-cm DSM, latitude,
and day of year [65]. Similar maps were derived for a simulated flat surface assuming DSM at the
average elevation of the study area (491 m). For each 30-cm pixel and each day, the ratio of the daily
30-cm solar radiation from the actual DSM over the simulated flat DSM was used to multiply the
corresponding CIMIS spatial solar radiation to derive solar radiation and PAR at 30-cm resolution.
We developed two semi-empirical statistical models to estimate LUE, depending on availability
of the input data. All parameters in LUE models were optimized with field measurements. We first
Remote Sens. 2019, 11, 595 8 of 22

parameterized LUE as a function of topographic variables (Equation (4), Model I), which were derived
from the DSM generated from sUAS data:


LUE x, y, t p = LUE0 ∗ f ( Topo, x, y) (4)

LUE0 is the LUE at the optimal topographic position. LUE0 is estimated by scaling f(Topo,x,y)
from 0 to 1. We used topographic variables to drive an empirical scalar because we did not have the
spatial soil moisture and air temperature data that are traditionally used for LUE down-regulators.
We explored a suite of topographic features including cosine aspect, slope, flow accumulation, IC,
and elevation, and enumerated the possible combinations of these variables in the equation.
We further added time-varying controls of soil moisture (W) and air temperature (T) on LUE,
in addition to significant topographic features (Equation (5), Model II), as shown below:





LUE x, y, t p = LUE0 ∗ g W, x, y, t p ∗ p T, x, y, t p ∗ h( Topo, x, y (5)



where g W, x, y, t p , p T, x, y, t p , and h( Topo, x, y) range from 0 to 1. Forms of the W and T scalar
functions were set to the same as those in Reference [22] but coefficients were optimized using
field-measured biomass. The temperature (T) stress scalar was based on deviation from optimal air
temperature (Topt ):

4
p T, x, y, t p =     (6)
a( Topt − Ta ( x,y,t p ))
1+e ∗ 1 + e a(−Topt +Ta (x,y,t p ))

where a is the coefficient to be calibrated, Ta is the air temperature (◦ C), and (x, y, t p ) denotes the location
and time. Topt is here defined as the average temperature during the month with the highest NDVI.
We used daily mean air temperature from the 2-km CIMIS spatial dataset. The T scalar decreases from
1 to 0 as temperature deviates from Topt , with the rate of decrease represented by a.
The water stress (W) scalar represents the down regulation of light use efficiency under drought
conditions, and was parameterized as a function of soil moisture (M) (Equation (7)):

1
g W, x, y, t p = (7)
1 + e−b( M( x,y,t p )−WP)

where b is the coefficient to be calibrated; M is the derived soil moisture, represented by volumetric
water content (VWC); and WP is the wilting point. W ranges from 0.5 to 1 as soil moisture varies
from the wilting point to field capacity. The wilting point was set to a water content of 0.14 cm3 cm−3
determined by averaging soil moisture content in June during the dry season when grass and forbs
had desiccated and soil moisture was at its water-year minimum [66]. For operational mapping
purposes, we required soil moisture for every pixel. We estimated the daily soil moisture using a
simple bucket model (SOM) following Reference [22], based primarily on precipitation data from
PRISM [67] and potential evapotranspiration from CIMIS Spatial [58]. The topographic control term is
a linear function of elevation and IC for Model II. Elevation and IC were selected because they had the
highest statistical significance for minimizing differences between measured and predicted biomass
among topographic variables.
As a comparison to Model II, we also built a Model III by replacing the air temperature and
derived soil moisture with measured soil temperature and soil moisture in the LUE parameterization.
As for the topographic variable, we selected elevation because it had the highest statistical significance
on minimizing model error.
We optimized the model coefficients using the Stochastic Gradient Descent (SGD) method [68,69].
We used 70% of the data for training and the remaining 30% for validation of the SGD LUE optimization.
Model performance was evaluated by comparing model estimates with measured biomass values in
the validation dataset. We used R-squared (R2 ) and RMSE to quantify the uncertainty of models for
 Remote Sens. 2019, 11, 595 9 of 22

estimating forage production. The LUE scalars and their corresponding coefficients may be site-specific
as theySens.
Remote were specifically
2018, optimized
10, x FOR PEER REVIEWfor our study area. 9 of 22

341 2.6. Forage Production Mapping

We implemented Model and
II toPatterns
generate daily forage production maps for the study area. The
342 mapsWewere
implemented Model II to generate daily RGB
compared against the concurrent forageimages frommaps
production sUASfortothe
determine if prediction
study area. The maps
343 were compared against the concurrent RGB images from sUAS to determine if prediction (aspect
captured a similar spatial pattern. To further study the interaction between topography capturedanda
344 slope) and forage production, we calculated the predicted biomass for topographic zones
similar spatial pattern. To further study the interaction between topography (aspect and slope) and divided by
345 aspectproduction,
forage (north andwe south) and the
calculated slope (flat, moderate,
predicted biomass forand steep) and
topographic performed
zones zonal
divided by statistical
aspect (north
346 analysis.
and south) and slope (flat, moderate, and steep) and performed zonal statistical analysis.

347 3. Results
3. Results

348 3.1.
3.1. Terrain
TerrainCorrection
Correction
349 The
The wavelength-specific
wavelength-specific C C model
model significantly
significantly removed
removed thethetopographic
topographic varying
varying illumination
illumination
350 effect
effectpresent
presentin inthe
theoriginal
originalsUAS
sUAS images
images (Figure
(Figure 3).
3). For
For example,
example, pixels
pixels on
on the
the north-facing
north-facing slopes,
351 when
when displayed
displayed in in true
truecolor
colorcomposite,
composite, appeared
appeared very
very dark
dark due
due tototerrain
terrainshadows,
shadows, especially
especially for
for
352 early
early season
season images
images whenwhen thethe sun
sun angle
angle was
was relatively
relatively low,
low, e.g.,
e.g., 16 January
January 16,2017.
2017.This
Thisshadowing
shadowing
353 effect
effectwas
was minimized
minimizedafter after applying
applying the
the correction.
correction. Raw
Raw reflectance
reflectance was
was significantly
significantly correlated
correlated with
with
354 IC
IC (p < 0.01ininmost
(p<0.01 most cases),
cases), but
but the
the correlation
correlation was minimized and not significantsignificant after
after correction
correction
355 (Figure
(Figure4).
4).For
Forexample,
example,RR 2 2was reduced from 0.33 (p < 0.01) to 0.12 (p = 0.02) in the
was reduced from 0.33 (p<0.01) to 0.12 (p=0.02) in the red and from from 0.80
0.80
356 (p < 0.01)to
(p<0.01) to 0.01
0.01 (p=0.5)
(p = 0.5)ininthe
theNIR
NIRfor
forJanuary
January2017
2017imagery
imagery (Figure
(Figure 4). The
The topographic illumination
illumination
357 effects
effectswere
werefound
foundto tobebemore
morepronounced
pronouncedin inthe
theNIR
NIRbandbandand
andearly
earlyininthe
thegrowing
growingseason.
season.

358
Figure 3. Original (a,c) and terrain corrected (b,d) example images, in true color composite, acquired
359 from the3.sUAS
Figure on (a,b)
Original 17 January
(a,c) and terrain 2017 and (c,d)
corrected (b,d)17 March images,
example 2017. Thein C model
true colorremoved most
composite, of the
acquired
360 topographic
from the sUASillumination effect (b,d)
on (a,b) January as presented
17, 2017 and (c,d) in the original
March 17, 2017.images
The C (a,c).
model The shadowmost
removed (pixels in
of the
361 very dark color) on the north-facing slopes is removed in the corrected images.
topographic illumination effect (b,d) as presented in the original images (a,c). The shadow (pixels in
362 very dark color) on the north-facing slopes is removed in the corrected images.
 Remote Sens.
Remote Sens. 2019,
2018, 11,
10, 595
x FOR PEER REVIEW 10
10 of 22
22

363 Figure 4. Scatterplots of red (top panels) and NIR reflectance (bottom panels) vs. the illumination
condition (IC) for 200 pixels with biomass clipping as shown in Figure 1. Raw reflectances are shown
364 Figure 4. Scatterplots of red (top panels) and NIR reflectance (bottom panels) vs. the illumination
in left column (a,c), and corrected reflectances in right column (b,d).
365 condition (IC) for 200 pixels with biomass clipping as shown in Figure 1. Raw reflectances are shown
366 in leftand
3.2. sUAS column (a,c), andData
PlanetScope corrected
Fusionreflectances in right column (b,d).

