GIScience & Remote Sensing

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tgrs20

Exploring the potential of multi-source

unsupervised domain adaptation in crop mapping
using Sentinel-2 images

Yumiao Wang, Luwei Feng, Weiwei Sun, Zhou Zhang, Hanyu Zhang, Gang
Yang & Xiangchao Meng

To cite this article: Yumiao Wang, Luwei Feng, Weiwei Sun, Zhou Zhang, Hanyu Zhang, Gang
Yang & Xiangchao Meng (2022) Exploring the potential of multi-source unsupervised domain
adaptation in crop mapping using Sentinel-2 images, GIScience & Remote Sensing, 59:1,
2247-2265, DOI: 10.1080/15481603.2022.2156123

To link to this article: https://doi.org/10.1080/15481603.2022.2156123

© 2022 The Author(s). Published by Informa Published online: 12 Dec 2022.

UK Limited, trading as Taylor & Francis
Group.

Submit your article to this journal Article views: 908

View related articles View Crossmark data

Citing articles: 2 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tgrs20
GISCIENCE & REMOTE SENSING
2022, VOL. 59, NO. 1, 2247–2265
https://doi.org/10.1080/15481603.2022.2156123

RESEARCH ARTICLE

Exploring the potential of multi-source unsupervised domain adaptation in crop

mapping using Sentinel-2 images
Yumiao Wanga,b, Luwei Fengc, Weiwei Suna, Zhou Zhangd, Hanyu Zhange, Gang Yanga and Xiangchao Menga
a
Department of Geography and Spatial Information Techniques, Ningbo University, Ningbo, China; bKey Laboratory of Urban Land Resources
Monitoring and Simulation, Ministry of Natural Resources, Shenzhen, China; cSchool of Remote Sensing and Information Engineering, Wuhan
University, Wuhan, China; dBiological Systems Engineering, University of Wisconsin–Madison, Madison, WI, USA; eDepartment of Electrical and
Computer Engineering, University of California, Los Angeles, CA, USA

ABSTRACT ARTICLE HISTORY

Accurate crop mapping is critical for agricultural applications. Although studies have combined Received 27 July 2022
deep learning methods and time-series satellite images to crop classification with satisfactory Accepted 01 December 2022
results, most of them focused on supervised methods, which are usually applicable to a specific KEYWORDS
domain and lose their validity in new domains. Unsupervised domain adaptation (UDA) was Crop mapping; multi-source
proposed to solve this limitation by transferring knowledge from source domains with labeled unsupervised domain
samples to target domains with unlabeled samples. Particularly, multi-source UDA (MUDA) is adaptation; time-series
a powerful extension that leverages knowledge from multiple source domains and can achieve remote sensing; transfer
better results in the target domain than single-source UDA (SUDA). However, few studies have learning; deep learning
explored the potential of MUDA for crop mapping. This study proposed a MUDA crop classification
model (MUCCM) for unsupervised crop mapping. Specifically, 11 states in the U.S. were selected as
the multi-source domains, and three provinces in Northeast China were selected as individual
target domains. Ten spectral bands and five vegetation indexes were collected at a 10-day interval
from time-series Sentinel-2 images to build the MUCCM. Subsequently, a SUDA model Domain
Adversarial Neural Network (DANN) and two direct transfer methods, namely, the deep neural
network and random forest, were constructed and compared with the MUCCM. The results
indicated that the UDA models outperformed the direct transfer models significantly, and the
MUCCM was superior to the DANN, achieving the highest classification accuracy (OA>85%) in each
target domain. In addition, the MUCCM also performed best in in-season forecasting and crop
mapping. This study is the first to apply a MUDA to crop classification and demonstrate a novel,
effective solution for high-performance crop mapping in regions without labeled samples.

1. Introduction
become a mainstream technique for crop classifica­
The growing population and extreme climate have tion (Wang et al. 2021; Hao et al. 2015; Peña and
made food security a hot topic of concern worldwide Brenning 2015). Compared with a single remote sen­
(Bajželj et al. 2014; Hao et al. 2020). For example, the sing image, time-series remote sensing images has
global population is expected to increase from 7.8 billion the advantage of capturing the dynamic spectral fea­
today to 9.7 billion by 2050 (De Wrachien et al., 2021), tures of crops, leading to increased accuracy in classi­
resulting in enormous pressure on food production. fication. Thus, using time-series images during the
Crop mapping can reflect the distribution and area of crop growth period for crop mapping has been an
crops, which helps governments make assessments and important approach (Löw et al. 2013; Skakun et al.
formulate policies to avoid food shortages (Cai et al. 2017; Vuolo et al. 2018; Zhong, Lina, and Zhou 2019).
2018; Lobell 2013). Moreover, an accurate crop map is Notably, machine learning methods have been com­
indispensable in crop yield prediction, crop insurance, bined with time-series satellite images to improve the
and land rental (Bolton and Friedl 2013; Lobell et al. accuracy of crop classification. (Hao et al. 2015; Kumar
2015). Therefore, accurate crop mapping is crucial for et al. 2015; Zheng et al. 2015). For example, Feng et al.
both scientific and practical applications. (2019) compared machine learning methods and sta­
With the increasing availability of satellite remote tistical-based methods to distinguish rice, corn, and
sensing data, using satellite data to identify crops has soybean using time-series images, the results showed

CONTACT Weiwei Sun [email protected]; Zhou Zhang [email protected]

© 2022 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
2248 Y. WANG ET AL.

