Remote Sensing For Assessing Drought Insurance Claims in Central Europe

In this study, the viability of assessing drought insurance claims via remote sensing is explored. Time series of satellite images from the Sentinel-2 mission and weather data from the European Climate Assessment & Dataset are used to fit classifiers on historical loss data from an agricultural insurance. Two different approaches for training classifiers are explored, designing neural networks to learn directly on the time series and transforming the data to a fixed-size representation to enable the use of other methods. It is shown that in this case the second approach yields much better results, as careful feature engineering combined with more rigid methods like gradient boosting leads to less overfitting compared to the neural network approach.Compared to existing approaches, the proposed methods allow for analyzing the situation on a per-field level while using high resolution imagery (Sentinel-2) over a large geographic area.