367 We combined
3.2. sUAS all available
and PlanetScope 30-cm sUAS NDVI images and satellite NDVI images from the two
Data Fusion
growing seasons to generate daily 30-cm NDVI time series from 11 November 2016 to 30 April 2017
368 We combined
and from 1 Januaryall available
to 14 30-cmThe
April 2018. sUAS NDVI
fused dailyimages
NDVIand satellite
curve had a NDVI
similarimages from to
magnitude thethose
two
369 growing seasons to generate daily 30-cm NDVI time series from November 11, 2016 to
from sUAS imagery, and a quantitative comparison showed that the predicted 30-cm NDVI agreed wellApril 30, 2017
370 and from
with JanuarysUAS
the original 1 to April
NDVI14, 2018. (R
images The
2 =fused dailyRMSE
0.87 and NDVI=curve had a similar
0.06) (SOM, Sectionmagnitude to those
2). The fused time
371 from sUAS imagery, and a quantitative comparison showed that the predicted 30-cm NDVI agreed
series followed similar temporal patterns with those from PS data. For a selected plant pixel within a
372 well with the original sUAS NDVI images (R2 = 0.87 and RMSE = 0.06) (SOM, section 2). The fused
relatively homogeneous patch, Figure 5 shows three sets of NDVI values from PS and sUAS sensors at
373 time series followed similar temporal patterns with those from PS data. For a selected plant pixel
30-cm resolution and data fusion. The sUAS NDVI values were typically higher than those from the
374 within a relatively homogeneous patch, Figure 5 shows three sets of NDVI values from PS and sUAS
concurrent PS imagery, e.g., 0.65 vs. 0.46 on 6 April 2017. This difference was most likely due to lower
375 sensors at 30-cm resolution and data fusion. The sUAS NDVI values were typically higher than those
atmospheric effects on surface reflectance from the low-altitude sUAS flights. The fused time series
376 from the concurrent PS imagery, e.g., 0.65 vs. 0.46 on April 6, 2017. This difference was most likely
was able to conserve the sUAS NDVI values, e.g., 0.65 on the same day. Due to its higher temporal
377 due to lower atmospheric effects on surface reflectance from the low-altitude sUAS flights. The fused
frequency, the PS NDVI time series captured more detailed temporal dynamics of plant growth than
378 time series was able to conserve the sUAS NDVI values, e.g., 0.65 on the same day. Due to its higher
the less frequent sUAS snapshots. Both fused and original PS NDVI time series in 2018 showed a
379 temporal frequency, the PS NDVI time series captured more detailed temporal dynamics of plant
unique bimodal shape, with peaks on January 25th and March 19th . This variation was consistent with a
380 growth than the less frequent sUAS snapshots. Both fused and original PS NDVI time series in 2018
gap in precipitation that greatly reduced plant growth in February, but growth resumed following late
381 showed a unique bimodal shape, with peaks on January 25th and March 19th. This variation was
March rainfall. The smoothed and fused NDVI time series successfully captured the rapid changes in
382 consistent with a gap in precipitation that greatly reduced plant growth in February, but growth
phenology, while preserving the magnitude of NDVI values with minimal atmospheric contamination
383 resumed following late March rainfall. The smoothed and fused NDVI time series successfully
(Figure 5).
384 captured the rapid changes in phenology, while preserving the magnitude of NDVI values with
385 minimal atmospheric contamination (Figure 5).
 Remote Sens. 2018, 10, x FOR PEER REVIEW 11 of 22

Remote Sens. 2019, 11, 595 11 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW 11 of 22

386
387 Figure 5. NDVI time series of a randomly selected grass pixel over a relatively homogeneous area,
386
388 from 10 sUAS missions (6 and 4 in the 2017 and 2018 growing season respectively), PlanetScope (PS),
389
387
Figure 5. NDVI
and data time
fusion. series
Also of a randomly
shown is daily selected grass (grey
precipitation pixel over
bar) afrom
relatively homogeneous
the on-site area,
rain gauge
Figure 5. NDVI time series of a randomly selected grass pixel over a relatively homogeneous area,
390
388
from 10 sUAS missions (6 and 4 in the 2017 and 2018 growing season respectively), PlanetScope (PS),
measurement.
from 10 sUAS missions (6 and 4 in the 2017 and 2018 growing season respectively), PlanetScope (PS),
and data fusion. Also shown is daily precipitation (grey bar) from the on-site rain gauge measurement.
389 and data fusion. Also shown is daily precipitation (grey bar) from the on-site rain gauge
391 NDVI maps derived from sUAS imagery showed similar spatiotemporal patterns to those from
390 measurement.
NDVI maps derived from sUAS imagery showed similar spatiotemporal patterns to those from
392 the PS satellite, as shown by the coincident images acquired by both sensors on November 19, 2016,
the PS satellite, as shown by the coincident images acquired by both sensors on 19 November 2016,
393
391 and January
NDVI 16, 2017
maps (Figure
derived 6).sUAS
from However, the showed
imagery sUAS NDVI map revealed greater spatialtodetails. The
and 16 January 2017 (Figure 6). However, the sUAS similar
NDVI spatiotemporal patterns
map revealed greater those
spatial from
details.
394
392 sUAS-resolution
the PS satellite, asNDVI
shown map on
theDecember 15, 2016 (Figure 6e),
bypredicted by the
onsimplified STARFM,
The sUAS-resolution NDVIby map coincident
on images
15 December 2016acquired
(Figure 6e), both sensors
predicted by the November 19, 2016,
simplified STARFM,
395
393 demonstrated
and January that
16, the
2017 prediction
(Figure 6). retained
However, sUAS
the spatial
sUAS resolution,
NDVI map as the
revealed road boundary
greater spatial and patterns
details. The
demonstrated that the prediction retained sUAS spatial resolution, as the road boundary and patterns
396
394 of NDVI heterogeneity
sUAS-resolution were
NDVIwere preserved,
map preserved,
on December and at the same time, captured phenological changes shown
of NDVI heterogeneity and15,
at2016 (Figure
the same 6e),captured
time, predicted by the simplified
phenological changes STARFM,
shown
397
395 in the PS time series.
demonstrated that the prediction retained sUAS spatial resolution, as the road boundary and patterns
in the PS time series.
396 of NDVI heterogeneity were preserved, and at the same time, captured phenological changes shown
397 in the PS time series.

398 Figure 6. NDVI maps at PlanetScope resolution (upper panels) and sUAS resolution (bottom panels).
Concurrent images were available on November 19, 2016 (a,d) and 16 January 2017 (c,f), while only
399 Figure 6. NDVI maps at PlanetScope resolution (upper panels) and sUAS resolution (bottom panels).
398
400 a PlanetScope image was available on 15 December 2016 (b). The predicted NDVI image from data
Concurrent images were available on November 19, 2016 (a and d) and January 16, 2017 (c, f), while
fusion at 30-cm resolution on December 15, 2016 is shown in (e).
401
399 only a PlanetScope image was available on December 15, 2016 (b). The predicted NDVI image from
Figure 6. NDVI maps at PlanetScope resolution (upper panels) and sUAS resolution (bottom panels).
402
400 data fusion atBetween
3.3. Relationships 30-cm resolution
Concurrent images were Remote on December
Sensing
available
15,and
Metrics
on November
2016 is shown
19,Biomass
in (e).
Measurements
2016 (a and d) and January 16, 2017 (c, f), while
401 only a PlanetScope
Daily fused Between image
NDVI was was available
positively on December 15, 2016 (b). The predicted NDVI image from
403 3.3. Relationships Remote Sensingrelated
Metrics(with varyingMeasurements
and Biomass significance) to biomass measurements
402 data fusion at 30-cm resolution on December 15, 2016 is shown in (e).
(mean = 1122 kg/ha; range = 17 - 4483 kg/ha) taken on the same day over all the clipping sites, with R2
404 Daily
ranging fused
from 0.01NDVI
to 0.33was
on positively related
various sUAS (with
image varying significance)
acquisition dates (FiguretoS4).
biomass measurements
The highest R2 , 0.33
403
405 3.3.< Relationships
(mean
(p = 1122
0.05, N = 21),Between
kg/ha;
wasrange Remote
found = for -Sensing
17 February Metrics
4483 kg/ha) andvarying
2018.taken
The Biomass
on RMeasurements
the same day over
2 among all the
different clipping
dates sites,
suggests with
limited
406
404 R 2 ranging from 0.01 to 0.33 on various sUAS image acquisition dates (Figure S4). The highest R2, 0.33
performance for using
Daily fused NDVIinstantaneous
was positivelyNDVI to predict
related biomasssignificance)
(with varying over the growing season.measurements
to biomass NDVI values
407
405 (p<0.05,
were N=21),
typically was
lower found
in for
2018 February
than those2018.
in The
2017 varying
for the R
same
2 among different dates suggests limited
month, consistent with
(mean = 1122 kg/ha; range = 17 - 4483 kg/ha) taken on the same day over all the clipping sites, with differences in
408
406 performance
corresponding for using
R2 ranging frombiomass instantaneous
measurements.
0.01 to 0.33 NDVI to predict biomass over the growing season. NDVI
on various sUAS image acquisition dates (Figure S4). The highest R2, 0.33 values
407 (p<0.05, N=21), was found for February 2018. The varying R2 among different dates suggests limited
408 performance for using instantaneous NDVI to predict biomass over the growing season. NDVI values
 Remote Sens. 2018, 10, x FOR PEER REVIEW 12 of 22

Remote Sens. 2018, 10,

595x FOR PEER REVIEW 1212
of of
409 Remote
were Sens. 2019,
typically 11,
lower in 2018 than those in 2017 for the same month, consistent with differences in22
22

410 corresponding biomass measurements.