that the machine learning methods were up to 10% (UDA) was proposed to overcome this problem
more accurate than the statistical-based methods. because it can transfer knowledge learned from
Recently, deep learning methods have been widely source domains with labeled samples to target
used in crop mapping (Meng et al. 2021; Turkoglu domains with unlabeled data (Ganin et al. 2016). The
et al. 2021; Xu et al. 2021; Zhong, Lina, and Zhou core of UDA is aligning feature distribution between
2019; Zhou et al. 2019). As an advanced extension of the source domain and the target domain to achieve
machine learning, deep learning methods have dee­ knowledge transfer. Most UDA models focus on sin­
per networks than the former that can automatically gle-source UDA (SUDA) (Sun and Saenko 2016),
learn information features with multiple levels, namely, learning and transferring knowledge from
achieving a better performance (Lecun, Bengio, and a single source domain to a target domain. The
Hinton 2015; Schmidhuber 2015). For instance, Domain Adversarial Neural Network (DANN) is
Zhong, Lina, and Zhou (2019) compared two widely a widely used SUDA model in speech and image
used deep learning models and three non-deep recognition tasks. For example, Wang et al. (2018)
machine learning classifiers in identifying crops applied the DANN in speaking recognition and alle­
using the time-series of Enhanced Vegetation Index viated the domain shift between two datasets. Han
(EVI). The result indicated that the deep learning et al. (2019) proposed a DANN-based model to iden­
models outperformed the non-deep machine learn­ tify mechanical faults and achieved good generaliza­
ing classifiers significantly. tion capability in conditions that did not occur during
However, most of the existing deep learning-based training.
crop classification models are supervised models, In practice, there are probably multiple source
which rely on many ground truth samples in the domains that may contain different useful knowledge
study area. A viable solution to reduce the depen­ for the target domain. Therefore, multi-source UDA
dence on sample labels is transfer learning, which (MUDA) is proposed to leverage all available data. The
can use the knowledge from a source dataset to typical method to achieve MUDA is treating all source
improve the prediction in the target dataset (Pan domains as one domain and then aligning data dis­
and Yang 2009). Specifically, direct transfer is tribution as a SUDA does. Nevertheless, the domain
a widely used strategy. For example, studies have shift is difficult to remove even for one source and
attempted to directly transfer crop classification mod­ one target domain, and mismatch might increase
els trained in sample-rich regions to sample-poor when aligning the multi-source and target domains,
regions (Ge et al. 2021; Hao et al. 2020). However, resulting in unsatisfactory performance in the target
direct transfer models may lose their validity when domain (Ganin et al. 2016). Consequently, novel
applied to different areas owing to domain shift (Ge MUDA approaches have been proposed and can be
et al. 2021). To improve the model, researchers have categorized into two groups (Sun, Shi, and Yuanbin
applied the fine-tuning technique. Fine-tuning uses 2015): feature representation (Chattopadhyay et al.
a few samples from the target domain to retrain all or 2012; Guo, Shah, and Barzilay 2018) and
part of the structure of the pre-trained models from a combination of pre-trained classifiers (Guo,
the source domain. For example, Chew et al. (2020) Pasunuru, and Bansal 2020; Peng et al. 2019). The
pre-trained the VGG16 model by using the ImageNet first group designs a feature space projection in
dataset, and then the pre-trained model was retrained which the distribution divergence between the multi-
with crop samples for crop classification. Similarly, source domains and the target domain is minimized.
Suh et al. (2018) fine-tuned six pre-trained models The second group trains classifiers in the source
and obtained a good performance in classifying domains and combines the pre-trained classifiers to
sugar beet and volunteer potato. obtain a final classifier in the target domain. Notably,
Despite successes, fine-tuning still requires labeled the Multiple Feature Spaces Adaptation Network
samples from the target domain. Because sample (MFSAN) combines the advantages of both groups
collection is labor-intensive and time-consuming, and shows valuable capabilities in many image classi­
some agricultural areas, especially in developing and fication tasks (Zhu, Zhuang, and Wang 2019).
undeveloped counties, might not have ground-truth Specifically, the MFSAN first aligns the distributions
data for fine-tuning. Unsupervised domain adaptation of each pair of the source and target domain to learn
 GISCIENCE & REMOTE SENSING 2249

the domain-invariant representations. Next, the distribution and area information for these two crops
MFSAN aligns pre-trained classifiers to minimize the are publicly released and updated annually in the
discrepancy among all classifiers. The combination of U.S. but not in China. Therefore, this study focused
the two alignments has been proven to guarantee on corn and soybeans, with 11 states in the U.S. as
remarkable performance in the target domain. source domains and three provinces in China as tar­
To the best of our knowledge, few studies have get domains.
applied UDA approaches in crop mapping, especially Target domains: Three provinces in Northeast
MUDA approaches. Due to differences in climate con­ China were selected as individual target domains:
ditions, precipitation and soil types in multiple Liaoning (LN), Jilin (JL) and Heilongjiang (HLJ)
domains, the phenology of the same crop differs (Figure 1(b)). The three provinces produce more
among the domains (Wang et al. 2021), which inten­ than 30% and 45% of China’s annual corn and soy­
sifies the domain shift between the source domain bean production, respectively (MOA 2022). As shown
and target domain. Full use of the knowledge from in Figure 2, the phenology dates of the crops in the
multi-source domains is necessary to achieve accurate target domains are more focused than those in the
crop identification in the target domain. This study source domains. For example, the harvesting date for
proposed a MUDA crop classification model (MUCCM) soybean in the target domains is concentrated in
based on MFSAN for unsupervised crop mapping October, which is within the range of the harvesting
using time-series Sentinel-2 images. Specifically, we date for soybean in the source domains.
attempted to answer two research questions: 1) Is the
crop classification accuracy of the MUCCM better than
that of the widely used machine learning methods 2.2. Sentinel-2 imagery
and SUDA models? and 2) how does the MUCCM
perform in in-season forecasting and crop mapping? Sentinel-2 comprises two satellites, Sentinel-2 A and
Sentinel-2 B, which provide high spatial resolution (up
to 10 m) and temporal resolution (up to 5 days)
2. Study areas and dataset images freely. Because studies have proven the relia­
bility of top-of-atmosphere (TOA) reflectance on clas­
2.1. Study areas sification tasks (d’Andrimont et al. 2020; Ge et al. 2021;
The U.S. and China are the largest agricultural produ­ Song et al. 2001) and there is few available Sentinel-2
cers worldwide and have a substantial impact on Level-2A data in the Google Earth Engine (GEE) plat­
global food security (Veeck, Veeck, and Hongyan form, this study collected the TOA reflectance images
2020). Among the various crops, corn and soybean covering the study area in 2018 to obtain the spectral
are widely planted in the two countries. The spatial features of crops. Specifically, 10 bands including Red,

Figure 1. Study area. (a) Source domains: 11 states in the U.S.; (b) Target domains: three provinces in north China. Source domains: 11
states in the U.S. were selected as multi-source domains: North Dakota (ND), Minnesota (MN), Wisconsin (WI), South Dakota (SD), Iowa
(IA), Nebraska (NE), Kansas (KS), Missouri (MO), Illinois (IL), Indiana (IN) and Ohio (OH) (Figure 1(a)). Notably, the 11 states produce
approximately 70% of the annual corn and soybean in the U.S. (USDA 2020). In addition, the planting and harvesting dates of the two
crops vary across states. For example, the planting dates of corn in the 11 states ranging from mid-April to mid-Jun (Figure 2),
providing an opportunity to assess the effect of multi-source domains.
2250 Y. WANG ET AL.

Figure 2. Range of planting and harvesting dates for corn and soybean in source and target domains.

Blue, Green, Red-edge1, Red-edge2, Red-edge3, Near- a 10-day time window; 2) the clouds of the images
Infrared (NIR), Red-edge4, Short wave infrared (SWIR) were masked by the adjusted cloud score algorithm; 3)
1 and SWIR2 were selected as spectral features. We 10-day composites were generated with the median
also selected five widely used vegetation indexes (VIs) values of the valid observations; 4) data gaps were
calculated from the Sentinel-2 image: the Normalized filled by the linear interpolation to achieve full cover­
Difference Vegetation Index (NDVI) (Tucker 1979), age throughout the temporal domain; 5) 10-day time
Green Vegetation Index (VIgreen) (Gitelson et al. series data were smoothed by using the Savitzky-
2002), Enhanced Vegetation Index (EVI) (Huete et al. Golay (SG) filter. To maintain the same spatial resolu­
2002), Normalized Difference Water Index (NDWI) tion, we resampled all images to 10 m. The image
(Gao 1996), and Normalized Difference Residue collection, preprocessing, and information extraction
Index (NDRI) (Gelder, Kaleita, and Cruse 2009). In addi­ were all processed on the GEE platform (Gorelick et al.
tion, the cloud mask of the images was conducted 2017).
refer to Oreopoulos, Wilson, and Tamás (2011).

