Soccer Fields as Rainfall Detectors using Machine Learning

Abstract

Agriculture is an important source of income for many countries in the Global South, where it may account for as much as 25% of GDP. Precipitation is crucial for agriculture in countries like Ghana, where ~95% of farming is rainfed. Accurate rainfall observations are limited in Ghana. The sparse rain gauge network and the lack of weather radars make remote sensing methods a potentially attractive alternative source of rainfall data. Radar satellites, such as Sentinel-1, emit radiation that passes through the atmosphere and is scattered back to the satellite by the Earth’s surface. The backscatter measured by the satellite is correlated with the wetness of the soil but the existence of vegetation hinders straightforward quantification of soil moisture. By choosing sites with a simple and, more or less, constant phenology, it may be possible to eliminate the effect of vegetation on backscatter. Soccer field may qualify as sites with such a simple and constant phenology. The main objective of this study is to use the Sentinel-1 data over soccer fields and assess them as rainfall detectors. A machine learning approach will be used to reach this objective.
This research assessed the stability and the generalization capabilities of a classification model (rain/no rain). The model was trained with and applied to different locations and periods (2019 & 2020). Ground observations from 53 Ghanaian (TAHMO) and 1 Greek stations were used. Soccer fields in Ghana and Greece were selected and their suitability as rainfall detectors was checked based on the correlation between modeled soil moisture and backscatter strength.
The rain/no rain classification of the soccer fields was made with a stacked classifier that was trained and validated with both spaceborne and ground data. The classifier was tested on six different datasets from Greece and Ghana 2019 and 2020. The stability of the model was assessed by a Leave-p out cross-validation approach. The generalization in space was tested by using different environments. The generalization in time was tested by using different time periods. The results showed that the classification was stable. The minimum and maximum performances for the different testing datasets were 0.43 to 0.85. The median performance of the algorithm in Ghana for 2020 is 67%. The stacked classifier was found to have the best performance compared to other classifiers. Finally, the performance of the stacked classifier was competitive in comparison with the performance of the well-known IMERG algorithm.
The study showed that there is a potential for using radar backscatter from suitable fields to detect rainfall. The classifier is stable and can be generalized in time and space under certain conditions.