409 were typically lower in 2018 than those in 2017 for the same month, consistent with differences in
411 When integrated from the germination date to the date of observation (using the fused time
410 When integrated
corresponding from
biomass the germination date to the date of observation (using the fused time
measurements.
412 series), the cumulative NDVI explained 68% of the variance in measured biomass among all field
411 series),When
the cumulative
integrated from explained
NDVI 68% ofdate
the germination the variance
to the datein measured biomass
of observation among
(using all field
the fused time
413 sampling plots during the two growing seasons (Figure 7a), and the cumulative APAR explained
412 sampling
series), plots during the NDVI
the cumulative two growing
explainedseasons
68%(Figure 7a), and the
of the variance cumulativebiomass
in measured APAR explained
among all69% field
414 69% of spatial and temporal variance (Figure 7b) (N = 220). The linear regression model, based on
413 of sampling
spatial and temporal
plots duringvariance
the two(Figure
growing 7b)seasons
(N = 220). The7a),
(Figure linear
andregression model,APAR
the cumulative based explained
on 70%
415 70% of the randomly selected data points, showed that cumulative APAR predicted biomass with an
414 of 69%
the randomly selected data points, showed that cumulative APAR predicted biomass
of spatial and temporal variance (Figure 7b) (N = 220). The linear regression model, based with an R2on
416 R2 of 0.67±0.06 and a RMSE of 631±82 kg/ha, when compared with the remaining 30% of the data
415 of 70% ± the
0.67 of 0.06randomly
and a RMSE of 631
selected ± 82
data kg/ha,
points, whenthat
showed compared with APAR
cumulative the remaining
predicted30% of thewith
biomass dataan
417 (N=66). Similarly, predictions from cumulative NDVI showed an R2 of 0.67±0.05 and an RMSE of
416 (NR2= 66). Similarly, predictions from cumulative NDVI showed an R2 of 0.67 ± 0.05
of 0.67±0.06 and a RMSE of 631±82 kg/ha, when compared with the remaining 30% of the data and an RMSE
418 627±73 kg/ha. However, the empirical NDVI-only and APAR-only methods may have large
417 627 ± 73
of (N=66). kg/ha. However,
Similarly, predictionsthefromempirical NDVI-only
cumulative and APAR-only
NDVI showed an R2 ofmethods
0.67±0.05may
and have large of
an RMSE
419 uncertainties when extrapolated to other areas or time periods.
418 uncertainties
627±73 kg/ha.whenHowever,
extrapolated
the toempirical
other areas or time periods.
NDVI-only and APAR-only methods may have large
419 uncertainties when extrapolated to other areas or time periods.

420
Figure 7. Scatterplots of measured biomass versus the cumulated NDVI (a) and cumulated APAR (b)
421
420 Figure 7. Scatterplots of measured biomass versus the cumulated NDVI (a) and cumulated APAR (b)
over 220 clip-plot samples in 2017 and 2018.
422 over 220 clip-plot samples in 2017 and 2018.
421 Figure 7. Scatterplots of measured biomass versus the cumulated NDVI (a) and cumulated APAR (b)
3.4. Forage Production Models
422 over 220 clip-plot samples in 2017 and 2018.
423 3.4. Forage Production Models
Optimization of the topography-based LUE model (Model I) with 70% of randomly selected data
424
423 3.4.Optimization
resultedForage
in the of theModels
Production
following topography-based
forage LUE model (Model I) with 70% of randomly selected data
production model:
425 resulted in the following forage production model:
424 Optimization of the topography-based LUE model . (Model I) with 70% of randomly selected data
57.81
426
425 𝑀𝑜𝑑𝑒𝑙Model
resulted in𝐼:the
𝐵𝑖𝑜𝑚𝑎𝑠𝑠 = ∑forage
I : Biomass
following (𝐴𝑃𝐴𝑅) tp
= ∑production
∗ 0.18 ∗ ( 0.18)∗(∗ (
t0 ( APAR ) ∗(model: )
+ 0.1) …… (8) + 0.1) (8)
(1 + IC ) ∗ ( Elevation − 430)
427
426 LUE decreased
𝑀𝑜𝑑𝑒𝑙 with=both
𝐼: 𝐵𝑖𝑜𝑚𝑎𝑠𝑠 elevation
∑ (𝐴𝑃𝐴𝑅) and IC.
∗ 0.18 ∗ ((When
.
compared with the ……
+ 0.1) remaining
(8) 30% of the data
)∗( )
428 (N=66),
LUEpredicted
decreasedbiomass showed
with both an Rand
elevation 2 of 0.70 and ancompared
IC. When RMSE ofwith 567 kg/ha (Figure 8a),
the remaining 30%suggesting
of the dataa
427 LUE decreased with both elevation2 and IC. When compared with the remaining 30% of the data
429 better
(N performance
= 66), (a decrease
predicted biomass showedof 56ankg/ha
R of in RMSE)
0.70 and anthan theof
RMSE linear model(Figure
567 kg/ha driven 8a),
onlysuggesting
by APAR
428 (N=66), predicted biomass showed an R2 of 0.70 and an RMSE of 567 kg/ha (Figure 8a), suggesting a
430 a(Figure 7b).
better performance (a decrease of 56 kg/ha in RMSE) than the linear model driven only by APAR
429 better performance (a decrease of 56 kg/ha in RMSE) than the linear model driven only by APAR
(Figure 7b).
430 (Figure 7b).

431
431
432 Figure 8. Biomass predicted by models with LUE optimized as functions of (a) elevation and IC
433 (Model8.I);Biomass
Figure and (b) elevation,
predicted IC,
by and
modelsair temperature and soil moisture
with LUE optimized stressofterms
as functions (Model II).
(a) elevation andEach
IC
432 Figure 8. Biomass predicted by models with LUE optimized as functions of (a) elevation and IC
434 point represents
(Model a pair of the
I); and (b) elevation, IC,predicted and observed
and air temperature and biomass fromstress
soil moisture the validation
terms (Model dataset.
II). Each point
433 (Model I); and (b) elevation, IC, and air temperature and soil moisture stress terms (Model II). Each
represents a pair of the predicted and observed biomass from the validation dataset.
434 point represents a pair of the predicted and observed biomass from the validation dataset.
 Remote Sens. 2019, 11, 595 13 of 22

RemoteThe
Sens.final optimized
2018, 10, x FOR PEERLUEREVIEW Model II was generated as a function of soil moisture 13(SM), of 22
air temperature (Ta), elevation and IC:
435 The final optimized LUE Model II was generated as a function of soil moisture (SM), air
tp 1 4
436 Model I(Ta),
temperature I : Biomass = ∑and
elevation IC: ∗ 1+e−46.16(SM−0.144) ∗ (1+e1.89(10.33−Ta) )∗(1+e1.89(−10.33+Ta) ) ) ∗ 0.032∗
t0 ( APAR
(9)
437 𝑀𝑜𝑑𝑒𝑙
( ( IC+1𝐼𝐼: 56.42
𝐵𝑖𝑜𝑚𝑎𝑠𝑠
)∗( Elevation ) ∑ (𝐴𝑃𝐴𝑅 ∗
−430=
+ 0.13) . ( . ) ∗ . ( . . ( ∗ . ) )) ∗ 0.032 ∗
.
438 (( The
)∗( soil moisture )
+ 0.13)
scalar…… showed (9) that decreasing moisture down-regulated LUE. The W scalar
439 The logistically
increases soil moisture fromscalar
0.5 showed
to 1 as soil that decreasing
moisture moisture
changes fromdown-regulated
wilting point (θvLUE. The
= 0.14 cmW −3 )
scalar
3 cm

440 increases logistically from 0.5 3to 1 −

as3 soil moisture changes from
to field capacity (θv = 0.29 cm cm ). The T scalar dropped to 0.5 when air temperature was 1 C wilting point (θ v = 0.14 cm 3 cm-3)◦ to

441 field capacity

below or above(θ v =optimal
the 0.29 cmtemperature of 10.33 ◦ C. This full
3 cm-3). The T scalar dropped to 0.5modelwhen air temperature
resulted in improved was 1˚C below
accuracy for
442 or above the optimal temperature of 10.33 ˚C. This full model
predicting biomass with greater explanation of variance (81%) and lower RMSE (542 kg/ha) compared resulted in improved accuracy for
443 predicting
with biomass with LUE
the topography-based greater model explanation
(Figure 8b). of This
variance
suggests (81%) thatand lower RMSE
time-varying (542 of
controllers kg/ha)
soil
444 compared
moisture andwith the topography-based
air temperature LUE model (Figure
captured environmental stress impacts8b). This suggests
on forage that time-varying
production.
445 controllers
When usingof soilfield-measured
moisture and soil air moisture
temperature andcaptured environmental
soil temperature, Model III: stress impacts on forage
446 production.
447 tp
When using field-measured
Biomass = ∑ ( APAR ∗
soil moisture
1 and soil temperature,
∗ 4 Model III:
) ∗ 0.0625∗
1+e−58.86(SM−0.144) ( 1+e0.84(14.21−ST ) )∗(1+e0.84(−14.21+ST ) ) (10)
t0 1 4
448 𝐵𝑖𝑜𝑚𝑎𝑠𝑠 = −1.56e
(𝐴𝑃𝐴𝑅 −3∗∗ Elevation
1+𝑒 . ( + 1.73
. )
∗


