A 3D Convolution Deep Learning Method For Semantic Segmentation Of Multispectral Remote Sensing Images

With the rapid development of remote sensing technology,hyperspectral remote sensing data is becoming more and more popular and widely used.Semantic segmentation based on hyperspectral data is one of the core applications of hyperspectral remote sensing technology.Hyperspectral remote sensing images have the characteristics of high dimensions and rich information,and can provide more spectral information than traditional RGB graphics in semantic segmentation tasks.However,due to technical reasons,current remote sensing satellites often cannot achieve both hyperspectral and same scale.That is,different spectra often have different spatial resolutions,which poses a challenge to semantic segmentation.On the other hand,satellites generate images in many bands.How to determine the specific bands to be used according to the specific segmentation goals,and how to make the best use of pixel information? High spectral resolution,difficult labeling,and different band scales determine the technical difficulties of the hyperspectral remote sensing image segmentation task.This paper takes water body extraction(the segmentation object is water body)as an example,and proposes a 3D convolutional neural network-based hyperspectral remote sensing image segmentation network D2S＿M3DUnet,which expands the traditional convolutional neural network into three dimensions and applies it to Sentinel 2 remote sensing image water body extraction.In this paper,the hyperspectral characteristics of Sentinel 2 data are used to input band images of different scales on different convolutional layers of the network and perform channel fusion without using the traditional resampling integration method,which can both reduce noise and use 3D convolution fully mines the information in the two dimensions of spectrum and space to improve the segmentation performance.It also uses the D2 S method to improve up-sample to obtain better segmentation accuracy and shorter training time.The experimental results show that the 3D convolutional neural network D2S＿M3DUnet has better performance than the 2D convolutional neural network in the water body extraction of Sentinel 2 remote sensing image,especially the prediction of small water bodies is more accurate.