Research progress of high-resolution remote sensing image scene classification
With the rapid advancement of remote sensing technology,the resolution of remote sensing satellites is improving,the number of spectral bands is increasing,and revisit periods are contracting.This progression empowers researchers to access more valuable data and information from remote sensing images.Concepts,such as remote sensing big data,remote sensing foundation models,and smart cities,have successively emerged in recent years,imposing increased demands on the intelligent extraction technology of massive remote sensing data,particularly regarding remote sensing image information.As an indispensable element of intelligent information extraction technology applied in fields,such as land use and cover,national land resource surveys,natural disaster observation,agricultural yield estimation,and forestry protection,remote sensing image classification exhibits substantial practical importance.Remote sensing image scene classification has been introduced in this context.The objective of scene classification in remote sensing images is to comprehensively and semantically categorize each given remote sensing image.This task entails summarizing and analyzing the extracted feature information at a high level and assigning different labels to areas of interest based on their features.In contrast with natural images,although they contain features,such as color,texture,and shape,remote sensing images encounter more challenges in classification due to the intricate scene content resulting from the overhead perspective,weak texture,and color information caused by low resolution.Nevertheless,as one of the technical means in remote sensing applications,remote sensing image scene classification technology plays a pivotal role in the development of practical application technologies.After years of development,numerous comprehensive review studies on remote sensing image scene classification have been conducted locally and abroad.However,the recent surge in remote sensing big data has introduced new challenges into scene classification.The ongoing evolution of deep learning technology,particularly the widespread application of Convolutional Neural Networks(CNNs)and transformers,has resulted in significant advancements in remote sensing image scene classification.In this context,self-supervised learning,as a method that is independent of annotated data,has become indispensable in the field of remote sensing image scene classification.Foundation models based on self-supervised learning have been successfully implemented in scene classification,presenting innovative solutions to this field.As the volume of remote sensing data continues to increase,the dataset scale for remote sensing image scene classification is expanding rapidly,giving rise to increasingly intricate classification tasks.Remote sensing image scene classification datasets are swiftly progressing toward the integration of multiple sources,the incorporation of multiple labels,and the inclusion of large-scale samples.Drawing from the findings of the current literature survey,this study systematically compiles a summary of deep learning methods within the domain of remote sensing image scene classification.Encompassing CNNs,visual transformers,and generative adversarial networks,this overview also introduces representative datasets and foundation models since the inception of scene classification.Several classical scene classification methods have undergone evaluation across various benchmark datasets.In addition,this study delves into primary challenges and prospects,paving the way for further research in the classification of scenes in remote sensing images.
万方数据
维普
