Land Cover Classification Of High Resolution Remote Sensing Images Based On Deep Learning

Satellite images are highly structured data source.The combination of deep learning and the analysis of remote sensing data is the focus of current research.Researches in this area could have critical impact to the way we understand our environment and lead to major breakthroughs in global urban planning or climate change research.Land cover classification is the cornerstone of applications above.It aims at classifying each pixel in a satellite image into a particular land cover category.In view of the characteristics of satellite remote sensing images such as large size and high pixels,the idea of using atrous convolution instead of pooling layer is first adopted.This can expand the convolution kernel of the neural networks which expand the kernel's receptive field without affecting the segmentation accuracy.In view of the characteristics of different sizes of remote sensing images in the same category,ASPP was introduced in the Encoder-Decoder module to obtain global and local features.In the final result,a breadth-first search based algorithm is used to design a fine tuning method for adjacent blocks,which reduces the fragment brought by image segmentation.In order to improve the consistency of the segmentation edges and reduce memory consumption,a global-local two-branch cooperative network structure was designed.In the global branch,the entire remote sensing image is down-sampled and then segmented,and the coarse-grained segmentation results are obtained while retaining the context information.In the local branch,the remote sensing image is cropped and then segmented to obtain fine-grained segmentation results.The two branches were combined by deep feature mapping and regularized branch fusion.According to the unbalanced distribution of categories in dataset,a linear task division strategy was designed.Coarse-grained segmentation was implemented in the global branch.Then an object detection was followed.The fine-grained segmentation was only carried out in the area within the detection frame.The network can improve the segmentation accuracy,reduce real-time memory consumption.According to the experiment on the Deep Globe and other datasets,the model proposed in this paper can achieve a satisfying remote land cover segmentation.What's more,the model is more memory-efficient than the mainstream segmentation networks.It can promote remote sensing image segmentation research while has value in land cover classification area.