(1 + 𝑒 . ( . ) ∗ (1 + 𝑒 . ( . ) )
)) ∗ 0.0625 ∗ (−1.56𝑒 ∗ 𝐸𝑙𝑒𝑣𝑎𝑡𝑖𝑜𝑛

449 + 1.73) … … (10)

predicted biomass with an RMSE of 472 kg/ha and R2 of 0.77 (Figure S5).
450 predicted biomass with an RMSE of 472 kg/ha and R2 of 0.77 (Figure S5).
3.5. Forage Production Mapping and Patterns
451 3.5. Forage Production Mapping and Patterns
We summarized forage production across the study area based on daily biomass maps generated
452 from WeModelsummarized forage production
II. Forage production increasedacross the during
gradually study area based on
the growing dailybut
season, biomass
followedmaps
an
453 generated from Model II. Forage production increased gradually during the growing
unstable trajectory with bursts and plateaus (Figure 9). Most rapid growth was from 20 March to season, but
454 2followed
April 2017an(averaging
unstable trajectory
107 kg/ha/d)with and
bursts
fromand10plateaus
March to(Figure 9). 2018
19 March Most(averaging
rapid growth was from
15 kg/ha/d).
455 March 20 to April 2, 2017 (averaging 107 kg/ha/d) and from March 10 to March 19,
The 2017 growing season had higher overall growth rates than 2018, which resulted in a mean peak 2018 (averaging
456 15 kg/ha/d).
standing The 2017
biomass growing
of 3216 seasonkg/ha
(±678SD) had higher
versusoverall
1054 (growth
±374SD)rates than 2018,
kg/ha, which resulted
respectively. in a
This large
457 mean peak standing biomass of 3216 (±678SD) kg/ha versus 1054 (±374SD) kg/ha,
difference in peak standing biomass was linked to higher precipitation in 2017 (287 mm) compared respectively. This
458 large
to 2018difference
(123 mm).inThe peak standing
standard biomassofwas
deviation linked
biomass fortothe
higher
sUASprecipitation in 20179)(287
flight dates (Figure mm)
showed
459 compared to 2018 (123 mm). The standard deviation of biomass
increasing spatial variability in biomass as the growing season progressed. for the sUAS flight dates (Figure 9)
460 showed increasing spatial variability in biomass as the growing season progressed.

461
462 Figure9.9. Daily
Figure Daily time
time series
series of
of the
the estimated
estimated biomass
biomass averaged
averaged across
across the
the study
study area
area in
in2017
2017and
and2018
2018
463 growing seasons. Error bars, denoting the standard deviation, were added to the eight
growing seasons. Error bars, denoting the standard deviation, were added to the eight sUAS flight sUAS flight
464 dates,
dates,indicating
indicatingthetheincrease
increaseof ofspatial
spatialvariability
variabilityin
inbiomass
biomassas
asplants
plantsgrow.
grow.

465 Predicted
Predictedbiomass
biomassmaps
mapsat at30-cm
30-cm resolution
resolutionprovided
providedaa visual
visual representation
representationofofspatial
spatial variation
variation
466 across
across the landscape (Figure 10). At peak biomass in 2017 (Figure 10d), high biomass production was
the landscape (Figure 10). At peak biomass in 2017 (Figure 10d), high biomass production was
467 predicted
predictedatatlow
lowelevation
elevation(Figure
(Figure10i)
10i)and
andhigh
highIC
IC(Figure
(Figure10j)
10j)regions,
regions,whereas
whereasinin2018,
2018,high
highbiomass
biomass
468 production
production was
was predicted
predicted in
inlow
lowelevation
elevation and
andlow
lowICICregions.
regions. Lower
Lower elevation
elevation regions
regions often
often have
have
469 higher soil moisture because lower slope positions receive runoff and throughflow from upper
470 hillslope positions. IC values are positively related to PAR with higher PAR leading to higher biomass
471 when plants are not water stressed. However, when water is limited, high PAR could contribute to
472 lower biomass production (i.e., greater plant water stress). The maps also captured human
 Remote Sens. 2018, 10, x FOR PEER REVIEW 14 of 22

473 disturbance
Remote on 595
Sens. 2019, 11, biomass production, such as the road having very low biomass compared to14the
of 22
474 surrounding area.
475 A zonal statistical analysis was performed on Biomass-Slope and Biomass-Aspect to examine
476 higher soil moisture
interactions because
between lower slope
topography positionsThe
and biomass. receive runoff
model and throughflow
predicted frominupper
higher biomass flatterhillslope
areas
477 positions. IC values are positively related to PAR with higher PAR leading to higher
(Figure 11a) while the steeper group had lower biomass than the other two groups in both years. biomass when
478 plants are notthe
However, water stressed.
model However,
predicted similar when
biomasswater is limited,
values for the high PARmoderate
flat and could contribute
groups into2017
lower
479 biomass
(wetterproduction (i.e., greater
year). The model plant
predicted water
lower stress).
biomass onThe maps also
north-facing captured
slopes human disturbance
than south-facing slopes inon
480 biomass production,
2017; however, such as the
an opposite road
effect washaving very
observed inlow
2018biomass compared to the surrounding area.
(Figure 11b).

481
Figure 10. Spatial distribution of the estimated biomass on days with sUAS flights during 2017 and
482 2018 growing
Figure seasons
10. Spatial (a–h). Alsoofshown
distribution are elevation
the estimated (i)on
biomass and illumination
days with sUAScondition (IC) maps
flights during (j).
2017 and
483 2018 growing seasons (a-h). Also shown are elevation (i) and illumination condition (IC) maps (j).
A zonal statistical analysis was performed on Biomass-Slope and Biomass-Aspect to examine
interactions between topography and biomass. The model predicted higher biomass in flatter areas
(Figure 11a) while the steeper group had lower biomass than the other two groups in both years.
However, the model predicted similar biomass values for the flat and moderate groups in 2017
 Remote Sens. 2019, 11, 595 15 of 22

(wetter year). The model predicted lower biomass on north-facing slopes than south-facing slopes
Remote Sens. 2018, 10, x FOR PEER REVIEW 15 of 22
in 2017; however, an opposite effect was observed in 2018 (Figure 11b).

484
485 Figure 11. Variation
Figure 11. Variation of
of the
the predicted
predicted peak
peak biomass
biomass byby (a)
(a) slope
slope and
and (b)
(b) aspect.
aspect. The
The area
area was
was divided
divided
◦ ), moderate (5~15◦ ), and steep (>15◦ ).
486 into three categories of slopes: flat terrain (0~5
into three categories of slopes: flat terrain (0~5º), moderate (5~15º), and steep ( >15º).

4. Discussion
487 4. Discussion
In this pioneering study, we developed two remote sensing data-driven LUE models that
488 In this pioneering study, we developed two remote sensing data-driven LUE models that
accurately predicted daily forage production at centimeter-scale resolution. We also streamlined
489 accurately predicted daily forage production at centimeter-scale resolution. We also streamlined
preprocessing and fusing approaches for raw sUAS data with high-resolution satellite remote
490 preprocessing and fusing approaches for raw sUAS data with high-resolution satellite remote sensing
sensing data to obtain high-spatiotemporal data. We successfully implemented preprocessing and
491 data to obtain high-spatiotemporal data. We successfully implemented preprocessing and fusing
fusing methods and frameworks (e.g., the C model for illumination correction and the simplified
492 methods and frameworks (e.g. the C model for illumination correction and the simplified STARFM)
STARFM) that were developed for satellite remote sensing data on the sUAS data. From the model
493 that were developed for satellite remote sensing data on the sUAS data. From the model predicted
predicted forage production maps, we observed unique patterns in productivity related to different
494 forage production maps, we observed unique patterns in productivity related to different
environmental and climatic drivers. The following discussion section expands on additional findings
495 environmental and climatic drivers. The following discussion section expands on additional findings
in sUAS data preprocessing (Section 4.1) and fusing with PlanetScope data (Section 4.2), and the
496 in sUAS data preprocessing (4.1) and fusing with PlanetScope data (4.2), and the response of forage
response of forage plants to precipitation (Section 4.3) and soil moisture and temperature (Section 4.4).
497 plants to precipitation (4.3) and soil moisture and temperature (4.4).
4.1. Variations in the Illumination Effect
498 4.1. Variations in the Illumination Effect
This study demonstrated a successful application of the C model for correcting sUAS data for
499 This study
illumination demonstrated
effects. Reflectancea in successful application
visible bands of the Cless
was generally model for correcting
affected by terrainsUASthan NIRdataand
for
500 illumination effects. Reflectance in visible bands was generally less
red-edge bands. The same phenomenon was found for topographic illumination effects on Landsataffected by terrain than NIR and
501 red-edge
images bands.
[50]. The The same
authors phenomenon
concluded waslower
that the foundcorrelation
for topographic illumination
for visible bands was effects
causedon Landsat
by their
502 imagesreflectance
lower [50]. The authors
rate and concluded that the lower
a more significant correlation
atmospheric for visible
scattering bands
effect onwas caused
visible bandsby than
their
503 lower reflectance rate and a more significant atmospheric scattering
longer-wavelength bands. From the temporal analysis, the topographic illumination effect intensity effect on visible bands than
504 longer-wavelength bands. From the temporal analysis, the topographic
was not consistent over the entire growing season for a given band, but rather diminished as the illumination effect intensity
505 was not consistent
growing over the entire
season progressed. growing
In addition to season
changesforinavegetation
given band, but rather
density, diminished as
the diminishing the
trend
506 growing season progressed. In addition to changes in vegetation density,
is probably related to changes in sun angle. Since the study site is at 35.5N, the solar-noon sun the diminishing trend is
507 probably angle
elevation relatedincreases
to changes fromin 31
sun angle.
◦ in Since the
December study
to >80 ◦ insite is at
June. 35.5N, thethe
Therefore, solar-noon
theoretical sun elevation
intensity of
508 angle increases from 31˚ in December to >80º in
illumination effects should decrease as the growing season progresses. June. Therefore, the theoretical intensity of
509 illumination effects should decrease as the growing season progresses.
Notably, the R2 between red band surface reflectance and IC was very low (0.02) for the March
510 flight,Notably,
and the the R2 between red
corresponding band surface
corrected R2 (0.06) reflectance
was higher and IC the
than wasoriginal
very low R2(0.02) for theboth
. Although MarchR2
511 flight, and the corresponding corrected R 2 (0.06) was higher than the original R2. Although both R2
values were small, this raises caution that applying the C model may lead to overcorrection when
512 values
the were small,
illumination this raises
effect is weak. caution that applying
Therefore, thethe
a test of C model may lead
illumination to overcorrection
effect intensity, for when
example,the
513 aillumination effect ison
correlation analysis weak. Therefore,
the surface a test and
reflectance of the illumination
IC, is recommended effect intensity,
before applying forillumination
example, a
514 correlation analysis on the surface reflectance and IC, is recommended before
correction to models. In this study, since the corrected red band surface reflectance-IC R2 value was applying illumination
515 correction
very low, weto used
models.the In this study,
corrected red since the corrected
band reflectance red band temporal
to maintain surface reflectance-IC
consistency among R2 value was
data.
516 very low, we used the corrected red band reflectance to maintain temporal consistency among data.
4.2. Fusing Satellite and sUAS Data
517 4.2. Fusing Satellite and sUAS Data
We found that the PS NDVI values were generally lower than those from sUAS, especially when
518 NDVIWe found
was higherthatthan
the 0.5.
PS NDVI values from
This results were PS generally lowercalculated
NDVI being than thosefrom fromthe sUAS,Topespecially
of Atmospherewhen
519 NDVI was higher than 0.5. This results from PS NDVI being calculated from the Top of Atmosphere
520 (TOA) reflectance while the sUAS NDVI is determined from the surface reflectance. Several studies
521 document a lower TOA NDVI compared to NDVI at the surface [70,71] due to the atmospheric
522 scattering effect. In the simplified STARFM, we used the temporal change of PS TOA NDVI, which
Remote Sens. 2019, 11, 595 16 of 22

