A Machine Learning Approach for Rainfall

Estimation Integrating Heterogeneous

Data Sources
abstract
• Providing an accurate rainfall estimate at individual points is a challenging
problem in order to mitigate risks derived from severe rainfall events, such as
floods and landslides. Dense networks of sensors, named rain gauges (RGs),
are typically used to obtain direct measurements of precipitation intensity in
these points. These measurements are usually interpolated by using spatial
interpolation methods for estimating the precipitation field over the entire
area of interest. However, these methods are computationally expensive, and
to improve the estimation of the variable of interest in unknown points, it is
necessary to integrate further information. To overcome these issues, this
work proposes a machine learning-based methodology that exploits a
classifier based on ensemble methods for rainfall estimation and is able to
integrate information from different remote sensing measurements. The
proposed approach supplies an accurate estimate of the rainfall where RGs
are not available, permits the integration of heterogeneous data sources
exploiting both the high quantitative precision of RGs and the spatial pattern
recognition ensured by radars and satellites, and is computationally less
expensive than the interpolation methods.
Introduction
• 1) Three heterogeneous data sources (i.e., RGs, radar, and
Meteosat) are integrated to generate more accurate estimates
of rainfall events.
• 2) Different classification methods are compared on a real
case concerning Calabria, a southern region in Italy, and a
hierarchical probabilistic ensemble approach is proposed.
• 3) Different ML-based methods, pre trained only on historical
data, with a widely used interpolation method in the
hydrological field (i.e., KED) are compared.
Literature survey
• Verdin et al. [23] also adopt Bayesian estimation in order to estimate the
parameters of the model; their system integrates RG observations and
satellite data and adopts an interpolation technique based on the Kriging
method. All these techniques are able to provide interesting results, but they
require a rather delicate phase of parameters estimation of the particular
model; therefore, as a side effect, usually, their flexibility and effectiveness
tend to be hampered. As the relations between sensors data, cloud
properties, and rainfall estimates are highly nonlinear, more flexible
approaches based on ML techniques have been investigated recently. For
instance, the problem of detecting convective events and closely related rainy
areas is addressed in [24] by using ANNs combined with support vector
machines. Data sets are obtained by processing data coming from optical
channels of the multispectral instrument onboard of Meteosat Second
Generation (MSG) satellites; different from our work, RG measures are used
only as a reference but not in the training phase of the algorithm. Sehad et al.
[25] propose an approach to rainfall estimation based on SVMs; the input data
are integrated from multispectral channels on MSG; and two models are
developed for daytime and nighttime respectively
Existing system
• An existing system is based on the ensemble paradigm include the work
in [12], which, similar to our work, employs a probabilistic ensemble and
merges two sources of data (i.e., rain gauges and radar) even if the aim of
this work is to develop a run-off analysis. Afterward, a blending technique
is applied to the results of the runoff hydrologic models to determine a
single runoff hydrograph. Experimental results show that the hydrologic
models are accurate and can help to make more effective decisions in the
flood warning. Frei and Isotta [13] define a technique for deriving a
probabilistic spatial analysis of daily precipitation from rain gauges. The
final model represents an ensemble of possible fields, conditional on the
observations, which can be explained as a Bayesian predictive
distribution measuring the uncertainty due to the data sampling from the
station network. An evaluation of a real case study, located in the
European Alps, proves the capability of the approach in providing
accurate predictions for a hydrological partitioning of the region.
•
Disadvantages of existing
system
• The system is not implemented hierarchical probabilistic
ensemble classifier (HPEC) for rainfall prediction.
• The system is implemented artificial neural networks (ANNs)
as a forecasting method in which prediction is not accurate.
Proposed system
• Our approach is an effective solution for real scenarios, as in the case
of an officer of the Department of Civil Protection (DCP), who has to
analyze the rainfall in a specific zone presenting risks of landslides or
floods. The experimental evaluation is conducted on real data
concerning Calabria, a region located in the South of Italy, and
provided by the DCP. Calabria is an effective test ground because of its
strong climate variability and its complex orography. Our contributions
can be summarized as follows.
• 1) Three heterogeneous data sources (i.e., RGs, radar, and Meteosat)
are integrated to generate more accurate estimates of rainfall events.
• 2) Different classification methods are compared on a real case
concerning Calabria, a southern region in Italy, and a hierarchical
probabilistic ensemble approach is proposed.
• 3) Different ML-based methods, pre trained only on historical data,
with a widely used interpolation method in the hydrological field (i.e.,
KED) are compared.
Advantages of proposed
system
• In the proposed system, raw data are preprocessed to make
them suitable for the analysis, and an under sampling strategy
is adopted to address the class unbalanced problem.
• The proposed system developed an Effect of Integrating RG,
Satellite, and Radar Measurements and are tested and trained
with an effective ML Classifiers.
System architecture
Novelty
• ungauged points (in a real usage scenario) in order to estimate
the severity class of these points and
• 2) a separated set of training data (i.e., not used in the
learning phase).
conclusion
• An ML-based approach for the spatial rainfall field estimation
has been defined. By integrating heterogeneous data sources,
such as RGs, radars, and satellites, this methodology permits
estimation of the rainfall, where RGs are not present, also
exploiting the spatial pattern recognition ensured by radars
and satellites. After a phase of preprocessing, a random
uniform under sampling strategy is adopted, and finally, an
HPEC permits the model used to be built to estimate the
severity of the rainfall events. This ensemble is based on two
levels: in the first level, a set of RF classifiers are trained,
while, in the second level, a probabilistic metal earner is used
to combine the estimated probabilities provided by the base
classifiers according to a stacking schema