The Research On Remote Sensing Image Ground Object Identification Method Based On Semi-supervised Deep Learning

Remote sensing image ground object idetification is related to important fields such as smart cities,precision agriculture,and environmental protection.Semantic segmentation,as one of the most advanced methods for object identification in remote sensing images,has achieved fruitful results based on the application of large-scale annotated datasets.However,the fact that the amount of remote sensing image data is huge leads to not establishing a large-scale annotated dataset similar to natural images.Semi-supervised learning aims to use a small amount of labeled data and a large amount of unlabeled data to optimize the model to improve its performance on specific tasks.Therefore,the study of semi-supervised remote sensing image feature recognition technology has research significance and application value.The core of semi-supervised semantic segmentation lies in how to mine the rich semantic features of data and how to establish the intrinsic relationship between unlabeled data and labeled data.The existing work has the following shortcomings: 1)Image-level confidence assessment is treated as pixel-level confidence assessment.For high-resolution remote sensing images,insufficiently refined confidence assessment will cause model optimization to be negatively affected by more pseudo-label noise.2)Constraining the consistency between predictions of unlabeled data under different perturbations can improve the performance of the model on specific tasks.The strong disturbance constraint is not stable enough,which makes it difficult for the model to maintain the correct optimization direction,and eventually leads to the phenomenon of model degradation.3)Under the condition of limited labeled data,the lack of feature encoding of model still limits the improvement of semi-supervised learning performance.Focusing on the above-mentioned problems of semantic segmentation of semi-supervised remote sensing images,this paper proposed a semi-supervised semantic segmentation method based on consistency self-training and a semi-supervised semantic segmentation method based on Visual Transformer(Vi T)on the basis of analyzing and summarizing the existing basic theoretical methods of semi-supervised learning,and developed a remote sensing image processing system to verify the proposed method.The main contents and innovations of this paper are as follows:1)In order to make up for the lack of refined confidence assessment and unstable constraints of strong perturbation adopted by existing methods,this paper proposes a consistency self-training semi-supervised semantic segmentation framework.The method is based on a generative adversarial network architecture,using a semantic segmentation network as a generator to generate prediction results,and a fully convolution-like network as a discriminator for pixel-level confidence assessment of the prediction results.Aiming at the lack of refined confidence evaluation,this method uses the pixel-level comparison between the annotated data and its predictions as a supervision signal for the training of the discriminator,so as to learn the semantic information of the correctly predicted points.Aiming at the insufficient stability of strong perturbations,this method introduces seven kinds of strong perturbations into the feature domain,and filters out the misclassified areas that cause model training deterioration under strong perturbations on the basis of refined confidence evaluation,so as to keep the model predictions consistency.The experimental results show that the consistency self-training semi-supervised semantic segmentation method outperforms other semi-supervised semantic segmentation methods,effectively solving the difficulty of how to establish the intrinsic relationship between unlabeled data and labeled data.2)In order to make up for the insufficient feature encoding of segmentation model in semi-supervised semantic segmentation,this paper proposes a semi-supervised semantic segmentation method based on Vi T.This method follows the normal form of self-training,and designs a discriminator with a full Vi T architecture to evaluate the confidence of the prediction results,so as to select high-confidence pseudo-labels to complete the self-training of unlabeled data.This method uses the Vi T model to design the discriminator,and presets a learnable confidence token parameter,which helps the discriminator establish the connection between the local difficult example block and the global easy example block,and makes accurate confidence evaluation for the difficult example area,thereby facilitating the segmentation network to learn the decision surface of the entire data domain and improving the feature encoding ability of the model.The experimental results show that the Vi T-based semi-supervised semantic segmentation method effectively solves the difficulty of how to mine the rich semantic features of data.3)This paper mainly uses python as the main programming language to design and develop a remote sensing image processing system.The front end adopts the Py Qt5 framework and is deployed on the Windows 10 platform.The back end adopts the producer-consumer model to realize concurrent data communication and data processing,and is deployed on the Ubuntu16.04 platform.On the remote sensing image processing system platform,the proposed semi-supervised semantic segmentation method has been fully verified.