(TOA) reflectance while the sUAS NDVI is determined from the surface reflectance. Several studies
document a lower TOA NDVI compared to NDVI at the surface [70,71] due to the atmospheric
scattering effect. In the simplified STARFM, we used the temporal change of PS TOA NDVI, which
reduced the noise from TOA reflectance, while the absolute value relies heavily on the sUAS. Smaller
differences between December and February sUAS instantaneous NDVI and PS NDVI datasets may
result from the use of a sun light irradiance (SLR) sensor on sUAS flights after March 2017. We initially
relied on a white calibration panel for radiometric and atmospheric correction when preprocessing
data from the first two flights before installation of the SLR sensor.
Data fusion of sUAS and high spatial-resolution images has great potential to lower the cost of
operating frequent sUAS flights, especially given increasing availability of high-spatial resolution
data (e.g., RapidEye data and PS data by Planet and WorldView data by DigitalGlobe). However,
much of the high-spatial resolution satellite (including PS) data is only available in DN or TOA
due to the lack of shortwave NIR bands for atmospheric correction. Therefore, it is important to
develop robust data fusion methods that work on multi-source data with systematic differences.
Future release of atmospherically-corrected PS surface reflectance products, i.e., using the concurrent
MODIS data, will further reduce uncertainties in sUAS-satellite data fusion methods. Moreover, a more
sophisticated fusion method is needed to take into account the spatial details embedded in sUAS data
for the temporal interpolation of PS data.

4.3. LUE Parameterization

We observed improvements in model accuracy when including additional model terms. RMSEs
of forage production estimates from APAR-only, Model I, Model II, and Model III were 624 kg/ha,
567 kg/ha, 542 kg/ha, and 472 kg/ha, respectively. We selected Model II to map forage production
because it had the highest accuracy and does not require ground measured soil moisture and
temperature. Model III, which uses locally measured soil moisture and soil temperature, achieved
a higher accuracy than Model II, which uses estimated soil moisture and coarse-resolution air
temperature. We believe that the main source of error in Model II is from the use of a simple bucket
model for estimating soil moisture and/or the coarse resolution of air temperature.
Comparing the stressing scalars in Model II and III, the W scalar functions have very similar
controls on LUE; however, the T scalar behaves quite differently. In Model II, the T scalar is driven
by air temperature and is more sensitive to deviations from optimal air temperature. The estimated
LUE0 across the three models showed large discrepancies, ranging from 0.32 to 1.8 g/MJ APAR with
Model I having the highest LUE0 and Model II the lowest. Over the years, maximal LUE has been
estimated using different methods for different biomes [72]. However, maximal LUE used in existing
LUE models varies a lot [31] because researchers used different methods to estimate this metric [72].
In addition, maximal LUE also varies across scales. Currently, the number of LUE studies at the
watershed scale is very limited. We expect to see increasingly more relevant studies on LUE modeling
with advances in sUAS technology.
Using measured soil temperature and moisture in Model III achieved the highest accuracy among
the three models, but this model cannot be applied to directly generate biomass maps because it
requires point measurement data to parameterize input variables. As of now, daily soil moisture
maps at high spatial resolution are not available, but with the continuous expanding constellations of
satellites, new remote sensing products are being produced that could provide daily spatial patterns for
soil moisture. For example, a past study demonstrated the potential for retrieving soil moisture from
C-band synthetic aperture radar (SAR) data [73]. Additional studies have developed algorithms for
mapping high-resolution (<1 km) soil moisture locally using the sentinel-1 C-band SAR data at a 10-m
resolution [74,75]. Therefore, inclusion of near real-time soil moisture estimates may be forthcoming as
new remote sensing technologies emerge.
 Remote Sens. 2019, 11, 595 17 of 22

4.4. Response of Forage Production and Plant Phenology to Moisture

Moisture is the primary controlling factor of forage production in Mediterranean annual range
systems [6,31,76]. The different precipitation regimes in the 2017 and 2018 growing seasons triggered
very different biomass-aspect relationships. South-facing slopes were expected to have a lower peak
biomass than north-facing slopes because of increased temperatures leading to higher ET and lower
Remote Sens. 2018, 10, x FOR PEER REVIEW 17 of 22
soil moisture on south-facing slopes. However, we observed a higher peak biomass on south-facing
573 slopes
slopes in inthe
the2017
2017growing
growingseason season (Figure
(Figure 10b). This
10b). resulted
This resultedfromfromthe the
highhigh
precipitation
precipitation(287 (287
mm)mm) that
574 was well distributed throughout the 2017 growing season. With
that was well distributed throughout the 2017 growing season. With sufficient precipitation to sufficient precipitation to maintain
575 soil moisture
maintain soilcontent
moisture (Figure
contentS6),(Figure
plants experienced little water deficit.
S6), plants experienced With sufficient
little water water
deficit. With supply,
sufficient
576 PAR
waterbecame
supply, the primary
PAR became limiting factor. Therefore,
the primary south-facing
limiting factor. slopes,
Therefore, where more
south-facing solarwhere
slopes, radiationmore is
577 received, were able to produce higher peak biomass than north-facing
solar radiation is received, were able to produce higher peak biomass than north-facing slopes. In slopes. In contrast, the 2018
578 growing
contrast, season
the 2018received
growingonly season108received
mm of precipitation
only 108 mm with large gaps with
of precipitation between largerainfall events.
gaps between
579 The
rainfall events. The majority of the precipitation was received in January, March and early in
majority of the precipitation was received in January, March and early April resulting an
April
580 extremely dry February (Figures S5 and S6). This precipitation distribution
resulting in an extremely dry February (Figure S6 & 5). This precipitation distribution hindered plant hindered plant growth on
581 the radiation-rich
growth south-facingsouth-facing
on the radiation-rich slopes by intensifying the soil moisture
slopes by intensifying the soildeficit. Therefore,
moisture a lower
deficit. peak
Therefore,
582 biomass was observed on south-facing slopes in the 2018
a lower peak biomass was observed on south-facing slopes in the 2018 growing season. growing season.
583 Time
Time series
series biomass
biomass maps maps provided
provided insights
insights on on plant
plant phenology
phenology changeschanges in in response
response to to different
different
584 precipitation
precipitation regimes during the 2017 and 2018 growing seasons. Germination and growthbe
regimes during the 2017 and 2018 growing seasons. Germination and growth can canvery
be
585 different
very different interannually in annual range systems. Peak growth is highly dependent on and
interannually in annual range systems. Peak growth is highly dependent on the amount the
586 timing
amountofand precipitation. The November
timing of precipitation. The2017 rainfall resulted
November in germination
2017 rainfall resulted in(11/11/2016)
germination of the plants
(11/11/2016)
587 about 2 months earlier than the 2018 growing season (1/1/2018)
of the plants about 2 months earlier than the 2018 growing season (1/1/2018) (Figure 12 & 5). The (Figures 5 and 12). The predicted
588 germination dates are close
predicted germination datesto actual
are closegemination
to actualdates (11/6/2016
gemination datesand 1/13/2018)
(11/6/2016 and captured
1/13/2018) bycaptured
the time
589 lapse
by thecamera.
time lapseAlthough
camera. the Although
two growing theseasons had a 1.5-month
two growing seasons had difference in theirdifference
a 1.5-month germination in date,
their
590 they both had peak growth in April. We predicted 4/6/2017 and
germination date, they both had peak growth in April. We predicted 4/6/2017 and 4/14/2018 as 4/14/2018 as the peak growth dates
the
591 for
peakthegrowth
2017 and 2018
dates forgrowing
the 2017season,
and 2018 which wereseason,
growing close towhich
the camera observed
were close to the peak growth
camera dates
observed
592 (4/12/2017
peak growth and 4/25/2018).
dates (4/12/2017Inand addition to having
4/25/2018). a longertogrowing
In addition having season,
a longerplants
growing alsoseason,
accumulatedplants
593 biomass faster in 2017 than in 2018 (Figure 9). This higher biomass
also accumulated biomass faster in 2017 than in 2018 (Figure 9). This higher biomass accumulation accumulation rate is largely
594 due
rate to the higher
is largely due and more
to the evenly
higher and distributed
more evenlyprecipitation pattern during
distributed precipitation 2017.during
pattern Plant growth
2017. Planton
595 south-facing slopes was very sensitive to soil moisture availability.
growth on south-facing slopes was very sensitive to soil moisture availability. During the 2017 During the 2017 growing season
596 and in January
growing season 2018,
andwhen moisture
in January waswhen
2018, sufficient, plantswas
moisture on south-facing
sufficient, plantsslopesonaccumulated
south-facing biomass
slopes
597 faster than those on north-facing slopes. However, when soil moisture
accumulated biomass faster than those on north-facing slopes. However, when soil moisture became became limiting, for example,
598 starting
limiting,infor February
example,2018, biomass
starting accumulation
in February on south-facing
2018, biomass accumulation slopeson was distinctly lower
south-facing slopesthanwas
599 north-facing
distinctly lower slopes.
than north-facing slopes.