3. Methodology
2.3. Ground reference
3.1. Multiple feature spaces adaptation network
The ground reference data of the source domains
were collected from the Cropland Data Layer (CDL), Multiple Feature Spaces Adaptation Network (MFSAN)
an annual land cover map with 30 m spatial resolution is an advanced MUDA model proposed by Zhu (Zhu,
depicting the distribution of more than 100 crops Zhuang, and Wang 2019). Unlike other models, the
across the U.S. (Boryan et al. 2011). For the samples core idea of this model is to extract domain-invariant
of target domains, no available ground reference data representations for each pair of source and target
has been officially released. Fortunately, You et al. domains separately, instead of extracting common
(2021) published 10 m spatial resolution maps of domain-invariant representations for all domains,
corn and soybean in Northeast China from 2017 to which makes the model perform satisfactorily when
2019. Thus, we regarded the 2018 crop maps from there are significant differences among multiple
CDL and You as the reference crop maps in this study. source domain data.
Because this study focused on soybean and corn, As shown in Figure 3, the MFSAN has a common
the other types were labeled as “other” (e.g. grass, feature extractor (CFE), several domain-specific fea­
water, buildings, and other crops). To obtain balanced ture extractors (DSFE), and several domain-specific
samples (Pott et al. 2021), we randomly selected 3000 classifiers (DSC). Specifically, the CFE maps all data
samples of corn, soybean, and other type in each from source domains and the target domain into
domain, for a total of 126,000 labeled samples. For a common feature space. DSFEs are built to map
each sample, 10 spectral bands and five VIs were the feature maps of each pair of target data and
extracted at a 10-day interval from DOY 90 to DOY source data to obtain the domain-invariant repre­
300, resulting in 315 features. Specifically, referring to sentations. To estimate the alignment of the two
You et al. (2021), five steps were implemented to domains, the Maximum Mean Difference (MMD)
generate the dataset: 1) all available images covering was adopted to calculate the discrepancy (Gretton
the whole study area in 2021 were collected within et al. 2012), and its formula is as follows:
 GISCIENCE & REMOTE SENSING 2251

Figure 3. Framework of the MFSAN. The framework receives samples with annotated ground truth from N source domains and adapts
to classifying the samples in the target domain. CEF is a common feature extractor for all source domains. DSFE 1-N are domain-
specific feature extractors and DSC 1-N are domain-specific classifiers.

1 Xns 1 Xnt � where λ and γ represent the weight of MMD loss and
MMD½X; Y � ¼ jj f ðxi Þ f yj jjH (1)
ns i¼1 nt j¼1 discrepancy loss, respectively.
where f refers to the mapping function, and
H represents the reproducing Hilbert space. The 3.2. MUDA based crop classification model
source domain data and the target domain data
were regarded as two random variables: X and Y, Although the original MFSAN has been successfully
where X contains ns samples, and Y contains nt applied in RGB image classification tasks, it cannot be
samples. directly implemented for crop classification in this
For each source domain, a DSC was designed, study. Specifically, the MFSAN is designed for hand­
which received the features from the DSFE and ling RGB image patches with 256 rows and 256 col­
ended with a Softmax layer to predict the labeled umns, and our dataset is pixel-based with 315
result. Intuitively, different classifiers should output features. Thus, we modified the shape of our dataset
the same results for a target sample. Hence, minimiz­ and the architecture of the MFSAN model to build
ing the discrepancy among the prediction results was a new MUDA based crop classification model
necessary. The discrepancy value among all pairs of (MUCCM).
prediction results was calculated as follows: First, the 315 features were transformed into
a matrix with 15 rows (spectral dimension) and 21
2 XN XN
Ldisc ¼ E columns (time dimension) to satisfy the input require­
N � ðN 1Þ j¼1 i¼jþ1 y,Y
�� �� � ments of convolution operations in the MUCCM
�DSCi ðDSFEi ðF ð yÞÞÞ DSCj DSFEj ðF ð yÞÞ � (2) model. Second, the first seven ResBlocks of ResNet-
where F refers to the CFE. 50 (He et al. 2016) were selected as the CFE in the
The total loss of the entire network contained the MUCCM. ResNet-50 is a widely used model in classifi­
MMD loss (Lcls ), classification loss (Lmmd ) and discre­ cation tasks and can extract useful high-level features
pancy loss (Ldisc ), and the loss expression is shown as from complex inputs. The MFSAN takes the whole
Formula 3. Specifically, minimizing MMD loss will ResNet-50 as the CFE, which is complex and will
make the distributions of the source domain data reduce the input to one-eighth of the original. Thus,
and the target data close, reducing classification loss it needs to be simplified to process our dataset.
will improve the performance of each classifier, and Finally, we designed 11 DSFEs and 11 DSCs in the
discrepancy loss reduction will make the prediction MUCCM according to the number of the source
results of N classifiers similar. After minimizing the domains in this study. Specifically, the DSFE consists
loss, the final prediction result of the target data was of three convolution layers, each of which is con­
regarded as the mean of the classifiers’ predictions. nected to a batch normalization (BatchNorm) layer
and an activation layer at the end. The BatchNorm
Lall ¼ Lcls þ λLmmd þ γLdisc (3) layer was adopted to solve internal covariate shift and
2252 Y. WANG ET AL.

overfitting (Ma et al. 2021), and the Rectified Linear process, the number of epochs and the batch size
Unit (ReLU) was used as the activation layer. For the were set to 150 and 512, respectively. We adapted
DSC, we used a Dense layer to predict crop types. the parameters of the MUCCM model according to
Figure 4 shows the detailed architecture of the the result of the validation. Finally, Adaptive Moment
MUCCM. Estimation (Adam) was set to the optimizer, and the
learning rates of the CFE, DSFE, and DSC were
initiated as 0.001, 0.01, and 0.01, respectively. The
3.4. Experimental setup
hyperparameters λ and γ of the MUCCM were initia­
The 11 states in the U.S. were treated as multi-source lized to 1.
domains, and each of the three provinces in Northeast To demonstrate the advantages of the MUDA
China was selected as an individual target domain. model in crop mapping, we compared the model
Samples from the multi-source domains were split with three other models. The first model was DANN,
into two parts randomly: 70% for training the one of the most widely used SUDA methods. For an
MUCCM model and 30% for validation. Samples accurate comparison, the DANN was designed to
from each target domain were selected as the testing have a similar architecture to that of the MUCCM.
dataset to evaluate the MUCCM. In the training The difference between the two models was that

