Flood susceptibility mapping using neural network based models in Morocco: Case of Souss Watershed

The global climate situation becomes more and more critical due to the impacts of climate change especially when dealing with flood hazard causing major human and economic losses every year. In Morocco, the Souss-watershed is one of the most vulnerable regions in term of flooding and land degradation. The climate conditions, population growth affect more the land use conditions. The present work introduces a novel approach to assess flood risk in Souss watershed using 4 neural network based models in Google Colab: Artificial neural networks (ANN), Recurrent Neural networks (RNN), One dimensional (1DCNN) and Two dimensional Convolutional neural networks (2DCNN). The models input were constructed using 4 features chosen from 17 of the most triggering flood factors that describe the characteristics of the watershed, including topography, vegetation and soil ones. The Pearsons correlation factor was applied to evaluate the correlation between the features, the variance inflation factor analysis (VIF) was applied to diagnose the collinearity and the Shapley Additive Explanations (SHAP) was applied to evaluate the importance of a factor in the prediction model. For the evaluation and validation process, the calculation of the Mean Absolute Error (MAE) and loss was used to evaluate the accuracy of predictions along with the calculation of the ROC (Receiver Operating Characteristic) and the AUC (Area Under the ROC Curve) to compare between the four models, the results demonstrated that the RNN has the highest performance with an accuracy of 96% and a validation loss of 0.0984 and a validation MAE of 0.2553, followed by the ANN with a slightly lower accuracy of 95% , 2DCNN and 1DCNN demonstrated lower accuracies of 87% and 81%. These findings have demonstrated that in the flood susceptibility mapping context, the application of complex neural networks such as 1DCNN and 2DCNN calls for more tuning and optimizing to overcome over-fitting issues, and that using simple neural networks such as RNN and ANN can be more effective in achieving more accurate predictions.