Prediction of Inland Excess Water Inundations Using Machine Learning Algorithms

Abstract

Regularly, large parts of the agricultural areas of the Great Hungarian Plain are inundated due to excessive rainfall and insufficient evaporation and infiltration. Climate change is expected to lead to increasingly extreme weather conditions, which may even increase the frequency and extent of these inundations. Shallow “floods”, also defined as inland excess water, are phenomena that occur due to a complex set of interrelated factors. Our research presents a workflow based on active and passive satellite data from Sentinel-1 and -2, combined with a large auxiliary data set to detect and predict these floods. The workflow uses convolutional neural networks to classify water bodies based on Sentinel-1 and Sentinel-2 satellite data. The inundation data were complimented with meteorological, soil, land use, and GIS data to form 24 features that were used to train an XGBoost model and a deep neural network to predict future inundations, with a daily interval. The best prediction was reached with the XGBoost model, with an overall accuracy of 86%, a Kappa value of 0.71, and an F1 score of 0.86. The SHAP explainable AI method showed that the most important input features were the amount of water detected in the satellite imagery during the week before the forecast and during the period two weeks earlier, the number of water pixels in the surroundings on the day before the forecast, and the potential evapotranspiration on the day of the forecast. The resulting inland excess water inundation time series can be used for operational action, planning, and prevention.