Figure 4. Architecture of the MUCCM. The modules in blue show the data flows in the MUCCM, and the detailed composition of each
sub-module is depicted in gray.
 GISCIENCE & REMOTE SENSING 2253

the DANN had merely one DSFE and one DSC. To train padded by 0 to preserve the original input size. This
the DANN, data from 11 states in the U.S. were com­ experiment allowed us to assess not only the in-
bined as one single source, and the other training season forecasting capability of the model but also
settings were identical to those of the MUCCM. To the robustness under different data deficiencies. The
compare the domain adaptation models, we mapping results are the best validation of crop classi­
designed two direct transfer learning models: includ­ fication models. Because the area of the three target
ing a deep neural network (DNN) model and domains is too large, we selected one county in each
a random forest (RF) model. The structure of the target domain to evaluate the models: Keshan County
DNN was almost identical to that of the DANN, except in HLJ, Yushu County in JL, and Pulandian County in
that the DNN had no feature alignment. The critical LN. We also selected a small area in each county to
parameters of RF were the number of trees and the verify the details of the maps generated by the mod­
maximum depth of each tree, and they were set to els. The locations of the selected counties and areas
400 and 20, respectively. Specifically, the deep learn­ are shown in Figure 5.
ing models were constructed by PyTorch 1.5 in
Python version 3.7, and the RF model was built by
scikit-learn, a Python module for machine learning. 4. Results and analysis
The MFSAN and the comparison models were eval­
4.1. Spectral characteristics of crops
uated based on three aspects: classification accuracy,
in-season forecasting capability and mapping perfor­ The NDVI time-series profiles for corn, soybean, and
mance. Specifically, the overall accuracy, macro- the other type in the 11 source domains and three
average of F1 score (F1) and confusion matrix were target domains are shown in Figure 6. The NDVI pro­
adopted as accuracy metrics. For evaluating the in- files varied by crop and domain. Notably, the other
season forecasting capability of the models, a fixed type displayed significantly different profiles from the
starting date was set to DOY 90 (planting date), and two crops in each domain. The reason for this finding
the ending date was varied from DOY 150 (early was that this category contained many land types
growing season) to DOY 300 (harvesting date). We except corn and soybean, such as grasses, forests,
used full time-series features to train each model and buildings, resulting in a large variance in the
and then tested the model using time-series features spectrum. The NDVI profiles for corn and soybean
of different periods (e.g. from DOY 90 to DOY 210). were similar, but slight differences in their peak mag­
Notably, the reduced time-series features were nitude and peak date were observed. For example,

Figure 5. Location of the counties and areas selected for testing mapping performance in the target domains.
2254 Y. WANG ET AL.

Figure 6. Time-series NDVI profiles of crops in different domains. Lines refer to average values, and buffers represent one standard
deviation from the average values.

the mean value of the peak NDVI for corn in KS and LN source domains to develop a high-quality unsuper­
was below 0.6, while most corn in SD and HLJ had vised transfer learning model is necessary.
a peak NDVI above 0.7. The NDVI peak dates for
soybean in ND and JL were very close to each other,
4.2. Accuracy comparison
concentrated at 220 DOY, while the NDVI peak date in
the target domain LN was more than 20 days later. Table 1 shows the classification accuracies of different
However, different spectral characteristics of the same models. The MUCCM was superior to the other mod­
crop can confuse the crop classification models and els in each target domain, especially in JL (87.52% OA
lead to poor accuracy. Notably, a model trained by the and 87.54% F1). The DANN was the second-best
data from a single source domain that differs too model, outperforming the DNN and RF, demonstrat­
much from the target domain may have low migra­ ing the efficacy of the domain adaptation. The accura­
tion performance. Therefore, using data from multiple cies of DNN in the target domains were significantly

Table 1. Classification accuracies of different models in target domains.

Target Metrics(%) MUCCM DANN DNN RF
HLJ OA 86.40 84.70 79.03 82.30
FI 86.46 84.69 78.98 82.63
JL OA 87.52 84.70 77.22 80.79
FI 87.54 84.69 77.12 81.11
LN OA 85.86 84.32 79.30 83.19
FI 85.90 84.32 79.32 83.36
 GISCIENCE & REMOTE SENSING 2255

lower than those of the other models, with OA less correct samples (above 83.33% OA) for each class in
than 80%, even less than that of RF. The reason for this each target domain. DANN identified slightly fewer
may be that deep learning models learn abstract and correct samples than the MUCCM did, but signifi­
specific laws that may not be directly applicable to cantly more than DNN and RF did, indicating that
new domains. By contrast, RF is an ensemble learning the domain adaptive models had good generalization
model with multiple decision trees to ensure its performance in target domains. However, all models
robust performance in different scenarios. exhibited mixed partitioning between corn and soy­
For further comparison, we calculated the confu­ beans. For example, the MUCCM classified 326 corn
sion matrix for each model in each target domain samples as soybeans and 303 soybeans samples as
(Figure 7). In general, the MUCCM exhibited the best corn in HLJ (Figure 7, (a1)). The RF model identified
identification performance, with more than 2500 440 corn samples as soybean and 606 soybean

Figure 7. Confusion matrix of (1) MUCCM, (2) DANN, (3) DNN, and (4) RF in (a) HLJ, (b) JL, and (c) LN.
2256 Y. WANG ET AL.

Table 2. Classification accuracies of different models in target domains with unbalanced

samples.
Target Metrics(%) MUCCM DANN DNN RF
HLJ OA 88.05 86.85 83.92 84.45
FI 83.36 81.36 78.17 79.96
JL OA 86.58 85.37 81.20 81.65
FI 81.38 79.77 74.27 75.19
LN OA 85.07 82.65 77.41 81.32
FI 75.94 72.30 66.24 69.89

samples as corn in JL (Figure 7, (b4)). These results October. Although the final accuracy of the RF
mainly occurred because the phenological informa­ model is not as good as that of the MUCCM and
tion of corn and soybean are similar such that they DANN, it achieved good accuracy (OA>75%) earlier
cannot be accurately distinguished even in the source than those of the two models. In addition, the perfor­
domains, and the severity of this problem increases in mance of RF is stable. For example, the accuracy
the target domain. curves of the deep learning models fluctuated sub­
Due to the unbalanced proportion of crops in the stantially and did not achieve accurate classification in
real environment, we collected an additional 6000 the early growing season in JL, while the accuracy of
random samples in each target domain to evaluate the RF model increased steadily over time. A possible
the models. The accuracy metrics and confusion reason for this finding is that deep learning models
matrix of the models are shown in Table 2 and are sensitive to inputs, and data distributed differently
Figure 8, respectively. The MUCCM model obtained from the training pattern can lead to model
the best OA and F1 in all target domains, followed by instability.
DANN, RF and DNN, which is consistent with the We also retrained and tested the MUCCM and RF in
results of the balanced samples. Figure 8 demon­ each target domain using the time-series features
strates a very unbalanced set of samples. The other between the starting date and different ending
type accounted for the largest proportion, followed dates. As shown in Figure 10, the retrained MUCCM
by corn, and soybean had the smallest proportion. model improved steadily over time, proving that
Although the MUCCM almost achieved the highest retraining is necessary for deep learning models. The
performance for each type in each target domain, it accuracy of the retrained MUCCM converged at
did not exhibit an absolute advantage over the DANN approximately DOY210, namely, three months before
in the other type. This is probably because the other harvest. In addition, the accuracy of the MUCCM was
type contains a variety of crops, some of which may better than that of the RF at almost every time point,
be similar to corn or soybeans, which makes classifi­ indicating that the MUCCM had advantages in early-
cation difficult for all the models. season crop classification in target domains.