600
601 Figure 12. Daily
Figure 12. Daily average
average biomass
biomass time
time series
series of
of biomass
biomass averaged
averaged over
over 500
500 randomly
randomly selected
selected pure
pure
602 forage
forage pixels (at 30cm resolution) in the study area during 2017-2018 growing season. The time
pixels (at 30cm resolution) in the study area during 2017-2018 growing season. The time series
series
603 shows
shows aa different
different phenology
phenology for
for plants
plants growing
growingon onsouth-
south-and
andnorth-facing
north-facingslopes.
slopes.

604 5. Conclusions
605 This pilot study demonstrates the synergistic use of complementary sUAS and satellite remote
606 sensing technology to map daily forage production at 30-cm resolution on a delayed grazing
Remote Sens. 2019, 11, 595 18 of 22

5. Conclusions
This pilot study demonstrates the synergistic use of complementary sUAS and satellite remote
sensing technology to map daily forage production at 30-cm resolution on a delayed grazing rangeland.
Remote sensing-based APAR estimates were the primary driver for both LUE-based models developed
here. LUE was optimized as a function of elevation and illumination conditions in Model I, achieving
an R2 of 0.70 and an RMSE of 567 kg/ha when comparing predicted forage production with
measured forage biomass. By further incorporating moisture and temperature stress terms to the LUE
parameterization (Model II), Model II predicted forage production with higher accuracy (R2 = 0.81)
and lower RMSE (542 kg/ha). Both LUE approaches showed improved prediction accuracy compared
to the univariate linear regression with cumulative APAR. Finally, the inclusion of measured soil
moisture and soil temperature into the LUE model (Model III) achieved the lowest RMSE (472 kg/ha)
with a similar R2 of 0.77.
The fusion of high spatial resolution sUAS data and high temporal PlanetScope satellite imagery
enabled us to generate daily forage production maps at a 30-cm resolution. The maps allowed
the analysis of landscape-derived zonal statistics to assess (a) interactions between topography and
biomass and (b) the response of forage production and plant phenology to changing precipitation
regimes. Our analysis showed higher forage production and biomass accumulation rates in landscape
positions experiencing less environmental stress (e.g., soil water deficit), with several differences
occurring between wet and dry years. The forage production maps can be implemented in decision
support tools to help ranchers better anticipate weather-driven changes in forage production and
optimize their decisions on proactive practices, such as stocking conservatively, resting pasture,
and incorporating yearling cattle. This study provides basic tools that can be further developed
and scaled to statewide regions to provide near real time forage availability delivered to users via
internet apps.

Supplementary Materials: The following are available online at http://www.mdpi.com/2072-4292/11/5/595/s1,

Figure S1: sUAS flight plan; Figure S2: (a) Observed and (b) predicted NDVI image on 4/6/2017, and the
(c) corresponding difference image; Figure S3: Scatter plot (N = 10000) of (a) observed and predicted NDVI
on 4/6/2017 and scatter plot of (b) observed NDVI on 4/6/2017 and 4/30/2017; Figure S4: Scatterplots of
instantaneous sUAS NDVI and measured biomass on the eight flight days. Each data point represents a single
ground sampling point on a specific day; Figure S5: Scatterplot of the Model III predicted biomass versus
measured biomass; Figure S6: Calibrated soil moisture time series for the 2017 and 2018 growing seasons.
Author Contributions: H.L., Y.J., and R.A.D. conceived and designed the research. H.L., S.M.D., R.E.L., R.A.D.,
S.C., and A.J.Y.W. carried out the field measurement. H.L. and A.J.Y.W. conducted the sUAS flights. H.L.
processed the data and developed the LUE models. H.L. and Y.J. analyzed the results and drafted the manuscript.
H.L., R.A.D., R.E.L., S.M.D., L.M.R., A.T.G., A.J.Y.W., S.C. and Y.J. provided critical feedback and helped edited
the manuscript.
Funding: This study was supported by the University of California—Davis Russell L. Rustici Rangeland and
Cattle Research Endowment.
Acknowledgments: We thank the Morrison Family for permitting access to Camatta Ranch for data collection
and their cooperation throughout this study period. We also thank Mui Lay and Karl Linsteadt for their help with
field measurements.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Roche, L.M.; Schohr, T.K.; Derner, J.D.; Lubell, M.N.; Cutts, B.B.; Kachergis, E.; Eviner, V.T.; Tate, K.W.
Sustaining Working Rangelands: Insights from Rancher Decision Making. Rangel. Ecol. Manag. 2015, 68,
383–389. [CrossRef]
2. Schwarzenegger, A.; Snow, L.; Walters, D. California’s Forests and Rangelands: 2010 Assessment.
Available online: http://frap.fire.ca.gov/data/assessment2010/pdfs/california_forest_assessment_nov22.
pdf (accessed on 17 April 2017).
3. California Department of Food & Agriculture. 2015 Crop Year Report; California Department of Food &
Agriculture: Sacramento, CA, USA, 2015.
Remote Sens. 2019, 11, 595 19 of 22

