Evaluation Of Debris Flow Disaster Susceptibility Based On Prototype Network

Debris flow disaster is one of the more frequent geological disasters in my country.It is usually caused by heavy rainfall and occurs mostly in semi-arid mountainous areas or plateau glacier areas,such as Sichuan,Yunnan,Tibet,Gansu,and other regions.Debris flow disasters have the characteristics of strong suddenness,fast flow,large flow,and strong destructive power.Once a debris flow disaster occurs,houses,roads and farmland will be destroyed,causing serious loss of life and property to the people.Therefore,how to identify the prone nature of debris flow disasters,formulate scientific and reasonable preventive measures in advance,and reduce the threat of debris flow disasters has become an urgent problem to be solved.In the investigation and research of debris flow,compared with the traditional field survey method,remote sensing technology has many incomparable advantages,such as large detection range,fast acquisition speed,high spatial and temporal precision,and contains a large amount of accurate objective information,etc.,which greatly saves the human and financial costs of investigation and research.Extracting debris flow-related factors from remote sensing to evaluate debris flow susceptibility has become a popular method,but there are still problems such as different watershed divisions,different selection factors,and large differences in evaluation results.With the development and wide application of deep learning,the neural network has made great progress in the recognition and classification of remote sensing images,but the effect of using the existing neural network model for debris flow disaster classification is relatively general.To solve this problem,this paper taking the valley as the evaluation unit,a remote sensing image dataset of valley debris flow disaster in Nujiang Prefecture,Yunnan Province was established,and a deep learning network model was rebuilt to learn data characteristics to conduct evaluation research on the susceptibility of valley debris flow.On this basis,Introducing Few-Shot Learning in deep learning as a model,and the attention mechanism is added to further improve the model performance,which improve the accuracy of the evaluation of valley debris flow susceptibility.This paper works as follows:1.Using DEM,Gaofen-1 remote sensing and Google Earth remote sensing,a remote sensing image dataset of valley debris flow in Nujiang Prefecture was established.In Arc GIS,use DEM to calculate the filling,flow direction and the river network in the research area,find out the water outlet(pour point)of each valley in the river network,and use the function of watershed to extract the mask of each valley.Using these masks extract the DEM image,Gaofen-1 remote sensing image and Google Earth remote sensing image of each valley in batches,resample the three images of each valley and combine them for channel merging to form a multi-source data remote sensing of valley debris flow Image dataset.The data set consists of all valleys in the study area,including positive samples,negative samples,and valley samples to be evaluated.2.Susceptibility evaluation of valley debris flow based on a Atrous Convolutional Neural Network model.In the case that the classic network structure cannot extract the debris flow characteristics of valleys for susceptibility evaluation,a Two Path Convolution is constructed based on the convolution block structure by combining the classic residual structure and the multi-branch structure,and adding a Atrous Space Convolution Pooling Pyramid to increase the perception field of the network,so that the newly constructed Atrous Convolutional Neural Network model can better identify the characteristics of debris flow in valleys and improve the accuracy of model predictions.3.Susceptibility evaluation of valley debris flow susceptibility based on SelfCalibration prototype network.Deep learning requires a large number of training samples,and the disaster data that can be obtained in a region is extremely limited,which is a typical Few-Shot Learning problem.When using deep learning to train few-shot data,the model is prone to overfitting and lower prediction performance.To solve this problem,the classic prototype network in Few-Shot Learning is used as the predictive model,and the coordinate attention mechanism is introduced to improve the Atrous Convolutional Neural Network structure,this improved structure is used as the feature extractor of the prototype network,and the Self-Calibration method is used to improve the calculation of the prototype network.The Self-Calibration prototype network model can better improve the accuracy of debris flow susceptibility evaluation.