4.3. In-season forecasting capability 4.4. Crop mapping performance

Early-season crop mapping is vital for agricultural The mapping results are the best validation of crop
applications. Thus, we also evaluated the model per­ classification models and Figure 11 shows the refer­
formance within the growing season. Specifically, the ence crop maps and the predicted crop maps gener­
models were trained using full time-series features ated by the models (trained using full time-series
and tested by the features between the starting date features) in the three selected counties. In Keshan
(DOY90) and different ending dates (DOY150 – County, the MUCCM and DANN generated results
DOY300). As shown in Figure 9, the three deep learn­ similar to the reference maps. By contrast, the DNN
ing models exhibited satisfactory performance with model misclassified many corn areas and the RF
more time-series features until the end of September, model identified the other type as soybean in north­
and then their accuracies stabilized. The MUCCM out­ ern Keshan County. However, all four crop classifica­
performed the other two deep learning models and tion models did not generate accurate crop maps in
achieved good performance (OA >85%) before Yushu and Pulandian. Although the MUCCM
 GISCIENCE & REMOTE SENSING 2257

Figure 8. Confusion matrix of (1) MUCCM, (2) DANN, (3) DNN, and (4) RF with unbalanced samples in (a) HLJ, (b) JL, and (c) LN.

outperformed the other models, the model- (Figure 12). In Keshan County, the maps generated by
generated crop maps misclassified many soybean the MUCCM and DANN depicted crop fields precisely.
fields as corn fields. The reason why the crop maps The DNN reversed the corn fields and soybean fields,
in Yushu and Pulandian seem inaccurate may be two­ and the RF produced serious “pepper noise” in each
fold: the model does have the limitation in mixing crop field. In Yushu County, the reference map showed
corn and soybeans, and this limitation is not spatially many other lands in the central region of the selected
consistent, and the crops in Yushu and Pulandian are area, but the DANN, DNN, and RF misclassified these
almost all corn, making visualizing which areas are lands as corn. Only the MUCCM correctly classified the
misclassified as soybeans easy. other type, indicating its high-quality performance in
To demonstrate the details of the crop maps, we actual mapping. The reference map of Pulandian
zoomed in on selected areas of the three counties showed that most of the land in the selected area was
2258 Y. WANG ET AL.

Figure 9. Performance against timelines of the models trained Figure 10. Performance against timelines of the models trained
using full time-series features in (a) HLJ, (b) JL, and (c) LN. using dynamic time-series features in (a) HLJ, (b) JL, and (c) LN.

planted with corn. However, all models incorrectly iden­ performance with reduced time-series features.
tified a substantial amount of corn as soybeans. The Figure 9 indicates that the accuracy of each model
average misclassification ratio (corn to soybean pixels/ did not increase consistently with the increasing num­
all pixels) of the MUCCUM, DANN, DNN, and RF in three ber of time series features. Compared to the RF
areas was 3.02%, 4.90%, 13.06%, and 7.51%. Maps gen­ model, the accuracy curves of the deep learning mod­
erated by the MUCCM model were closer to the refer­ els were more fluctuating. This could be due to the
ence map than those generated by other models. more complex structures of the deep learning mod­
Specifically, the MUCCM also produced fewer errors els. A deep learning model consists of multiple layers,
than the DANN, illustrating the efficacy of multi-source each of which consists of numerous neurons.
domains. Changing the model input may have an unpredict­
able effect on the model output. Therefore, the gra­
dual increase of time-series features cannot produce
5. Discussion a smooth increase in the accuracy of the deep learn­
ing models. The second experiment used time-series
5.1. Model interpretation
features of different lengths to retrain and test the
This study conducted two experiments to discover models, and the results (Figure 10) indicated that the
the performance of the models against timelines in models improved steadily over time until the later
target domains. The first one aimed to test if models stage in the season, and the deep learning model
trained by full time-series features can keep good significantly outperformed the RF model for each
 GISCIENCE & REMOTE SENSING 2259

Figure 11. Comparison of the maps of (a) Keshan County, (b) Yushu County, and (c) Pulandian County from (1) reference crop maps,
(2) MUCCM, (3) DANN, (4) DNN, and (5) RF.

feature combination. These two experiments demon­ the original features of corn and soybean and the
strated the importance of data pattern alignment extracted features from the crop classification models.
between the training phase and the testing phase, Because there were no extracted features from RF, we
and our model can keep the best performance in each only applied the algorithm for the MUCCM, DANN, and
alignment situation. DNN. The visualization results (Figure 13) indicated that
To explore how models classify crops, we applied the distributions of the original features of the two
the t-distributed Stochastic Embedding (t-SNE) algo­ crops were disordered and chaotic. After processing
rithm (Van der Maaten and Hinton 2008) to visualize through the models, the two distributions of the
2260 Y. WANG ET AL.

Figure 12. Crop maps of the selected areas in (a) Keshan County, (b) Yushu County, and (c) Pulandian County from (1) reference crop
maps, (2) MUCCM, (3) DANN, (4) DNN, and (5) RF.

extracted features exhibited varying degrees of sepa­ the MUCCM formed two distinct aggregates, with each
rateness. Compared with the domain adaptive models, representing one crop and containing fewer wrong
the DNN generated two distributions with weak separ­ samples than those of DANN. However, although
ability, which explained its low classification accuracy. there were errors in the two distributions of the
The two distributions from DANN differ, but each cate­ DANN both were more aggregated than those of the
gory contained a substantial number of samples from MUCCM. Two reasons may explain this finding. First,
the other. Moreover, the two distributions formed the DSFEs of MUCCM had only one-eleventh of the
three aggregates, increasing the difficulty of crop iden­ training data and epoch of DANN, and the inadequate
tification. The distributions of extracted features from training conditions can make the distribution difficult
 GISCIENCE & REMOTE SENSING 2261

Figure 13. T -SNE visualization results of (1) original features and extracted features from (2) MUCCM, (3) DANN, and (4) DNN in (a) HLJ,
(b) JL, and (c) LN.

to aggregate. Second, the result of the MUCCM was have an advantage over single-source models, we
obtained by averaging the results of 11 extractors, retrained DANN, DNN, and RF using samples from
which may also increase the sparseness of the the 11 source domains. For an accurate comparison,
distribution. the samples for the single-source model were
expanded to as many as that for the multi-source
model, namely, 99,000 samples. Table 3 shows that
5.2. Multi-source model versus single-source model
the performance of the single-source models trained
This study used data from multi-source domains to in different source domains varied in the same target
train DANN, DNN and RF (called multi-source models). domain. In the case of HLJ, for example, the accuracy
However, these models are usually trained by data of the models trained in ND was above 80%, and the
from a single-source domain (called single-source accuracy of the models trained in SD decreased sig­
models). To discover whether multi-source models nificantly. Different performances may be caused by
2262 Y. WANG ET AL.

Table 3. OA (%) in target domains of the models trained using samples from a single source domain.
HLJ JL LN
Source Domain DANN DNN RF DANN DNN RF DANN DNN RF
ND 87.86 83.05 85.13 86.41 79.07 83.37 85.79 76.68 81.23
SD 85.12 76.29 77.68 83.14 74.89 79.84 85.52 75.53 80.50
KS 85.79 76.24 78.39 87.30 70.05 70.62 85.41 73.62 78.72
NE 85.08 78.18 78.76 86.54 76.85 79.61 85.10 79.99 84.41
MN 85.53 83.88 84.17 86.68 81.42 75.86 83.78 78.04 73.40
IA 85.51 77.45 79.31 83.49 76.31 71.26 82.44 75.68 71.27
WI 86.31 76.79 80.76 87.14 78.73 79.56 86.2 78.39 73.94
IL 86.75 76.75 69.54 88.51 79.68 64.89 85.99 79.64 69.64
IN 86.93 74.36 76.88 88.31 75.26 76.73 87.21 77.05 74.48
OH 87.06 78.66 77.88 86.84 79.68 78.47 86.55 75.33 70.82
MO 84.22 62.69 54.87 81.17 60.87 56.43 82.51 64.21 60.70

