Multi-scale extreme climate disaster prediction model integrated with ConvLSTM: taking rainstorm and flood disaster as an example

As one of the most influential and harmful disasters in extreme climate events, rainstorm and flood threaten people’s lives, property and social stability seriously. The research applied the historical disaster data to predict future risk assessment with deep learning method for timely response. Firstly, four disaster indicators (population disaster index, housing disaster index, agricultural disaster index and economic disaster index) were introduced to reflect different losses, which could form a comprehensive disaster index to quantify the overall loss degree; Second, with raster data and VGGNet, a lightweight regression convolutional neural network model VGG-Light was proposed to solve these problem; Third, focused on impact of precipitation on disaster situations, the ConvLSTM module was used to capture the spatiotemporal characteristics of precipitation data, and then the TSVGG-Light model was presented for feature fusion. Finally, the comprehensive disaster index obtained on the validation set had MSE and MAE of 0.012 and 0.007, R2 increased to 0.918, CS (0.05) increased to 97.50%, while CS (0.1) and CS (0.2) were both 100.00%. Therefore, TSVGG-Light model integrated with ConvLSTM has been proved efficiently for extracting the spatio-temporal characteristics of precipitation, which can accurately predict rainstorm and flood disasters compared with statistics characteristics of precipitation.