Research On Lithology Classification Of UAV Image Based On Deep Semantic Segmentation

It is difficult to carry out geological mapping in high-altitude and hard-to-access access areas such as the Qinghai-Tibet Plateau.However,the rapid development of UAV remote sensing technology can quickly acquire high-resolution remote sensing image in hard-to-access areas,and can automatically identify lithology,thus assisting in large-scale geological mapping.At the same time,the traditional image classification methods are mainly based on the features extracted manually.It need not only rich expert knowledge,but also difficult to find the optimal features,which results in the low classification accuracy.Therefore,in order to realize intelligent extraction of formation lithology information,not only UAV remote sensing image data,but also more intelligent information extraction methods are needed.In recent years,thanks to the development of deep convolution neural networks,breakthroughs have been made in semantic segmentation method.The semantic segmentation method based on CNN can automatically extract image feature information in the process of network training.Through the continuous iterative optimization of the network,the optimal feature combination can be learned,so as to improve the accuracy of semantic segmentation.Practice has proved that the image segmentation accuracy based on CNN is much higher than that based on traditional semantics under the condition of large data,which provides a new technical method for automatic lithology classification of UAV images.In this paper,FCN,SegNet and Unet network are selected to extract lithology information from UAV image in Bannu Belt of Tibet.The experimental process is as follows:Firstly,acquiring the high-resolution UAV remote sensing image in the research area.Secondly,preprocessing the acquired UAV image data.Thirdly,training model and extracting lithology information.Finally,using conditional random field to classify the lithology of UAV remote sensing and post-process the classification result.The experimental results show that the lithology classification accuracy of UAV remote sensing images based on FCN,SegNet and Unet network model is 90.33%,86.69%and 82.85%,respectively.And the FCN network has the shortest convergence time amount three model,so FCN model is the optimal model of the formation lithology semantics segmentation,and satisfy the requirement of accuracy.The research results prove that the semantic segmentation method based on CNN is feasible for lithology mapping of UAV remote sensing image in Plateau area,can not only correct the boundary of different lithology in geological map,but also achieve higher precision lithology division to assist in large-scale geological mapping.The research in this paper is of great significance and practical value for carrying out large-scale geological mapping of hard-to-access areas such as high altitude of the Qinghai-Tibet Plateau.