Table 4. OA (%) of the DANN and MUCCM in target domains. The domain. However, there are several limitations. First,
OA of the single-source DANN is the average result of the 11 this study focused on soybean and corn, and the other
DANN models in Table 3, and the OAs of the multi-source DANN
and MUCCM are the same as those of DANN and MUCCM in crop types are combined into one “other” class.
Table 1. However, to improve transfer accuracy, the crop species
Target Domain in the source domain and target domain should be
Model HLJ JL LN similar, and each crop should be sampled for modeling.
Single-source DANN 86.01 85.96 85.14 The purpose of this study is to verify the advance of the
Multi-source DANN 84.7 84.7 84.32
MUCCM 86.4 87.52 85.86 proposed model, and the results indicated that the
proposed model has a more generalized transfer perfor­
mance than other models under the same experimental
different data distributions in target domains and environment. In practical applications, to make the
source domains. For instance, time-series NDVI pro­ transfer results more accurate, we should ensure that
files of the crops in JL (Figure 6(m)) and OH (Figure 6 the crop species of the source and target domains are
(j)) were more similar than those of MO (Figure 6(k)), consistent, and divide the “other” type into specific
thus the classification accuracy of the model trained crops so that the accuracy will be higher. Second, the
in OH was high in JL. resulting maps generated by models were evaluated
The performance of the single-source DNN and RF based on the 2018 existing crop map. Because the
was weaker than that of the MUCCM, but some single- existing map is not 100% correct, the accuracy of the
source DANN models (such as from ND to HLJ) exhibited resulting maps can only be judged from the general
good accuracy. To improve the evaluation of the trend, and determining whether a mismatch is an error
MUCCM and DANN, Table 4 shows the accuracies of in the resulting maps or the existing map is not possible.
the multi-source DANN and MUCCM, and the average Third, this study only considered two main crops in the
accuracy of the 11 single-source DANNs. The results U.S. and China, while there are various crops in the two
demonstrated that although MUCCM was not better counties. In future work, we will explore the transfer
than every single-source DANN, it was better than the performance of the proposed model in heterogeneous
average DANN and the multi-source DANN. This finding regions with multiple crops. Although the proposed
is of practical significance because without the sample model has a flexible structure that can easily be adapted
labels of the target domain, we cannot determine which to multiple crops, the model’s performance may
source domain can be best migrated to the target degrade because crop mapping in heterogeneous
domain, whereas MUCCM does not need to consider regions is very challenging (Zhang et al. 2021; Chen
this problem and can achieve good performance. et al. 2016). Thus, we will optimize the model by 1)
adding new features, such as backscattering coefficients
from Sentinel-1 image, to enhance the separability
5.3. Limitations and future work
among multiple crops. 2) selecting time-series features
This study illustrated that the MUCCM model used the based on crop phenology to improve the alignment
knowledge of the multi-source domains to achieve between the source and target domains. 3) embedding
accurate unsupervised crop mapping in the target attention modules into the model to classify crops more
 GISCIENCE & REMOTE SENSING 2263

effectively. Finally, the study areas have close latitudes and Simulation (No. KF-2021-06-089), the Public Projects of
and seasons, providing the possibility for a good trans­ Ningbo City (No. 2022S101), Ningbo Science and Technology
fer. However, whether the present method can maintain Innovation 2025 Major Special Project (No. 2021Z107), and the
National Natural Science Foundation of China (No. 42201354,
good transfer performance in two regions with large
42171326).
seasonal differences requires further research. In light
of the previous study (Zhao et al. 2020), a potential idea
to improve the transfer between two distinct regions is References
to use prior phenological knowledge to extract the
spectral features of crops at key phenological stages Bajželj, B., K. S. Richards, J. M. Allwood, J. S. Pete Smith,
E. C. Dennis, and C. A. Gilligan. 2014. “Importance of
(e.g. sowing and flowing stages) in the two regions
Food-Demand Management for Climate Mitigation.”
separately and then use these features as the input to Nature.Com.
the transfer model. This method can avoid the feature Bolton, D. K., and M. A. Friedl. 2013. “Forecasting Crop Yield
shift caused by seasons and increase the transfer Using Remotely Sensed Vegetation Indices and Crop
performance. Phenology Metrics.” Agricultural and Forest Meteorology
173: 74–84. doi:10.1016/j.agrformet.2013.01.007.
Boryan, C., Z. Yang, R. Mueller, and M. Craig. 2011. “Monitoring
6. Conclusion US Agriculture: The US Department of Agriculture, National
Agricultural Statistics Service, Cropland Data Layer
This study proposed the MUCCM to explore the Program.” Geocarto International 26 (5): 341–358. doi:10.
potential of MUDA in crop mapping. Specifically, 11 1080/10106049.2011.562309.
states in the U.S. were selected as the multi-source Cai, Y., K. Guan, J. Peng, S. Wang, C. Seifert, B. Wardlow, and
L. Zhan. 2018. “A High-Performance and in-Season
domains, and three provinces in Northeast China
Classification System of Field-Level Crop Types Using
were selected as individual target domains. Ten spec­ Time-Series Landsat Data and A Machine Learning
tral bands and five vegetation indexes were collected Approach.” Remote Sensing of Environment 210: 35–47.
from time-series satellite imagery. To evaluate the doi:10.1016/j.rse.2018.02.045.
MUCCM model, we compared it with the SUDA Chattopadhyay, R., Q. Sun, W. Fan, I. Davidson,
model, DANN, and two direct transfer models: DNN S. Panchanathan, and Y. Jieping. 2012. “Multisource
Domain Adaptation and Its Application to Early Detection
and RF. The results showed that the MUCCM model
of Fatigue.” ACM Transactions on Knowledge Discovery from
achieved good performance in each target domain, Data (TKDD) 6 (4): 1–26. doi:10.1145/2382577.2382582.
especially for in-season forecasting and actual map­ Chen, Y., X. Song, S. Wang, J. Huang, and L. R. Mansaray. 2016.
ping. Furthermore, we employed t-SNE to interpret “Impacts of Spatial Heterogeneity on Crop Area Mapping in
the models and found that the corn and soybean Canada Using MODIS Data.” ISPRS J. Photogramm. Remote
features extracted by the MUCCM had the best Sens 119: 451–461. doi:10.1016/j.isprsjprs.2016.07.007.
Chew, R., J. Rineer, R. Beach, M. O’Neil, N. Ujeneza, D. Lapidus,
separation. Finally, we compared the multi-source
T. Miano, M. Hegarty-Craver, J. Polly, and D. S. Temple. 2020.
models and single-source models, and the results “Deep Neural Networks and Transfer Learning for Food Crop
indicated that the performances of the single-source Identification in UAV Images.” Drones 4 (1): 7. doi:10.3390/
models were very unstable, and that the MUCCM can drones4010007.
exhibit good performance by using multi-source data. d’Andrimont, R., M. Taymans, G. Lemoine, A. Ceglar,
This study demonstrated the great potential of apply­ M. Yordanov, and M. van der Velde. 2020. “Detecting
Flowering Phenology in Oil Seed Rape Parcels with
ing the MUDA model to unsupervised crop mapping.
Sentinel-1 and −2 Time Series.” Remote Sensing of
Environment 239: 111660. doi:10.1016/j.rse.2020.111660.
De Wrachien, D., B. Schultz, and M. B. Goli. 2021. “Impacts of
Disclosure statement
Population Growth and Climate Change on Food Production
No potential conflict of interest was reported by the author(s). and Irrigation and Drainage Needs: A World-wide View.”
Irrigation and Drainage 70 (5): 981–995.
Feng, S., J. Zhao, T. Liu, H. Zhang, Z. Zhang, and X. Guo. 2019.
Funding “Crop Type Identification and Mapping Using Machine
Learning Algorithms and Sentinel-2 Time Series Data.” IEEE
This study is supported by the Zhejiang Provincial Natural Journal of Selected Topics in Applied Earth Observations and
Science Foundation of China (No. LQ22D010007), the Open Remote Sensing 12 (9): 3295–3306. doi:10.1109/JSTARS.2019.
Fund of Key Laboratory of Urban Land Resources Monitoring 2922469.
2264 Y. WANG ET AL.

