A new procedure for optimizing neural network using stochastic algorithms in predicting and assessing landslide risk in East Azerbaijan

Abstract

This study utilized artificial neural network (ANN) optimization techniques including biography-based optimization (BBO), earthworm optimization (EWA), shuffled complex evolution (SCE), and stochastic fractal search (SFS) to predict landslide susceptibility mapping. The ANN model was optimized using hybrid algorithms based on BBO-MLP, EWA-MLP, SCE-MLP, and SFS-MLP. A large dataset consisting of 3211 training and testing datasets from the eastern Azerbaijan province in west Iran was used to prepare the ANN network. The variables of the algorithms were optimized, including network parameters and weights, to create reliable maps of landslide susceptibility. The layers for preparing the landslide susceptibility map included 16 environmental, geographical, hydro-geomorphological, and climatic factors. The accuracy of the probabilistic models was evaluated using the area under the curve criterion within the context of predictive modeling for landslide susceptibility mapping. Numerous algorithms and swarm sizes were employed to assess the results. The area under the curve (AUC) was used to measure the accuracy of these algorithms. For the BBO-MLP and EWA-MLP models, AUC values were calculated for different population sizes in training databases. The optimal hybrid model for the two algorithms was determined to have a swarm size of 500. Similarly, the SCE-MLP and SCE-MLP models were assessed and determined to possess remarkable precision, as evidenced by AUC scores that varied between 0.9965 and 0.9997 for both training and testing. Notably, the SFS-MLP model exhibited the most exceptional accuracy overall, with an AUC value that surpassed all others. The SFS-MLP model had the highest overall accuracy with an AUC value of 0.9879 compared to SCE-ANN (AUC = 0.9887) and BBO & EWA-ANN (AUC = 0.9865, 0.9793). These algorithms proved effective in optimizing artificial neural networks and improving performance in landslide risk zoning.