Research On The Application Of Multi-source Information Fusion Technology In Landslide Hazard Assessmen

Landslide is a complex and uncertain geological disaster,which is closely related to geological conditions and environmental factors.There are significant differences in landslide displacement prediction for different geological conditions and environmental factors,therefore landslide monitoring and research should be carried out in specific areas to establish a more complete landslide displacement prediction system,in order to accurately assess the likelihood of a disaster and reduce geological disaster risks.With the continuous improvement of artificial intelligence and sensor technologies,more and more nonlinear methods and devices are used for landslide hazard assessment,but due to differences in basic data,the effectiveness of different model methods also varies.Choosing effective algorithm models is important for improving landslide warning.Traditional landslide displacement prediction have the shortcomings of specialization and singularity,making it difficult to replicate and adopt on a large scale.Therefore,this paper collected remote sensing satellite data(snow depth,coverage of high vegetation,coverage of lower vegetation and landslide displacement,etc.),ground data(surface temperature,surface runoff,underground runoff and soil moisture,etc.),and meteorological data(temperature,precipitation,snowfall and relative humidity,etc.).This paper applies multi-source information fusion technology to the neural network algorithm,conducting a series of key technical research on landslide displacement prediction,with the following main work contents:(1)For the classic landslide prediction area Bazimen,this paper analyzes multiple potential factors with multi-source data analysis.Then,the regression prediction performances of multiple algorithms such as MARS,RF,ANN,and SVM are compared,selecting SVM for further study.Four algorithm models,PSO-SVM,CV-SVM,PSO-LSSVM,and CV-LSSVM,are established under different kernel functions for landslide displacement regression prediction.Finally,the LSSVM model with the PSO parameter optimization method and RBF kernel function has the best regression prediction performance.(2)For the complex prediction environment of Zara hydropower station,this paper proposes an ASTF-LSTM model based on spatiotemporal multi-source information fusion for predicting landslide displacement.The model constructs a new ASTF-LSTM module on the basis of the LSTM module.Firstly,variable screening and grouping are conducted by combining the spatial information significance of multiple source data variables,and then,grouping feature weighting is carried out by designing input feature fusion gates.Finally,the proposed landslide displacement prediction method based on spatio-temporal multi-source information fusion of long and short-term memory neural networks is experimented quantitatively on the test set.The experimental results show that the proposed model outperforms the conventional method and the traditional LSTM method in terms of quantitative metrics,with average MSE and MAE metrics of 0.055 and 0.095 respectively,which are significantly better than the traditional machine learning method and the conventional LSTM method.(3)Although using the ASTF-LSTM model for prediction can achieve better prediction results,the long-short-term contextual memory of LSTM is implemented sequentially,but the hidden relationships between the compositional variables of complex multi-source information are nonlinear.Therefore,in order to enhance adaptive global contextual modeling for multi-source information,this paper proposes an ASTF-Transformer model.For this model,two encoder branches are used,each containing two Tr-Encoder modules for feature encoding.Then,the network performs a weighted fusion of multi-source features based on temporal saliency,which can effectively fuse features in a sequence to learn an extended depth representation of the input features in order to obtain a feature representation with spatio-temporal information embedding.Finally,the experimental results show that the proposed model outperforms traditional methods in terms of quantitative metrics,with average MSE and MAE metrics of 0.012 and 0.079,respectively,and achieves better landslide displacement prediction to some extent due to the ASTF-LSTM model proposed in this paper.