Ganin, Y., E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, Remote Sensing of Environment 83 (1–2): 195–213. doi:10.
F. Laviolette, M. Marchand, and V. Lempitsky. 2016. 1016/S0034-4257(02)00096-2.
“Domain-Adversarial Training of Neural Networks.” The Kumar, P., D. K. Gupta, V. N. Mishra, and R. Prasad. 2015.
Journal of Machine Learning Research 17 (1): 2030–2096. “Comparison of Support Vector Machine, Artificial Neural
Gao, B.-C. 1996. “NDWI—A Normalized Difference Water Index Network, and Spectral Angle Mapper Algorithms for Crop
for Remote Sensing of Vegetation Liquid Water from Space.” Classification Using LISS IV Data.” International Journal of
Remote Sensing of Environment 58 (3): 257–266. doi:10.1016/ Remote Sensing 36 (6): 1604–1617. doi:10.1080/2150704X.
S0034-4257(96)00067-3. 2015.1019015.
Gelder, B. K., A. L. Kaleita, and R. M. Cruse. 2009. “Estimating Lecun, Y., Y. Bengio, and G. Hinton. 2015. “Deep Learning.”
Mean Field Residue Cover on Midwestern Soils Using Nature 521 (7553): 436–444. doi:10.1038/nature14539.
Satellite Imagery.” Agronomy Journal 101 (3): 635–643. Lobell, D. B. 2013. “The Use of Satellite Data for Crop Yield Gap
doi:10.2134/agronj2007.0249. Analysis.” Field Crops Research 143: 56–64. doi:10.1016/j.fcr.
Ge, S., J. Zhang, Y. Pan, Z. Yang, and S. Zhu. 2021. “Transferable 2012.08.008.
Deep Learning Model Based on the Phenological Matching Lobell, D. B., D. Thau, C. Seifert, E. Engle, and B. Little. 2015.
Principle for Mapping Crop Extent.” International Journal of “A Scalable Satellite-Based Crop Yield Mapper.” Remote
Applied Earth Observation and Geoinformation 102: 102451. Sensing of Environment 164: 324–333. doi:10.1016/j.rse.
doi:10.1016/j.jag.2021.102451. 2015.04.021.
Gitelson, A. A., Y. J. Kaufman, R. Stark, and D. Rundquist. 2002. Löw, F. U., S. D. Michel, C. Conrad, and C. Conrad. 2013. “Impact
“Novel Algorithms for Remote Estimation of Vegetation of Feature Selection on the Accuracy and Spatial Uncertainty
Fraction.” Remote Sensing of Environment 80 (1): 76–87. of Per-Field Crop Classification Using Support Vector
doi:10.1016/S0034-4257(01)00289-9. Machines.” ISPRS Journal of Photogrammetry and Remote
Gorelick, N., M. Hancher, M. Dixon, S. Ilyushchenko, D. Thau, Sensing 85: 102–119. doi:10.1016/j.isprsjprs.2013.08.007.
and R. Moore. 2017. “Google Earth Engine: Planetary-Scale Ma, Y., Z. Zhang, H. L. Yang, and Z. Yang. 2021. “An Adaptive
Geospatial Analysis for Everyone.” Remote Sensing of Adversarial Domain Adaptation Approach for Corn Yield
Environment 202: 18–27. doi:10.1016/j.rse.2017.06.031. Prediction.” Computers and Electronics in Agriculture 187:
Gretton, A., K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and 106314. doi:10.1016/j.compag.2021.106314.
A. Smola. 2012. “A Kernel Two-Sample Test.” The Journal of Meng, S., X. Wang, H. Xin, C. Luo, and Y. Zhong. 2021. “Deep
Machine Learning Research 13 (1): 723–773. Learning-Based Crop Mapping in the Cloudy Season Using
Guo, H., R. Pasunuru, and M. Bansal. 2020. “Multi-Source One-Shot Hyperspectral Satellite Imagery.” Computers and
Domain Adaptation for Text Classification via Electronics in Agriculture 186: 106188. doi:10.1016/j.compag.
Distancenet-Bandits.” Pp. 7830–7838 in Proceedings of the 2021.106188.
AAAI Conference on Artificial Intelligence, New York, New MOA. 2022. “The Ministry of Agriculture and Rural Affairs.”
York, USA, Vol. 34. Retrieved 20 June 2022 http://zdscxx.moa.gov.cn:8080/mis
Guo, J., D. J. Shah, and R. Barzilay. 2018. “Multi-Source Domain portal/public/publicationRedStyle.jsp.
Adaptation with Mixture of Experts.” ArXiv Preprint ArXiv Oreopoulos, L., M. J. Wilson, and V. Tamás. 2011. “Implementation
1809: 02256. on Landsat Data of a Simple Cloud-Mask Algorithm Developed
Han, T., C. Liu, W. Yang, and D. Jiang. 2019. “A Novel for MODIS Land Bands.” IEEE Geoscience and Remote Sensing
Adversarial Learning Framework in Deep Convolutional Letters 8 (4): 597–601. doi:10.1109/LGRS.2010.2095409.
Neural Network for Intelligent Diagnosis of Mechanical Pan, S. J., and Q. Yang. 2009. “A Survey on Transfer Learning.”
Faults.” Knowledge-Based Systems 165: 474–487. doi:10. IEEE Transactions on Knowledge and Data Engineering
1016/j.knosys.2018.12.019. 22 (10): 1345–1359. doi:10.1109/TKDE.2009.191.
Hao, P., L. Di, C. Zhang, and L. Guo. 2020. “Transfer Learning for Peña, M. A., and A. Brenning. 2015. “Assessing Fruit-Tree Crop
Crop Classification with Cropland Data Layer Data (CDL) as Classification from Landsat-8 Time Series for the Maipo
Training Samples.” Science of the Total Environment 733: Valley, Chile.” Remote Sensing of Environment 171: 234–244.
138869. doi:10.1016/j.scitotenv.2020.138869. doi:10.1016/j.rse.2015.10.029.
Hao, P., Y. Zhan, L. Wang, Z. Niu, and M. Shakir. 2015. “Feature Peng, X., Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang. 2019.
Selection of Time Series MODIS Data for Early Crop “Moment Matching for Multi-Source Domain Adaptation.”
Classification Using Random Forest: A Case Study in Pp. 1406–1415 in Proceedings of the IEEE/CVF international
Kansas, USA.” Remote Sensing 7 (5): 5347–5369. doi:10. conference on computer vision, Long Beach, California, USA.
3390/rs70505347. Pott, L. P., T. J. C. Amado, R. A. Schwalbert, G. M. Corassa, and
He, K., X. Zhang, S. Ren, and J. Sun. 2016. “Deep Residual I. A. Ciampitti. 2021. “Satellite-based Data Fusion Crop Type
Learning for Image Recognition.” Pp. 770–778 in Classification and Mapping in Rio Grande Do Sul, Brazil.”
Proceedings of the IEEE conference on computer vision ISPRS J. Photogramm. Remote Sens 176: 21–196. doi:10.
and pattern recognition, Las Vegas, Nevada, USA. 1016/j.isprsjprs.2021.04.015.
Huete, A., K. Didan, E. Tomoaki Miura, P. Rodriguez, X. Gao, and Schmidhuber, J. 2015. “Deep Learning in Neural Networks: An
L. G. Ferreira. 2002. “Overview of the Radiometric and Overview.” Neural Networks 61: 85–117. doi:10.1016/j.neu
Biophysical Performance of the MODIS Vegetation Indices.” net.2014.09.003.
 GISCIENCE & REMOTE SENSING 2265