4. Becchetti, T.; George, M.; McDougald, N.; Dudley, D.M.; Connor, M.; Flavell, D.K.; Vaughn, C.E.; Forero, L.C.;
Frost, B.; Oneto, S.R.; et al. Annual Range Forage Production. Univ. Calif. Agric. Nat. Resour. Publ. 2016, 8018,
1–12.
5. Sloat, L.L.; Gerber, J.S.; Samberg, L.H.; Smith, W.K.; Herrero, M.; Ferreira, L.G.; Godde, C.M.; West, P.C.
Increasing Importance of Precipitation Variability on Global Livestock Grazing Lands. Nat. Clim. Chang.
2018, 8, 214–218. [CrossRef]
6. Larsen, R.; Striby, K.; Horney, M. Fourteen Years of Frage Monitoring on the California Central Coast Shows
Tremendous Variation; General Technical Reports PSW-GTR-251; US Department of Agriculture, Forest Service,
Pacific Southwest Research Station: Berkeley, CA, USA, 2014; pp. 273–281.
7. Intergovernmental Panel on Climate Change [IPCC]. Climate Change 2013: The Physical Science Basis.
Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change;
Stocker, T., Qin, D., Plattner, G., Tignor, M., Allen, S., Bosching, J., Nauels, A., Xia, Y., Bex, V., Midgley, P.,
Eds.; Cambridge University Press: Cambridge, UK, 2013.
8. Coppock, D.L.; Fernández-Giménez, M.; Hiernaux, P.; Huber-Sannwald, E.; Schloeder, C.; Valdivia, C.;
Arredondo, J.T.; Jacobs, M.; Turin, C.; Turner, M. Rangeland Systems in Developing Nations: Conceptual
Advances and Societal Implications. In Rangeland Systems; Briske, D.D., Ed.; Springer: Cham, Switzerland,
2017; pp. 569–641. ISBN 978-3-319-46707-8.
9. George, M.R.; Williams, W.A.; McDougald, N.K.; Clawson, W.J.; Murphy, A.H. Predicting Peak Standing
Crop on Annual Range Using Weather Variables. J. Range Manag. 1989, 42, 508–512. [CrossRef]
10. Liacos, L.G. Soil Moisture Depletion in the Annual Grass Type. J. Range Manag. 1962, 15, 67–72. [CrossRef]
11. O’ Geen, A.T. Soil Water Dynamics. Nat. Educ. Knowl. 2012, 3, 12.
12. Beaudette, D.E.; O’Geen, A.T. Quantifying the Aspect Effect: An Application of Solar Radiation Modeling
for Soil Survey. Soil Sci. Soc. Am. J. 2009, 73, 1755. [CrossRef]
13. Wang, J.Y. A Critique of the Heat Unit Approach to Plant Response Studies. Ecology 1960, 41, 785–790.
[CrossRef]
14. George, M.R.; Raguse, C.A.; Clawson, W.J.; Wilson, C.B.; Willoughby, R.L.; McDougald, N.K.; Duncan, D.A.;
Murphy, A.H. Correlation of Degree-Days with Annual Herbage Yields and Livestock Gains. J. Range Manag.
1988, 41. [CrossRef]
15. Murphy, A.H. Predicted Forage Yield Based on Fall Precipitation in California Annual Grasslands.
J. Range Manag. 1970, 23, 363–365. [CrossRef]
16. Woodmansee, R.G. Critique and Analyses of the Grassland Ecosystem Model ELM. In Grassland Simulation
Model; Innis, G.S., Ed.; Springer: New York, NY, USA, 1978; pp. 257–281. ISBN 978-1-4612-9929-5.
17. Wright, J.; Skiles, J. SPUR Simulation of Production and Utilization of Rangelands: Documentation and User Guide;
USDA-Agricultural Research Service: Washington, DC, USA, 1987.
18. Pierson, F.B.; Carlson, D.H.; Spaeth, K.E. A Process-based Hydrology Submodel Dynamically Linked to the
Plant Component of the Simulation of Production and Utilization on Rangelands SPUR Model. Ecol. Model.
2001, 141, 241–260. [CrossRef]
19. Stout, W.L.; Vona, L.C.; Skiles, J.W.; Shaffer, J.A.; Jung, G.A.; Reid, R.L. Evaluating SPUR Model for Predicting
Animal Gains and Biomass on Eastern Hill Land Pastures. Agric. Syst. 1990, 34, 169–178. [CrossRef]
20. Corson, M.S.; Skinner, R.H.; Rotz, C.A. Modification of the SPUR Rangeland Model to Simulate Species
Composition and Pasture Productivity in Humid Temperate Regions. Agric. Syst. 2006, 87, 169–191.
[CrossRef]
21. Monteith, J.L. Solar Radiation and Productivity in Tropical Ecosystems. J. Appl. Ecol. 1972, 9, 747. [CrossRef]
22. Potter, C.S.; Randerson, J.T.; Field, C.B.; Matson, P.A.; Vitousek, P.M.; Mooney, H.A.; Klooster, S.A. Terresrial
Ecosystem Production: A Process Model Based on Global Satellite and Surface Data. Glob. Biogeochem. Cycles
1993, 7, 811–841. [CrossRef]
23. Li, S. Monitoring of Net Primary Production in California Rangelands Using Landsat and MODIS Satellite
Remote Sensing. Nat. Resour. 2012, 3, 56–65. [CrossRef]
24. Goetz, S.J.; Prince, D.; Goward, N.; Thawley, M.M.; Small, J.; Johnston, A. Mapping Net Primary Production
and Related Biophysical Variables with Remote Sensing: Application to the BOREAS region. J. Geophys. Res.
1999, 104, 27719–27734. [CrossRef]
25. Grigera, G.; Oesterheld, M.; Pacín, F. Monitoring Forage Production for Farmers’ Decision Making.
Agric. Syst. 2007, 94, 637–648. [CrossRef]
Remote Sens. 2019, 11, 595 20 of 22

26. Running, S.W.; Thornton, P.E.; Nemani, R.; Glassy, J.M. Global Terrestrial Gross and Net Primary Productivity
from the Earth Observing System. In Methods in Ecosystem Science; Springer: New York, NY, USA, 2000;
pp. 44–57.
27. Turner, D.; Gower, S.; Cohen, W.; Gregory, M.; Maiersperger, T. Effects of Spatial Variability in Light Use
Efficiency on Satellite-based NPP Monitoring. Remote Sens. Environ. 2002, 80, 397–405. [CrossRef]
28. Lobell, D.B.; Hicke, J.A.; Asner, G.P.; Field, C.B.; Tucker, C.J.; Los, S.O. Satellite Estimates of Productivity and
Light Use Efficiency in United States Agriculture, 1982-98. Glob. Chang. Biol. 2002, 8, 722–735. [CrossRef]
29. Nestola, E.; Calfapietra, C.; Emmerton, C.A.; Wong, C.Y.S.; Thayer, D.R.; Gamon, J.A. Monitoring Grassland
Seasonal Carbon Dynamics, by Integrating MODIS NDVI, Proximal Optical Sampling, and Eddy Covariance
Measurements. Remote Sens. 2016, 8, 260. [CrossRef]
30. Hilker, T.; Coops, N.C.; Wulder, M.A.; Black, T.A.; Guy, R.D. The Use of Remote Sensing in Light Use
Efficiency Based Models of Gross Primary Production: A Review of Current Status and Future Requirements.
Sci. Total Environ. 2008, 404, 411–423. [CrossRef]
31. Yuan, W.; Liu, S.; Zhou, G.G.; Zhou, G.G.; Tieszen, L.L.; Baldocchi, D.; Bernhofer, C.; Gholz, H.;
Goldstein, A.H.; Goulden, M.L.; et al. Deriving a Light Use Efficiency Model from Eddy Covariance
Flux Data for Predicting Daily Gross Primary Production Across Biomes. Agric. For. Meteorol. 2007, 143,
189–207. [CrossRef]
32. Prince, S.D.; Goward, S.N. Global Primary Production: A Remote Sensing Approach. Source J. Biogeogr. 1995,
22, 815–835. [CrossRef]
33. Fisher, J.I.; Mustard, J.F.; Vadeboncoeur, M.A. Green Leaf Phenology at Landsat Resolution: Scaling from the
Field to the Satellite. Remote Sens. Environ. 2006, 100, 265–279. [CrossRef]
34. Gärtner, P.; Förster, M.; Kleinschmit, B. The Benefit of Synthetically Generated RapidEye and Landsat 8 data
Fusion Time Series for Riparian Forest Disturbance Monitoring. Remote Sens. Environ. 2016, 177, 237–247.
[CrossRef]
35. Hunt, E.R.; Everitt, J.H.; Ritchie, J.C.; Moran, M.S.; Booth, D.T.; Anderson, G.L.; Clark, P.E.; Seyfried, M.S.
Applications and Research Using Remote Sensing for Rangeland Management. Photogramm. Eng.
Remote Sens. 2003, 69, 675–693. [CrossRef]
36. Walker, J.J.; De Beurs, K.M.; Wynne, R.H. Dryland Vegetation Phenology across An Elevation Gradient in
Arizona, USA, Investigated with Fused MODIS and Landsat Data. Remote Sens. Environ. 2014, 144, 85–97.
[CrossRef]
37. Yu, N.; Li, L.; Schmitz, N.; Tian, L.F.; Greenberg, J.A.; Diers, B.W. Development of methods to improve
soybean yield estimation and predict plant maturity with an unmanned aerial vehicle based platform.
Remote Sens. Environ. 2016, 187, 91–101. [CrossRef]
38. Guan, S.; Fukami, K.; Matsunaka, H.; Okami, M.; Tanaka, R.; Nakano, H.; Sakai, T.; Nakano, K.; Ohdan, H.;
Takahashi, K.; et al. Assessing Correlation of High-Resolution NDVI with Fertilizer Application Level and
Yield of Rice and Wheat Crops using Small UAVs. Remote Sens. 2019, 11, 112. [CrossRef]
39. Yue, J.; Yang, G.; Li, C.; Li, Z.; Wang, Y.; Feng, H.; Xu, B.; Yue, J.; Yang, G.; Li, C.; et al. Estimation of Winter
Wheat Above-Ground Biomass Using Unmanned Aerial Vehicle-Based Snapshot Hyperspectral Sensor and
Crop Height Improved Models. Remote Sens. 2017, 9, 708. [CrossRef]
40. Näsi, R.; Viljanen, N.; Kaivosoja, J.; Alhonoja, K.; Hakala, T.; Markelin, L.; Honkavaara, E.; Näsi, R.;
Viljanen, N.; Kaivosoja, J.; et al. Estimating Biomass and Nitrogen Amount of Barley and Grass Using UAV
and Aircraft Based Spectral and Photogrammetric 3D Features. Remote Sens. 2018, 10, 1082. [CrossRef]
41. Brocks, S.; Bareth, G.; Brocks, S.; Bareth, G. Estimating Barley Biomass with Crop Surface Models from
Oblique RGB Imagery. Remote Sens. 2018, 10, 268. [CrossRef]
42. Sarron, J.; Malézieux, É.; Sané, C.; Faye, É.; Sarron, J.; Malézieux, É.; Sané, C.A.B.; Faye, É. Mango Yield
Mapping at the Orchard Scale Based on Tree Structure and Land Cover Assessed by UAV. Remote Sens. 2018,
10, 1900. [CrossRef]
43. Planet. Planet Imagery Product Specification: PlanetScope and RapidEye; Planet Labs Inc.: San Francisco, CA,
USA, 2018.
44. Bartolome, J.; Frost, W.; Mcdougald, N. Guidelines for Residual Dry Matter on Coastal and Foothill
Rangelands in California. Agric. Nat. Resour. Publ. 2002, 8092, 1–7.
45. Ball, G.H.; Hall, D.J. ISODATA, A Novel Method of Data Analysis and Pattern Classification; Stanford Research
Institute: Menlo Park, CA, USA, 1965.
Remote Sens. 2019, 11, 595 21 of 22

