Application of Multiple Geographical Units Convolutional Neural Network based on neighborhood effects in urban waterlogging risk assessment in the city of Guangzhou, China

Abstract

Abstract Waterlogging has recently occurred in cities across China, causing serious economic and property losses. It is imperative to assess the susceptibility of cities to waterlogging, especially those in urban areas, to reduce the damages caused by waterlogging. Artificial Neural Network algorithms have been used to assess the hazards of waterlogging because of their simplicity, objectivity and efficiency. Considering the neighborhood effects, this research proposes an integrated Multiple Geograpical Units Convolutional Neural Network (MGCNN) model for assessing the waterlogging risk in the primarily urban area of Guangzhou, China. Ten spatial factors were considered to be added to the MGCNN, which are DEM, Slope, Std, NDVI, NDBI, MNDWI, Entropy, Contrast, Correlation and Energy based on a Gray-level Co-occurrence Matrix (GLCM). The results show that 1) the texture features of remote sensing image plays an important role in the identification of urban waterlogging area through significance analysis of the ten impact factors; 2) the accuracy rate of the MGCNN trained with the input data of the four risk levels is higher than 0.88, while the maximum accuracy rate of the two risk levels is only about 0.6; 3) the MGCNN has a lower cross-entropy loss value of 0.82 than the SGCNN of 1.13 after 20,000 iterations of training; 4) The Area Under the Curve (AUC) value of the MGCNN was about 0.95, highest among the MGCNN,SGCNN, SVM-BRF and SVM-Poly models; 5) Extremely high risk areas for waterlogging were mainly found in low-lying areas with a high density of urban buildings, including the Yuexiu district, the northern Liwan district, the central and southern sections of the Tianhe district, and the western Haizhu district. Therefore, fully considering neighborhood effects of the surrounding geographical units, the constructed MGCNN model improves not only the accuracy but also the interpretability of the black box.