Skakun, S., B. Franch, E. Vermote, J.-C. Roger, I. Becker-Reshef, Wang, Q., W. Rao, S. Sun, L. Xie, E. S. Chng, and H. Li. 2018.
C. Justice, and N. Kussul. 2017. “Early Season Large-Area “Unsupervised Domain Adaptation via Domain Adversarial
Winter Crop Mapping Using MODIS NDVI Data, Growing Training for Speaker Recognition.” Pp. 4889–4893 in 2018
Degree Days Information and a Gaussian Mixture Model.” IEEE International Conference on Acoustics, Speech and
Remote Sensing of Environment 195: 244–258. doi:10.1016/j. Signal Processing (ICASSP), Calgary, Alberta, Canada.
rse.2017.04.026. Wang, Y., Z. Zhang, L. Feng, M. Yuchi, and D. Qingyun. 2021.
Song, C., C. E. Woodcock, K. C. Seto, M. P. Lenney, and “A New Attention-Based CNN Approach for Crop Mapping
S. A. Macomber. 2001. “Classification and Change Detection Using Time Series Sentinel-2 Images.” Computers and
Using Landsat TM Data: When and How to Correct Electronics in Agriculture 184: 106090. doi:10.1016/j.com
Atmospheric Effects?” Remote Sensing of Environment 75 (2): pag.2021.106090.
230–244. doi:10.1016/S0034-4257(00)00169-3. Xu, J., J. Yang, X. Xiong, L. Haifeng, K. C. Jingfeng Huang, Y. Y. Ting,
Suh, H. K., J. Ijsselmuiden, J. W. Hofstee, and E. J. van Henten. 2018. and T. Lin. 2021. “Towards Interpreting Multi-Temporal Deep
“Transfer Learning for the Classification of Sugar Beet and Learning Models in Crop Mapping.” Remote Sensing of
Volunteer Potato under Field Conditions.” Biosystems Environment 264: 112599. doi:10.1016/j.rse.2021.112599.
Engineering 174: 50–65. doi:10.1016/j.biosystemseng.2018.06. You, N., J. Dong, J. Huang, G. Du, G. Zhang, Y. He, T. Yang, Y. Di,
017. and X. Xiao. 2021. “The 10-m Crop Type Maps in Northeast
Sun, B., and K. Saenko. 2016. “Deep Coral: Correlation China during 2017–2019.” Scientific Data 8 (1): 1–11. doi:10.
Alignment for Deep Domain Adaptation.” Pp. 443–450 in 1038/s41597-021-00827-9.
European conference on computer vision, Amsterdam, The Zhang, H., Y. Wang, J. Shang, M. Liu, and Q. Li. 2021. “Investigating
Netherlands. Springer. the Impact of Classification Features and Classifiers on Crop
Sun, S., H. Shi, and W. Yuanbin. 2015. “A Survey of Multi-Source Mapping Performance in Heterogeneous Agricultural
Domain Adaptation.” Information Fusion 24: 84–92. doi:10. Landscapes.” Int. J. Appl. Earth Obs. Geoinformation 102:
1016/j.inffus.2014.12.003. 102388.
Tucker, C. J. 1979. “Red and Photographic Infrared Linear Zhao, S., X. Liu, C. Ding, S. Liu, W. Changshan, and W. Ling. 2020.
Combinations for Monitoring Vegetation.” Remote Sensing “Mapping Rice Paddies in Complex Landscapes with
of Environment 8 (2): 127–150. doi:10.1016/0034-4257(79) Convolutional Neural Networks and Phenological Metrics.”
90013-0. GIScience & Remote Sensing 57 (1): 37–48. doi:10.1080/
Turkoglu, M. O., S. D’Aronco, G. Perich, F. Liebisch, C. Streit, 15481603.2019.1658960.
K. Schindler, and J. D. Wegner. 2021. “Crop Mapping from Zheng, B., S. W. Myint, P. S. Thenkabail, and R. M. Aggarwal.
Image Time Series: Deep Learning with Multi-Scale Label 2015. “A Support Vector Machine to Identify Irrigated Crop
Hierarchies.” Remote Sensing of Environment 264: 112603. Types Using Time-Series Landsat NDVI Data.” International
doi:10.1016/j.rse.2021.112603. Journal of Applied Earth Observation and Geoinformation 34:
USDA. 2020. “United States Department of Agriculture National 103–112. doi:10.1016/j.jag.2014.07.002.
Agricultural Statistics Service.” Retrieved 6 December 2021 Zhong, L., H. Lina, and H. Zhou. 2019. “Deep Learning Based
(https://quickstats.nass.usda.gov). Multi-Temporal Crop Classification.” Remote Sensing of
Van der Maaten, L., and G. Hinton. 2008. “Visualizing Data Environment 221: 430–443. doi:10.1016/j.rse.2018.11.032.
Using T-SNE.” Journal of Machine Learning Research 9: 11. Zhou, Y., J. Luo, L. Feng, Y. Yang, Y. Chen, and W. Wu. 2019.
Veeck, G., A. Veeck, and Y. Hongyan. 2020. “Challenges of “Long-Short-Term-Memory-Based Crop Classification Using
Agriculture and Food Systems Issues in China and the High-Resolution Optical Images and Multi-Temporal SAR
United States.” Geography and Sustainability 1 (2): 109–117. Data.” GIScience & Remote Sensing 56 (8): 1170–1191.
doi:10.1016/j.geosus.2020.05.002. doi:10.1080/15481603.2019.1628412.
Vuolo, F., M. Neuwirth, M. Immitzer, C. Atzberger, and N. Wai- Zhu, Y., F. Zhuang, and D. Wang. 2019. “Aligning Domain-Specific
Tim. 2018. “How Much Does Multi-Temporal Sentinel-2 Data Distribution and Classifier for Cross-Domain Classification from
Improve Crop Type Classification?” International Journal of Multiple Sources.” Proceedings of the AAAI Conference on
Applied Earth Observation and Geoinformation 72: 122–130. Artificial Intelligence Honolulu, Hawaii, USA, EAAI 2019
doi:10.1016/j.jag.2018.06.007. 5989–5996.