46. Pix4D Support Radiometric Corrections. Available online: https://support.pix4d.com/hc/en-us/articles/

202559509-Radiometric-corrections#%23label2 (accessed on 26 September 2018).
47. Teillet, P.M.; Guindon, B.; Goodenough, D.G. On the Slope-Aspect Correction of Multispectral Scanner Data.
Can. J. Remote Sens. 1982, 8, 84–106. [CrossRef]
48. Meyer, P.; Itten, K.I.; Kellenberger, T.; Sandmeier, S.; Sandmeier, R. Radiometric Corrections of
Topographically Induced Effects on Landsat TM Data in an Alpine Environment. ISPRS J. Photogramm.
Remote Sens. 1993, 48, 17–28. [CrossRef]
49. Soenen, S.A.; Peddle, D.R.; Coburn, C.A. SCS+C: A Modified Sun-canopy-sensor Topographic Correction in
Forested Terrain. IEEE Trans. Geosci. Remote Sens. 2005, 43, 2148–2159. [CrossRef]
50. Tan, B.; Masek, J.G.; Wolfe, R.; Gao, F.; Huang, C.; Vermote, E.F.; Sexton, J.O.; Ederer, G. Improved Forest
Change Detection with Terrain Illumination Corrected Landsat Images. Remote Sens. Environ. 2013, 136,
469–483. [CrossRef]
51. Gao, F.; Masek, J.; Schwaller, M.; Hall, F. On the Blending of The landsat and MODIS Surface Reflectance:
Predicting Daily Landsat Surface Reflectance. IEEE Trans. Geosci. Remote Sens. 2006, 44, 2207–2218. [CrossRef]
52. Paruelo, J.M.; Epstein, H.E.; Lauenroth, W.K.; Burke, I.C. ANPP Estimates from NDVI for the Central
Grassland Region of the United States ANPP. Ecology 1997, 78, 953–958. [CrossRef]
53. Anderson, G.; Hanson, J.; Haas, R. Evaluating Landsat Thematic Mapper Derived Vegetation Indices for
Estimating Above-Ground Biomass on Semiarid Rangelands. Remote Sens. Environ. 1993, 45, 165–175.
[CrossRef]
54. Moulin, S.; Kergoat, L.; Viovy, N.; Dedieu, G. Global-scale Assessment of Vegetation Phenology using
NOAA/AVHRR Satellite Measurements. J. Clim. 1997, 10, 1154–1170. [CrossRef]
55. Tucker, C.J.; Sellers, P.J. Satellite Remote Sensing of Primary Production. Int. J. Remote Sens. 1986, 7, 1395–1416.
[CrossRef]
56. Zhang, X.; Friedl, M.A.; Schaaf, C.B.; Strahler, A.H.; Hodges, J.C.F.; Gao, F.; Reed, B.C.; Huete, A. Monitoring
Vegetation Phenology using MODIS. Remote Sens. Environ. 2003, 84, 471–475. [CrossRef]
57. Sellers, P.J.; Los, S.O.; Tucker, C.J.; Justice, C.O.; Dazlich, D.A.; Collatz, G.J.; Randall, D.A. A Revised
Land Surface Parameterization (SiB2) Ffor Atmospheric GCMs. Part II: The Generation of Global Fields of
Terrestrial Biophysical Parameters from Datellite Data. J. Clim. 1996, 9, 706–737. [CrossRef]
58. Hart, Q.J.; Brugnach, M.; Temesgen, B.; Rueda, C.; Ustin, S.L.; Frame, K. Daily Reference Evapotranspiration
for California using Satellite Imagery and Weather Station Measurement Interpolation. Civ. Eng. Environ. Syst.
2009, 26, 19–33. [CrossRef]
59. Akitsu, T.; Kume, A.; Hirose, Y.; Ijima, O.; Nasahara, K.N. On the Stability of Radiometric Ratios of
Photosynthetically Active Radiation to Global Solar Radiation in Tsukuba, Japan. Agric. For. Meteorol.
2015, 209, 59–68. [CrossRef]
60. Blackburn, W.J.; Proctor, J.T.A. Estimating Photosynthetically Active Radiation from Measured Solar
Irradlance. Sol. Energy 1983, 3183, 233–234. [CrossRef]
61. Li, R.; Zhao, L.; Ding, Y.; Wang, S.; Ji, G.; Xiao, Y.; Liu, G.; Sun, L. Monthly Ratios of PAR to Global Solar
Radiation Measured at Northern Tibetan Plateau, China. Sol. Energy 2010, 84, 964–973. [CrossRef]
62. Papaioannou, G.; Papanikolaou, N.; Retalis, D. Relationships of Photosynthetically Active Radiation and
Shortwave Irradiance. Theor. Appl. Climatol. 1993, 48, 23–27. [CrossRef]
63. Tsubo, M.; Walker, S. Relationships between Photosynthetically Active Radiation and Clearness Index at
Bloemfontein, South Africa. Theor. Appl. Climatol. 2005, 80, 17–25. [CrossRef]
64. Frouin, R.; Pinker, R.T. Estimating Photosynthetically Active Radiation (PAR) at the Earth’s Surface from
Satellite Observations. Remote Sens. Environ. 1995, 51, 98–107. [CrossRef]
65. ArcGIS for Desktop Area Solar Radiation. Available online: http://desktop.arcgis.com/en/arcmap/10.3/
tools/spatial-analyst-toolbox/area-solar-radiation.htm (accessed on 11 March 2019).
66. Klute, A.; Cassel, D.K.; Nielsen, D.R. Field Capacity and Available Water Capacity. In Methods of Soil Analysis:
Part 1—Physical and Mineralogical Methods; Soil Science Society of America, American Society of Agronomy:
Madison, WI, USA, 1986; pp. 901–926.
67. Gibson, W.; Daly, C.; Kittel, T.; Nychka, D.; Johns, C.; Rosenbloom, N.; McNab, A.; Taylor, G. Development
of A 103-year high-resolution Climate Data Set for the Conterminous United States. In AMS Conference on
Applied Climatology; American Meteorological Society: Portland, OR, USA, 2002; pp. 181–183.
68. Robbins, H.; Monro, S. A Stochastic Approximation Method. Annu. Math. Stat. 1951, 22, 400–407. [CrossRef]
Remote Sens. 2019, 11, 595 22 of 22

69. Bottou, L. Stochastic gradient descent tricks. In Neural Networks: Tricks of the Trade; Springer:
Berlin/Heidelberg, Germany, 2012; pp. 421–436. ISBN 9783642352881.
70. Trishchenko, A.P.; Cihlar, J.; Li, Z. Effects of Spectral Response Function on Surface Reflectance and NDVI
Measured with Moderate Resolution Satellite Sensors. Remote Sens. Environ. 2002, 81, 1–18. [CrossRef]
71. Roy, D.P.; Kovalskyy, V.; Zhang, H.K.; Vermote, E.F.; Yan, L.; Kumar, S.S.; Egorov, A. Characterization of
Landsat-7 to Landsat-8 Reflective Wavelength and Normalized Difference Vegetation Index Continuity.
Remote Sens. Environ. 2016, 185, 57–70. [CrossRef]
72. Li, A.; Bian, J.; Lei, G.; Huang, C. Estimating the Maximal Light Use Efficiency for Different Vegetation
Through the CASA Model Combined with Time-series Remote Sensing Data and Ground Measurements.
Remote Sens. 2012, 4, 3857–3876. [CrossRef]
73. Paloscia, S.; Pettinato, S.; Santi, E.; Notarnicola, C.; Pasolli, L.; Reppucci, A. Soil Moisture Mapping using
Sentinel-1 Images: Algorithm and Rreliminary Validation. Remote Sens. Environ. 2013, 134, 234–248.
[CrossRef]
74. Wagner, W.; Sabel, D.; Doubkova, M.; Bartsch, A.; Pathe, C. The Potential of Sentinel-1 for Monitoring Soil
Moisture With a High Spatial Resolution At Global Scale. Earth Obs. Water Cycle Sci. 2009, 2009, 18–20.
75. Balenzano, A.; Mattia, F.; Satalino, G.; Pauwels, V.; Snoeij, P. SMOSAR Algorithm for Soil Moisture Retrival
using Sentinel-1 Data. In Proceedings of the2012 IEEE International Geoscience and Remote Sensing
Symposium (IGARSS), Munich, Germany, 22–27 July 2012; pp. 1200–1203.
76. Chaplin-Kramer, R.; George, M.R. Effects of Climate Change on Range Forage Production in the San Francisco
Bay Area. PLoS ONE 2013, 8, 1–11. [CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).