Research On Semantic Segmentation Of Remote Sensing Image Based On Deep Learning

Satellite remote sensing image reflects the spatial distribution of electromagnetic radiation characteristics of ground objects.It has been widely used in environmental monitoring,urban construction and other fields,exhibiting both research value and economic benefits.Satellite remote sensing image contains abundant ground information so that the processing procedures are very complex.How to realize intelligent interpretation of remote sensing image has always been difficult for academia and industry.Image segmentation is the basis of remote sensing image processing,the key factor of interpretation quality.With the improvement of software and hardware level,the remote sensing images become more and more complex,limiting the application of traditional segmentation methods.In recent years,image segmentation methods based on deep learning have been developed rapidly.It is of great practical significance to apply deep learning to solve the intelligent interpretation of remote sensing image.The purpose of this study is to investigate the deep learning based semantic segmentation technology for high-resolution remote sensing image.It has been discussed in the stages of remote sensing image acquisition,quality improvement of the image,and semantic segmentation network design and training,with more focuses on the optimization strategies.Our data sources include open image competition datasets and Harbin satellite remote sensing image datasets.In the stage of quality improvement,we have done denoising,haze removal and super-resolution processing for remote sensing images.Specifically,spatial denoising methods such as gaussian filtering and mean filtering have been applied in the phase of denoising processing;the dark channel prior method has been used in the phase of haze-removal processing;and an image super-resolution method based on Boundary Equilibrium Generative Adversarial Networks(BEGAN)is proposed in the phase of super-resolution processing.In the design and training stage of semantic segmentation network,we briefly describe the basic principles and components of deep convolutional neural network,and then describe the proposed semantic segmentation network based on U-Net improvement.Taking DeepLab v3 parallel training as an case study,we discuss the optimization strategy based on regularization,transfer learning and ensemble learning.Using the improved U-Net in a new training,90%of the test accuracy was achieved on the open dataset "CCF Satellite Image AI Classification and Recognition Competitio".Fine-tuning of the training by using DeepLab v3 ensembled with the weighting factors of pre-training model of ResNet-50,93%of the test accuracy was achieved on the Harbin remote sensing image dataset.The experimental results show that the proposed methods achieved high segmentation accuracy and good generalization ability,suitable for the application in practical engineering.