Improving deep learning-based flood susceptibility modeling by integrating data balancing technique and dual-input convolutional neural network

Flood susceptibility maps are fundamental tools for identifying flood-susceptible areas to facilitate planning and resource allocation. This paper examines the capabilities of two types of convolutional neural networks, dual-input convolutional neural networks (DICNN) and single-input convolutional neural networks (SICNN), in flood susceptibility modeling. Hence, radar satellite images from Sentinel-1 were employed to extract flooded areas in the Gorganrood watershed located in Golestan Province, Iran. Furthermore, fifteen predictive factors encompassing altitude, normalized difference moisture index (NDMI), slope, land use, topographic wetness index (TWI), soil texture, lithology, drainage density, normalized difference vegetation index (NDVI), aspect, rainfall, distance from rivers, stream power index (SPI), profile and plan curvature were identified and prepared for the modeling process. This study also used a spatial flood data balancing technique, generative adversarial networks (GAN), to correct imbalanced flood datasets. Moreover, the information gain ratio (IGR) analysis was performed to identify the effectiveness of each predictive factor. The area under the receiver operating characteristic curve (AUC) was analyzed to validate the developed models. The validation outcomes demonstrated that the DICNN model outperforms the SICNN model in both training (AUCtraining=0.974; AUCtraining=0.966) and testing (AUCtesting=0.954; AUCtesting=0.933), respectively. Results also showed that about 15.17% of the land, mostly concentrated in downstream areas, exhibits high and very high levels of susceptibility, with altitude being the most influential factor in flood occurrence within the study area.