Research On Extracting Pegmatite Dike Information Of Remote Sensing Image Based On Depth Semantic Segmentation

Lithium ore is an important strategic resource.The reserves of granitic pegmatite type Lithium ore in China are huge,but due to multiple factors,it is difficult to carry out field work.The application of high-resolution remote sensing image can quickly obtain the information of prospecting indicator(granitic pegmatite dike)in the study area,which greatly speeds up the progress of exploration work.Traditional information extraction methods based on human need rich expert knowledge,usually high accuracy but can not achieve automatic extraction,sometimes will be disturbed by complex background information.In view of these shortcomings,this paper introduces deep learning semantic segmentation method to carry out research on automatic extraction of prospecting sign information.In this paper,two deep learning semantic segmentation network models,FCN network and U-NET network,were selected to carry out a comparative experiment of automatic extraction of prospecting sign information.After preprocessing the acquired high resolution Wordl View-3 satellite images,the manually extracted information and geological data were taken as prior data,and the data set was made by combining remote sensing images.The data from these data sets are input into the constructed semantic segmentation networks such as FCN and U-NET for model training.Finally,the automatic information extraction model is obtained to extract the prospecting mark information.During the experiment,the following improvements were made for the specific experimental operation and network model:(1)During the production of the data set,the original data is divided into small images of fixed size and input into the training model.High resolution remote sensing images are made full use of to increase the amount of training data and the number of training iterations and improve the training accuracy.(2)In view of the large amount of background data in the image,the Focal-loss loss function was used to replace the traditional cross-entropy loss function in the selection of loss function,and the parameters were adjusted through several comparative tests to achieve the best combination.Finally,the experimental results of FCN network and U-NET network show that the extraction accuracy of high resolution remote sensing image granitic pegmatite diff information based on FCN and U-NET network model is 98.37% and 95.34%,respectively.Both of the two network models achieve high accuracy of information extraction and can meet the requirements of information extraction.The experimental results of this paper fully demonstrate that deep learning semantic segmentation network can learn from remote sensing images and successfully extract the relevant feature information of granitic pegmatite dikes,and also prove the feasibility of using deep learning semantic segmentation method to extract geological information from remote sensing images.