Research On Classification Of Ground Objects Based On Multi-source Remote Sensing Data

In recent years,with the development of sensor technology and the launch of remote sensing satellites,the quantity and quality of remote sensing images have been significantly improved,and the difficulty of acquisition has gradually decreased.Different types of remote sensing images have been widely used in resource exploration,environmental protection,ground feature classification,military and other fields.Because of the difference in the resolution and imaging principle of different sensors,the application effects of different types of remote sensing images are both good and bad.In the field of remote sensing image classification,optical remote sensing images based on high resolution have achieved certain success.However,because optical images are easily affected by the natural environment and different types of ground objects have similar spectral information,it is difficult to classify the types of ground objects.Synthetic aperture radar is not limited by the natural environment,and can obtain remote sensing data all-weather,thus obtaining the backscattering characteristics of different ground objects.In order to make comprehensive use of the advantages of different types of remote sensing images,this paper combines the optical remote sensing images and synthetic aperture radar(SAR)remote sensing images to realize the information complementarity between the images,and then carries out the ground object classification research on the optical remote sensing images,SAR remote sensing images and their combined images through the depth learning method.The main work and research results of this paper are as follows:(1)Data preprocessing and dataset production.The research areas selected in this article are Jingyuan County and Longde County in Ningxia Hui Autonomous Region.This region has a rich variety of terrain types and is a tourist destination in Ningxia.Studying the classification of terrain features in this region is of great significance for urban development and ecological civilization construction.In terms of data,Sentinel-1 SAR data,Sentinel-2 optical data from the European Space Agency(ESA),and Digital Elevation Model(DEM)data from the Geospatial Data Cloud were selected as the data sources for this article.After the data download is completed,the remote sensing data is first preprocessed using ENVI 5.3 and SARscape plugins,and then cropped and combined with data bands.Due to the average elevation of the research area being between 1608m and 2942m,it is necessary to eliminate terrain distortion in the corresponding area’s DEM data during the preprocessing operations of geocoding and radiometric calibration.Finally,Label me software is used to annotate and enhance the three cropped images,providing technical support and data preparation for subsequent land object classification experiments.(2)Research on multi-source remote sensing data ground object classification based on fully convolutional neural networks.The research content of this section is based on pixel level ground object classification.The classification network selects representative U-Net and DeepLabV3 networks from convolutional neural networks,and uses these two networks to conduct ground object classification experiments and analysis on Sentinel-1 dataset,Sentinel-2 dataset,and their band combination dataset.The classification accuracy of U-Net network on Sentinel-2 optical dataset is 87.21%,on Sentinel-1 SAR dataset is 83.68%,and on a combination of optical and SAR dataset is 90.38%;The classification accuracy of DeepLabV3 network on Sentinel-2 optical data is 73.69%,on Sentinel-1 SAR data is 74.53%,and on combined optical and SAR data is 80.23%.(3)Research on multi-source remote sensing data ground object classification based on improved UNet network.Although U-Net and DeepLabV3 networks have achieved good classification results on three datasets,there are still cases of misclassification and unsatisfactory recognition accuracy for small area terrain types.In view of this situation,based on the U-Net network,this paper proposes the C-S-UNet network from three aspects:feature extraction structure,activation function and Max Pool.The experimental results show that the accuracy of the combined dataset is better than that of the single dataset,and the verification accuracy of the network proposed in this paper in the combined dataset,Sentinel-1 dataset and Sentinel-2 dataset reaches 90.89%,86.09%and 89.05%respectively,compared with the UNet,Increased by 0.51%,2.41%,and 1.84%respectively.