Study On Gully Classification Of Debris Flow Disaster Based On Knowledge Distillation

Debris flow disaster is one of the most frequent geological disasters in China.It is usually caused by heavy rainfall and occurs in semi-arid mountainous areas or plateau glacier areas,such as Sichuan,Yunnan,Tibet and Gansu provinces.Debris flow disaster is characterized by large scale,serious harm and wide crisis area.Once debris flow disaster occurs,it will bring serious loss of life and property to the residents in the disaster area.How to identify the gully of debris flow disaster efficiently and accurately,formulate corresponding preventive measures in time and reduce the threat of debris flow disaster has become an urgent problem to be solved.Compared with traditional field investigation methods,remote sensing technology has many incomparable advantages in debris flow hazard identification,such as dynamic monitoring,high spatial and temporal accuracy and low cost.With the development of deep learning and its wide application in various fields,neural networks have made great breakthroughs in remote sensing image recognition and classification.At the same time,the number of layers of neural network is deepening,the structure is more complex,the number of parameters increases exponentially,huge parameters need to consume a lot of system resources.To solve this problem,this paper introduced the idea of knowledge distillation,selected Resnet101 and Resnet18 models with good classification performance,combined with the attention mechanism and model pruning method,carried out experiments under the data set of debris flow disaster gulfs in Yunnan Province.Based on the above steps,a gully classification model of debris flow disaster is proposed based on the integration of pruning,CBAM and knowledge distillation.The work of this paper is as follows:1.The image data set of debris flow disaster gullies in DEM and remote sensing images is established.Firstly,the DEM images are filled with depressions and the flow direction is calculated by Arc GIS software.Then,by manually labeling the drainage outlet points,MATLAB software is used to extract the gullies corresponding to the drainage outlet points in DEM images.Finally,the mask processing function of Arc GIS software is used to obtain remote sensing images of corresponding positions from extracted DEM images.2.Based on the concept of knowledge distillation,multiple groups of experiments were carried out under RSSCN7 data to obtain a combination of knowledge distillation models with better performance.Firstly,six traditional convolutional neural networks Alex Net,VGG16,Googlenet,Resnet18,Resnet50,Resnet101 were divided into teacher and student models according to different performance,and then the optimal teacher and student combination was obtained according to the analysis of experimental results.The DEM data extracted from the best teacher-student combination model are compared with the remote sensing data of four channels.Through the experimental analysis,it is concluded that the near-infrared channel image is more suitable for image classification.Finally,experiments are carried out at different temperatures to select the best temperature value for debris flow data classification,and dark knowledge visualization is carried out under RSSCN7 data.3.Construct a knowledge distillation classification model based on the combination of model pruning and attention mechanism.Firstly,several common convolutional neural networks and residual networks are added into the attention mechanism module to analyze its feasibility.Then experiments are carried out on the attention modules with different functions to analyze their differences in extracting features.After the attention mechanism was used to improve the classification accuracy of the model,the model was pruned at the channel level to reduce redundant parameters and reduce the model size.Finally,knowledge distillation technology is used to transfer the knowledge from the teacher model to the pruned student network to improve the model recognition rate again.Finally,it can improve the efficiency of student network classification and reduce the model size.