Development of a Neural Network Algorithm for automatic operational detection of Amur River basin floods based on Sentinel-2 satellite data

Abstract

In this paper, the authors propose an algorithm for automatic operational flood detection of the Amur River basin based on data from the MSI instrument installed on the Sentinel-2 satellite. To solve the problem, a U-net convolutional neural network is used, improved due to the specifics of the data by adding an additional layer that reduces sampling and normalization layers after each neuron activation. As a training set, Sentinel-2 Level-2A data was used, which underwent the atmospheric correction procedure and represents 4 channels of the MSI instrument with a spatial resolution of 10 m, as well as index images built on their basis. As reference information, vector maps of river floods were used, built in an interactive mode by decoder specialists from the Far-Eastern Center of State Research Center for Space Hydrometeorology “Planeta”. The results of the neural network algorithm after the learning process were evaluated according to the metrics that amounted to: Precision – 94.91%, Recall – 90.76%, F1-measure – 92.79%. High accuracy ratings and fast operation speed make it possible to use the developed algorithm for automatic operational flood detection of the Amur River basin floods in the tasks of integrated monitoring of flood conditions. The work is a complete technical solution and has been put into trial operation at the Far-Eastern Center of State Research Center for Space Hydrometeorology “Planeta”. In the future, the results obtained by the authors are planned to be adapted to the data of the Russian satellite of the Meteor-M series with the KMSS-2 multi-zone satellite imagery complex installed on board. The expected results will improve the quality of the thematic products provided and will make it possible to switch over to the use of domestic data when creating our own technologies for processing satellite information.