Implication of novel hybrid machine learning model for flood subsidence susceptibility mapping: A representative case study in Saudi Arabia

Flash floods are among the most dynamic and complex hydrological phenomena, posing a significant challenge for detection in hydrological studies. Accurate and timely predictions of flash floods are difficult to achieve with current approaches of detection. These difficulties are attributed to the complex hydrological processes, the dynamic character of flash floods, and the requirement for real-time data interpretation. To address these challenges, advanced machine learning (ML) techniques offer great potential. This research proposed a new version of the hybrid ML model and tested its potential and effectiveness for flood subsidence susceptibility mapping. The model combines the Levenberg-Marquardt (LM) algorithm with different metaheuristic algorithms as an optimization algorithm for training the deep neural network (DNN) model. The proposed model is called Gbest-guided Artificial Bee Colony with directed scout_LM (GABCds_LM). The primary advantage of the proposed model is that during each iteration of the learning process, the GABCds algorithm actively explores different search areas and forces the LM algorithm to exploit promising findings. In addition, the proposed hybrid ML model also delivers a balanced establishment between exploration and exploitation for better learning. By integrating the strengths of exploration from the GABCds algorithm and the convergence speed of the LM algorithm, the GABCds_LM algorithm overcomes the limitations of existing methods and significantly improves flood subsidence susceptibility mapping accuracy. A representative case study at Jeddah city in Saudi Arabia was selected to be evaluated. The area under the curve (AUC) and the receiver operating characteristic (ROC) curve. The research findings demonstrated that the GABCds_LM model achieves the highest accuracy (AUC = 0.95), followed by the GABC_LM (AUC = 0.94) and ABC_LM (AUC = 0.94) models. In comparison, the LM model had a lower accuracy (AUC = 0.8). Furthermore, the GABCds_LM model exhibits excellent precision (94.87 %), sensitivity (94.87 %), specificity (95.56 %), F1 score (94.87 %), and overall accuracy (95.24 %). This indicates that using GABCds_LM for training DNN outperforms other approaches and can serve as a robust algorithm in flood subsidence susceptibility mapping. The success of this approach suggests its potential for application of flood control, watershed management and water resources management.