Enhancing flood susceptibility modeling using integration of multi-source satellite imagery and multi-input convolutional neural network

Flood susceptibility maps are vital tools for implementing prevention and mitigation strategies. This study describes the potential application of convolutional neural networks (CNN) from two standpoints, single-input and multi-input CNN, to improve flood susceptibility modeling. Firstly, optical (Sentinel-2 and Landsat-8) and radar (Sentinel-1) satellite images were integrated to identify flooded and non-flooded areas. Moreover, a geospatial database with thirteen geo-environmental features including altitude, slope, rainfall, land use and land cover (LULC), normalized difference vegetation index (NDVI), normalized difference moisture index (NDMI), aspect, curvature, drainage density, topographic wetness index (TWI), stream power index (SPI), soil texture, and distance from the river, was created in Aqqala County, Golestan Province, Iran. This study concentrates on improving the prediction performance by enhancing the feature extraction capabilities of the CNN model. To achieve this, a multi-input CNN model is developed and compared with the single-input CNN model. The validation results in terms of the area under the receiver operating characteristic (ROC) curve (AUC) showed that the multi-input CNN model in training (AUC = 0.998) and testing (AUC = 0.946) performed better than the single-input CNN model in training (AUC = 0.987) and testing (AUC = 0.896). The results also demonstrated the potential of the multi-input CNN model as a promising flood susceptibility prediction model.