A new combined approach of two neural-metaheuristic techniques based on the Cuckoo optimization algorithm and backtracking search algorithms for predicting and appraisal of landslide susceptibility mapping

Abstract In this research, a hybrid Backtracking Search Algorithm (BSA) and Cuckoo Optimization Algorithm (COA)-based artificial neural network (ANN) model (BSA-MLP and COA-MLP) was used to predict landslide susceptibility mapping (LSM) in an area in the province of Kurdistan, west of Iran. The input dataset includes elevation, slope angle, rainfall, and land use. The output is a value that shows how likely a landslide will happen. The parameters and weights of the BSA and COA algorithms were fine-tuned to produce the most accurate LSM. Table 2 illustrates the effect of the number of layers and neurons on the accuracy of models produced using the standard ANN approach. Root-mean-squared error (RMSE) and correlation coefficient (R2) were used to compare different network designs. Table 2 shows that feed-forward back-propagation with six hidden layers (a transit function and six neurons in the hidden layer) gave the best results. The model got more accurate as the number of hidden layers and neurons increased, but it stopped getting better when there were four hidden layers. After a certain point, the model became too complicated to be worth the benefit of being more accurate. A model may be overfitted if it employs several hidden layers and neurons. The results of the first optimization stage serve as the basis for further operations. The succeeding parts thus use the findings of these networks. In the next step, to make the models better, the number of neurons in each hidden layer will be changed. This suggests that the number of neurons may increase or decrease depending on the results of the first phase. Many networks with different types of neurons and combinations of layers have been built to find the best architecture. The most successful network design was generated via a feed-forward back-propagation algorithm with six hidden layers. A model may be overfitting if it has several hidden layers and neurons.