Research On Remote Sensing Image Semantic Segmentation Based On Deep Learning And 3D Scene Generation

In recent years,with the continuous development of remote sensing technology,the application of remote sensing images in urban planning,agricultural planning and military training has great prospects.At the same time,with the continuous development of deep learning,the ability of its network to extract image features is continuously enhanced.Therefore,how to use deep learning to accurately and quickly segment high-precision remote sensing images and take full advantage of the segmentation results is of great significance.This paper mainly studies the semantic segmentation of remote sensing images and 3D scene modeling based on deep learning,aiming at the complex problems of large and small objects in highresolution remote sensing images and the complex problems of 3D scene modeling based on2 D data.The specific work of this paper is as follows:（1）This paper proposed a remote sensing image semantic segmentation method based on improved Deep Lab v3.By improving a single upper sampling layer and using the residuals obtained in the backbone network for multi-layer upper sampling,the semantic integrity of the image in resolution is guaranteed;At the same time,the expansion rate of 4-layer expansion convolution in ASPP layer is modified to improve the segmentation effect and robustness of the network to small objects.The experimental results show that the improved Deep Lab v3 semantic segmentation algorithm achieves 94.92% and 98.01% accuracy of MIou and pixels on the self-made data set,which is 3.77% and 2.40% higher than the original algorithm respectively.It not only has higher accuracy,but also has better robustness to all kinds of terrain segmentation;At the same time,the segmentation method proposed in this paper is also better than the current mainstream segmentation methods such as U-Net,Seg Net,HR-Net and DANet.（2）This paper proposed a method to generate buildings.Firstly,extracting the outline of each building by using Canny edge detection algorithm.Then,match the model top view projection in the model library with the actual building projection by using a hash algorithm,so as to make the generated urban scene more realistic.（3）This paper proposed a 3D scene modeling method based on Unity3 D which uses the location information obtained from different kinds of mask images obtained by semantic segmentation to model with the elevation map of remote sensing image.This method uses terrain and mesh components in Unity3 D to automate the modeling process through script editing.Use the elevation data of remote sensing map to establish the basic terrain,use the segmented mask image to locate different materials of the scene,add different maps for different materials,and adjust the shader of water surface map to make the water surface more realistic.The experimental results show that the above method can correctly and quickly generate three-dimensional scenes that are more consistent with the corresponding remote sensing images,and the generation time of roads,vegetation,water systems and buildings are less than 3 mins.