=Paper=
{{Paper
|id=Vol-2540/paper44
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2540/FAIR2019_paper_31.pdf
|volume=Vol-2540
}}
==None==
https://ceur-ws.org/Vol-2540/FAIR2019_paper_31.pdf
Precipitation forecasting using satellite images
and SVMs?
Eslam Hussein[0000−0001−6040−8516] and Mehrdad
Ghaziasgar[0000−0002−7713−677X]
Department of Computer Science, University of the Western Cape, South Africa
eslamhuss34@gmail.com mghaziasgar@uwc.ac.za
Abstract. The prediction of rainfall is important for planning; it can
help individuals plan their days ahead; more importantly it can help
governments prepare for potential disasters. This research aims to in-
vestigate a data-driven approach to rainfall intensity prediction using
support vector machines (SVMs) and sequences of daily satellite precip-
itation images as input. The primary aim of the work is to accurately
predict one day ahead, but is also extended to predict several days into
the future.
Keywords: Machine Learning· Rainfall · Satellite · Sequence · Images.
1 Introduction
Weather forecasting aims to predict the state of the atmosphere at a specific
location and time in the future. The accurate prediction of the weather and
climate is essential to the production of crops [1]. Weather can also cause natural
disasters, such as typhoon in Mozambique in April 2019. An accurate weather
forecasting system can provide early warnings which can help individuals and
governments to better prepare for such events.
Recently, a number of papers studied the use of different machine learning
techniques for weather prediction [[2], [3], [4], [5]]. Due to space constraints,
we only explain Boonyuen et al’s [6] work here. Boonyuen et al proposed the
use of the standard “Inception 3” model to forecast rain fall up to three days
ahead. Satellite image data taken every ten minutes was used as input. The
system used one image as input to predict up to three days ahead on a binary
classification scale of “rain” versus “no rain” with an accuracy of 63%. This
accuracy represents an 13% advantage over random guessing which statistically
yields a 50% accuracy for two-class problem. The authors recommended the use
of more than one image as input, and to attempt to predict up to one week
ahead. They also proposed to extend the classification scale to more classes,
such as light rain, moderate rain, and heavy rain, which can serve as a more
useful source of information.
?
Supported by the Openserve/Telkom/Aria Technologies Centre-of-Excellence at
UWC and the use of the ilifu cloud computing facility.
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0)
2 E. Hussein and M. Ghaziasgar.
Fig. 1. Satellite precipitation images: (left) An image from the National Centers for
Environmental Prediction (NCEP) data set; (centre) 5 grid overlay; (right) Grid cell
13.The NCEP data were provided by the NOAA at /https://www.esrl.noaa.gov/psd/
This research aims to investigate precipitation forecasting on the NCEP data
set using SVMs.The data set consists of images of the United States (US) taken
at 7am daily. One image of the data set is shown on the left of Figure 1. The
parameters that will be investigated are: i) the length of the input image se-
quence i.e. the number of input days ∈ {2, 4, 6, 8} and ii) the size of the input
images ∈ {100%, 50%, 25%}. Furthermore, we propose to divide up the US us-
ing a 5×5 grid seen in the centre of Figure 1; the US is very large and it is
expected that the weather patterns that we aim to model and predict are more
region-specific. We therefore aim to develop a separate model for, and inves-
tigate each of the two parameters mentioned previously on, each grid cell. Fi-
nally, we aim to predict a a multi-class output where we classify rainfall into
three classes ∈ {no rain, light rain, heavy rain}, and to predict k days ahead
∈ {1, 2, 3, 4, 5, 6, 7, 14, 30}.
2 Preliminary Results
As a start, we ran an experiment on grid cell 13, indicated in the centre of Figure
1 and shown more clearly on the right of the figure, using sequences of 2 days
as input and image sizes of 50%. The NCEP data was class-reduced into three
classes, and all sequences of length 2 which started with a minimum threshold
of precipitation were determined and used. This resulted in 2548 sequences for
training and 284 sequences for testing. Training involved selecting the best cost
parameter C of the SVM model using 10-fold cross-validation on the training
set. Table 1 illustrates the prediction F1-score on the testing set for 1–30 days
ahead (DA) on this grid cell.
Noting that the statistical accuracy of an untrained system on a 3-class prob-
lem is about 33%, the results are very encouraging. We are able to predict one
DA with a 22% advantage over an untrained system. The F1 score fluctuates
between 40% and 48% from 2 DA to 14 DA. Finally, we are able to predict at a
4% advantage over an untrained system even for a full month ahead.
Table 1. Testing F1-Score for predicting k days ahead (DA).
F1-Score of Days Ahead (DA) (%)
1DA 2DA 3DA 4DA 5DA 6DA 7DA 14DA 30DA
55 46 48 46 41 40 48 46 37
Precipitation forecasting using satellite images and SVMs 3
References
1. M. Reyniers, Quantitative Precipitation Forecasts Based on Radar Observations:
Principles, Algorithms and Operational Systems. Institut Royal Météorologique de
Belgique Brussel, Belgium, 2008.
2. H. S. Jang, K. Y. Bae, H.-S. Park, and D. K. Sung, “Solar power prediction based
on satellite images and support vector machine,” IEEE Transactions on Sustainable
Energy, vol. 7, no. 3, pp. 1255–1263, 2016.
3. S. S. Aung, Y. Senaha, S. Ohsawa, I. Nagayama, and S. Tamaki, “Short-term pre-
diction of localized heavy rain from radar imaging and machine learning,” IEIE
Transactions on Smart Processing & Computing, vol. 7, no. 2, pp. 107–115, 2018.
4. B. Klein, L. Wolf, and Y. Afek, “A dynamic convolutional layer for short range
weather prediction,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pp. 4840–4848, 2015.
5. X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.-c. Woo, “Convolu-
tional lstm network: A machine learning approach for precipitation nowcasting,” in
Advances in Neural Information Processing Systems, pp. 802–810, 2015.
6. K. Boonyuen, P. Kaewprapha, and P. Srivihok, “Daily rainfall forecast model from
satellite image using convolution neural network,” in 2018 IEEE International Con-
ference on Information Technology, pp. 1–7, 2018.