Research On Semantic Segmentation And Object Detection Of Remote Sensing Images Based On Deep Learning

In recent years,the development of remote sensing science and technology has been rapid,and the resolution of optical remote sensing images has been increasing,which can be used to accurately observe various targets and all kinds of features on the earth’s surface,and therefore has a wide range of applications in such fields as intelligent agriculture,disaster assessment and urban planning.With the significant breakthroughs in theories and methods of deep learning,it is of great value to use deep learning methods for target detection and semantic segmentation of massive high-resolution remote sensing images in practical scenarios.The main research of this paper is to use deep learning methods for target detection and semantic segmentation of remote sensing images,and to improve the deep learning methods to improve their interpretation accuracy of remote sensing images.The main work of the paper is as follows:(1)The NDVI-RSU-Net model is proposed.Based on the U-Net model,the RSU-Net model is constructed by adding the residual network and attention mechanism.In order to explore the best input band combination of RSU-Net model,the NDVI-RSU-Net model is constructed by adding normalized vegetation index(NDVI)to the input feature channels.The model was used to extract land cover types from the remote sensing images of GF-2 and the extraction results were compared with those of FCN,U-Net,LinkNet,VNet,ResU-Net and RSU-Net models to verify the effectiveness of the model in extracting land cover types.Based on the experimental results,the following conclusions were drawn:(1)The RSU-Net model achieved the best classification results with the overall segmentation accuracy(MPA)and the mean intersection ratio(MIOU)of 87.24%and 67.69%,respectively,which were higher than those of FCN(84.38%and 65.36%),U-Net(85.34%and 65.10%),LinkNet(81.89%and 55.43%),VNet(79.63%and 57.69%),ResU-Net model(86.37%and 66.09%),and RSU-Net model(87.24%and 67.69%).(2)The band combination of visible,near-infrared and normalized vegetation index(NDVI)was the best input band for RSU-Net.The overall segmentation accuracy and average intersection ratio of 88.36%and 71.18%were higher for this method to extract land cover types than the RSU-Net segmentation method without adding NDVI.(2)A remote sensing image target detection model with improved YOLOv5 is proposed.Firstly,the small target information contained in the feature map is less or disappeared due to the down-sampling of convolutional neural network,and feature reuse is introduced to increase the small target feature information in the feature map;secondly,the feature fusion network of EMFFN(Efficient Multi-scale Feature Fusion Network)is used in the feature fusion stage instead of the original Secondly,in the feature fusion stage,the original PANet(Path Aggregation Network)is replaced by EMFFN(Efficient Multi-scale Feature Fusion Network),which efficiently fuses feature map information at different scales by adding jump connections and cross-layer connections;finally,in order to cope with the problem of poor detection due to complex backgrounds,a Bidirectional Feature Attention Mechanism(BFAM)is proposed,which includes both channels and pixels,to improve the detection performance.Feature Attention Mechanism)to improve the detection effect of the model in complex backgrounds.Experiments are conducted on DIOR dataset and RSOD dataset,and the mAP(Mean average precision)is 87.8%and 96.6%,respectively,which are 5.2%and 1.6%better than YOLOv5,and both are higher than other classical target detection models.The experimental results show that the model proposed in this paper has better detection effect.