综合负样本优化指数与CNN-LSTM-ATT模型的滑坡易发性评价OA

Traditional random sampling of non-landslide introduces high uncertainty with low accuracy during landslide susceptibility assessment.Moreover,computational models often struggle to effectively process multidimensional data and improve prediction accuracy.To address these challenges,this study proposes the negative sample optimization index(NSI)to enhance the selection strategy for non-landslide samples and employs a deep neural network model integrating attention mechanism(ATT),convolutional neural network(CNN),and long short-term memory(LSTM)for landslide susceptibility assessment in Yihe Town,Suide County,located in the Loess Plateau of Shaanxi Province.Firstly,a total of fourteen landslide influencing factors were considered as the input variables to estimate the probability of landslide occurrence.The NSI was calculated by weighting,summing,and normalizing the outputs of three machine learning base models according to their Matthews correlation coefficient(MCC)scores.Then,the NSI and historical landslide records were used to construct the model training database.Finally,the CNN-LSTM-ATT model was applied for landslide susceptibility assessment,with factor importance analyzed using Shapley additive explanations(SHAP).The results indicate that NSI improves the quality of non-landslide samples by constraining the sampling space,thereby mitigating prediction errors induced by excessively biased negative samples,with the model accuracy increasing by up to 7%.Meanwhile,compared to individual models,the CNN-LSTM-ATT model,which integrates multiple complex layers,demonstrates a high performance up to 0.925.Additionally,slope,elevation,and distance to buildings are the key factors influencing the landslide susceptibility mapping in the study area.The proposed sampling strategy and CNN-LSTM-ATT model provide valuable technical support for the spatial prediction of landslides in the Loess Plateau region.