Research On Feature Extraction Method Of Coal Mine Area Based On The Fusion Of UAV Images And Elevation Data

Coal mine resources are of great significance to the sustainable development of China's economy,and accurate acquisition of mine site feature information is a prerequisite for safe coal mine production.The traditional survey method based on manual field survey is time-consuming and labor-intensive,with high cost and low accuracy,etc.To address this problem,this paper investigates the method of extracting mine features based on deep learning of UAV mine HD images,and on this basis,it investigates the method of fusing mine DSM elevation data to increase the accuracy of mine feature extraction.In this paper,three methods are proposed to fuse the mine area HD image data and DSM elevation data.Firstly,the pixel fusion of RGB images and DSM data is performed using imitate IHS transform to obtain three-channel feature data of RGB-like images,which are input to a deep learning network for feature extraction.Secondly,without changing RGB and DSM data,they are directly processed into four-channel feature data,which are input to the deep learning network,and the data are fused at the feature layer by using the feature fusion function of the network itself.Finally,a two-branch network is proposed to extract features of RGB image and DSM elevation data separately,and fuse the extracted features of each branch for feature extraction in the final decision stage.Three current mainstream deep learning networks,U-Net,bilateral segmentation network(Bi Se Net)and high resolution network(HRNet),are used in the paper,and the extraction effects of features in mining areas between different networks are compared horizontally through experiments,and the extraction effects of the networks before and after fusing DSM elevation data are compared vertically.Meanwhile,the paper adds the Squeeze-and-Excitation block(SE Block)into the deep learning model to verify the enhancement effect of attention mechanism on the extraction of mine features.The experimental results show that the three DSM elevation data fusion methods proposed in this paper can improve the accuracy of feature extraction in mining areas compared with the original RGB image data processing,and the indexes of model recognition accuracy are all improved.Among them,the U-Net model improved the overall pixel accuracy by 9.36 percentage points and the average pixel accuracy by 63.25 percentage points compared with the RGB data for the four channels.Among the three fusion methods,the two-branch network has the highest overall pixel accuracy of 95.51% and the average pixel accuracy of 90.44%.SE Block also improves the final accuracy for most of the models,and the Bi Se Net model improves the overall pixel accuracy by 11.16 percentage points and the average pixel accuracy by 26.77 percentage points after adding SE Block.The SE Bi Se Net model segmented the best in the experiment,with an overall pixel accuracy of 95.35%,an average pixel accuracy of 93.11%,and a kappa coefficient of: 0.8301 in the test set.