Abstract: Standard BP (back propagation) neural network suffers from inherent drawbacks such as easy trapping in local optimal solutions and high sensitivity to initial weights. To address these limitations, the crayfish optimization algorithm (COA) was introduced to optimize the BP neural network, and a COA-BP landslide susceptibility evaluation model was constructed. Taking Wuyi County in Zhejiang Province as the study area, the landslide susceptibility evaluation was carried out. Firstly, four unimodal test functions and one multimodal test function were adopted to test the performance of COA, and comparative analysis was conducted with the multi-verse optimizer (MVO), sparrow search algorithm (SSA), and marine predators algorithm (MPA). Subsequently, 13 environmental conditioning factors including elevation, slope, aspect, and precipitation were extracted from the study area. The proposed COA-BP model was applied to landslide susceptibility evaluation, with BP, MVO-BP, MPA-BP and SSA-BP models set as comparison models. The classification accuracy was verified via the confusion matrix and two-level index radar chart. The results demonstrate that COA outperforms the other three optimization algorithms. The COA-BP model presents superior predictive capability with clearer zoning of landslide susceptibility levels and high recognition accuracy for both landslide and non-landslide samples. The area under curve (AUC) value of the model exceeds 0.9, which meets the accuracy requirements for landslide susceptibility evaluation in the study area.

Key words: crayfish optimization algorithm, BP neural network, landslide susceptibility, confusion matrix, Wuyi County, Zhejiang Province