Research On Semantic Segmentation Of Large-scale Collapsed And Slide Mass In High Resolution Image Based On Deep Learning

With the development of China's aerospace industry,the use of remote sensing technology to accurately and quickly acquire geospatial information has gradually become a major method of acquisition.This thesis focuses on the topic of semantic segmentation of high-resolution remote sensing images based on deep learning.Based on systematically reviewing the literature on semantic segmentation methods and identification of collapsed and slide mass at home and abroad,it is difficult for traditional remote sensing image classification methods to achieve better classification effect,and the situation that deep learning has made breakthrough progress in natural image classification,combined the semantic segmentation task in deep learning,high-resolution remote sensing image information extraction and collapsed and slide mass recognition,and carried out research attempts.The main research contents and results of this thesis are as follows:(1)Considering that a small area cannot reflect the complete characteristics of a large collapsed and slide mass,this thesis cuts the image of the 2560 * 2560 size area and then resize it to 256 * 256 size.A semantic segmentation dataset for high-resolution remote sensing images of collapsed and slide mass in a large scale range was produced.Ensure the integrity of the image features of large collapsed and slide mass.In this thesis,using this data set has achieved good experimental results.(2)In order to explore the feasibility of applying deep learning methods to the task of large collapsed and slide mass semantic segmentation of high-resolution image.This thesis uses several classic deep learning semantic segmentation models and the models based on VGG16,transfer learning,encoding and decoding,jump connection methods proposed in this thesis,attempted to recognition large collapsed and slide mass in high-resolution remote sensing image.Experiments show that several classic deep learning methods and the methods in this thesis can effectively identify large collapsed and slide mass on high-resolution images.It shows that it is feasible to use deep learning method to segment semantic large collapsed and slide mass of high-resolution images.(3)In order to demonstrate the effectiveness of the method in this thesis,this thesis compares the experimental results of several classic deep learning semantic segmentation methods with the method of this thesis according to the method of holistic and then partial,qualitative and quantitative.The results show that the results predicted by the method in this thesis are significantly more complete,with fewer broken plaques and higher internal consistency.At the same time,the method proposed in this thesis has an F1 score of 82.02%and an Intersection over Union(Io U)of 70.55%,which is superior to other methods.This shows that the method proposed in this thesis has the best effect when considering the accuracy rate and the recall rate,and the overall effect is also the best.It provides a method reference for large-scale geological disaster assessment,emergency rescue and other work,and has certain research and application value.