北京测绘2025，Vol.39Issue(6)：843-848,6.DOI:10.19580/j.cnki.1007-3000.2025.06.015
结合注意力机制和LSTM的降雨型滑坡沉降预测
Precipitation-induced landslide subsidence prediction combining attention mechanism and LSTM
摘要
Abstract
To accurately simulate and predict the evolution trend of precipitation-induced landslides,this paper proposed a combined prediction model that started with multi-feature inputs and increased the number of adjacent monitoring points.Firstly,convolutional neural networks(CNN)were utilized to extract spatial features from multi-dimensional time-series data.Secondly,within a multi-layer perceptron(MLP),these spatial features underwent Hadamard product operations through attention mechanism and feature inputs,enabling the allocation of feature attention weights.By leveraging the superior time-series modeling capabilities of long short-term memory(LSTM)networks,the subsidence trend was precisely predicted for attention-enhanced features.Finally,a fully connected layer integrated all processed information to output high-precision displacement predictions for specific monitoring points.An analysis was conducted using a landslide case study.Experimental results demonstrate that compared to CNN-LSTM and LSTM models without the integration of an attention mechanism,the CNN-Attention-LSTM model designed in this paper significantly improves prediction accuracy.Specifically,compared to the LSTM model,the mean absolute error(MAE)and root mean square error(RMSE)are reduced by 84.6%and 79.5%,respectively,and compared to the CNN-LSTM model without the integration of an attention mechanism,the MAE and RMSE are reduced by 49.57%and 47.98%,respectively,proving the superior performance and broad application prospects of this model in landslide displacement prediction.Additionally,the advantages of increasing feature inputs are discussed through experiments,revealing that incorporating features closely related to the prediction can further significantly enhance the model's predictive performance.In particular,the MAE and RMSE of the LSTM model are reduced by 19.4%and 19.3%,respectively,and those of the CNN-LSTM model are reduced by 12.1%and 13.2%,respectively;the MAE and RMSE of the CNN-Attention-LSTM decline by 66.1%and 56.3%,respectively.关键词
阵雨型滑坡/沉降预测/注意力机制/长短期记忆Key words
shower-induced landslide/subsidence prediction/attention mechanism/long short-term memory分类
天文与地球科学引用本文复制引用
付军,查燕萍..结合注意力机制和LSTM的降雨型滑坡沉降预测[J].北京测绘,2025,39(6):843-848,6.基金项目
江西省地质局科技研究项目(2022JXDZKJKJ07) （2022JXDZKJKJ07）
