Study And Application Of Categories Representation Augmentation In Multiple Supervised Remote Sensing Imagery Segmentation Tasks

Remote sensing imagery segmentation is one of the important tasks in remote sensing imagery understanding.In the application of remote sensing Earth observation data,using imagery segmentation methods can obtain various types of land surface information,including ecosystem changes,urban area expansion and natural resource distribution.The use of optical remote sensing imageries for surface information extraction has the advantages of wide detection area,high real-time and large amount of information,which has high economic and social values.However,archived remote sensing data usually come from different loading platforms or different sensors,and there is a lack of sufficient labeling knowledge for batch interpretation of multi-source data within the same framework.On the other hand,the diversity of Earth observation data distribution makes it difficult to extend the existing information extraction and interpretation methods to the interpretation tasks of image data with different distributions.Based on these problems,this paper develops a segmentation model based on category representation enhanced query and explores a lightweight and efficiently labeled image segmentation method under multi-supervised conditions,aiming to make the remote sensing image segmentation method using artificial intelligence have stronger robustness and generalization capability as well as fast batch production capability of high-grade multi-criteria land coverage products when performing massive multi-source heterogeneous data interpretation.The main work is divided into three main areas:(1)Land surface information extraction method based on end-to-end codec architecture for image segmentation of Earth observation data.The role of the encoder component is to learn imagery features,which are simplified and learned by projecting features into a highdimensional space,and the decoder component optimizes and classifies the features learned by the encoder using label knowledge.In order to explore the intrinsic feature links between remote sensing images and achieve accurate segmentation between target categories in images,this paper designs a categories representation augmentation module for the decoder component of the semantic segmentation model.This module optimizes and updates the high-dimensional category representation by cross-attention during the learning process.Unlike the local feature extraction at the pixel level,the category representation focuses more on the global contextual information of the target category objects.The category recognition features stored in the category object representation query can help the model to enhance some of the indistinguishable target category features in the sample during the pixel-by-pixel segmentation process to achieve the target object-based imagery enhancement.In addition,the expression of the category object representation query is only related to the number of target categories but not to the amount of sample pixels,so that the number of parameters of the model as a whole will be significantly reduced,effectively reducing the time for sample interpretation and improving the model production efficiency.(2)Combining category pixel representation enhancement module to verify the feasibility of unsupervised domain adaptive segmentation in multi-source remote sensing image segmentation tasks and reduce the reliance on specific label knowledge for remote sensing image segmentation tasks.The unsupervised domain adaption semantic segmentation uses a self-training method to construct a teacher model to generate pseudo-labels,and feature alignment and reconstruction for different kinds of imageries to enhance the generalization ability of the student model through the transfer of label information.The experimental results of cross-domain segmentation of satellite imagery with two different imaging methods,GF-2and Zhuhai-1,show that the segmentation accuracy of the domain adaptation segmentation model using a categories representation augmentation module combined with Zhuhai-1 images for domain adaptation learning is 18% higher than that of the fully supervised learning method using only GF-2 images.This demonstrates that the unsupervised domain adaptation method,coupled with category representation enhancement,can be effectively applied in the same land cover classification system to interpret remote sensing images from different sources.(3)Self-supervised learning with massive archived Earth observation data and a learning framework for downstream image segmentation tasks.the self-supervised learning semantic segmentation method achieves efficient labeling by performing self-supervised pre-training with a small amount of sample fine-tuning.The self-supervised approach fully utilizes the massive amount of archived remote sensing data,as no manual labeling is required during the pre-training phase.Furthermore,only partially labeled data and a small number of training batches are required for segmentation head fine-tuning to achieve segmentation accuracy that is no less than that of fully supervised learning.As a result,the self-supervised approach can be highly efficient for learning different classification systems or different tasks within the same data distribution.In this paper,design an efficient label segmentation model for remote sensing image segmentation based on the unsupervised method of denoising and diffusion probability model combined with the category representation query module,and use 5% of labels on the Self-France dataset to approach the accuracy of fully supervised learning using all labels.It is fully demonstrated that the features learned by the self-supervised method can be effectively applied to the remote sensing image segmentation task.