Research On Scene Classification Of High-resolution Remote Sensing Images Based On Deep Learning

The classification and recognition of high-resolution remote sensing images have significant importance in understanding and transforming nature.With the development of technology,remote sensing technology has a broader and more detailed application.Therefore,it demands higher recognition accuracy for high-resolution remote sensing image scenes.The details in high-resolution remote sensing images are abundant,and the structures of the objects are complex and diverse.Traditional algorithms are difficult to break through in terms of accuracy,and manual identification is expensive and unsustainable due to the massive amount of images.Deep neural networks,inspired by the human brain,have opened up a new era.Unlike traditional algorithms,the semantic information obtained by deep networks can greatly improve the classification accuracy of high-resolution remote sensing images.In addition,the replicability of deep networks makes their initial training investment cheaper than the uninterrupted cost of human identification.Therefore,utilizing deep neural networks for high-resolution remote sensing image recognition has become popular.This article explores the understanding modes of deep networks for high-resolution remote sensing images through experiments on remote sensing datasets.It proposes corresponding methods and strategies to improve the network’s understanding of images and thus improve recognition accuracy.The main research includes the following:(1)To address the problem of large intra-class difference and high inter-class similarity in high-resolution remote sensing images,a high-resolution remote sensing image scene classification method based on multi-feature fusion and Fisher criteria is proposed.Highresolution remote sensing images have large intra-class differences and high inter-class similarities,and the extracted features reflect this characteristic.The Fisher criteria effectively solve this problem by maximizing the ratio of inter-class and intra-class scatter matrices.In addition to the features extracted by deep convolutional networks,the method also combines the texture and color features extracted by traditional methods to obtain more comprehensive features of high-resolution remote sensing images.Experiments prove the effectiveness of this method.Experiments on larger datasets and the selection of other backbone networks also demonstrate the wide applicability of this method.Moreover,comparative experiments show that fusing lower-dimensional features can reduce running(inference)time while achieving better results.(2)To address the problem of incomplete network understanding of high-resolution remote sensing images,a high-resolution remote sensing image classification method based on the fusion of commonality and individuality information is proposed.Visualizing the network’s interest regions in images through class activation heatmaps shows that the network’s understanding of high-resolution remote sensing images is often different from our experiential judgment.In different image categories,foreground and background can both assist the network in judging.For example,when identifying a baseball field,the network’s interest region is concentrated on the grass rather than the usual batting area.This inspires us to fuse commonality and individuality information to enable the network to obtain a more comprehensive understanding and achieve higher accuracy.Experimental results show that fusing commonality and individuality information can usually achieve higher recognition accuracy on different networks and datasets.(3)To address the problem that the complexity and diversity of objects in highresolution remote sensing images lead to incomplete network understanding,relevant solutions are explored.Different network structures have different understandings of the same image,and this difference affects recognition accuracy,especially in high-resolution remote sensing images.This is mainly because high-resolution remote sensing images have complex and diverse object categories and more abundant details.For example,when distinguishing between medium-sized and dense residential areas,networks with higher recognition accuracy pay more attention to the medium-sized residential areas’ surroundings.This article proposes an improved residual network structure that improves the network’s understanding of images and achieves higher recognition accuracy.