Research On High-resolution Remote Sensing Land Use Classification Method Based On Small Sample Image Enhancement And Semantic Segmentation

Land use information is an advanced expression of human activities in space,and it is the foundation for understanding the dynamic changes of the earth’s surface and the interaction between social and ecological systems.It is crucial for urban planning and environmental impact assessment.Accurate classification is the main way to obtain land use information and a necessary means to deepen the understanding of the relationship between human activities and the spatial environment.The use of remote sensing images to achieve high-precision land use classification mapping has been a research hotspot.In recent years,deep learning methods represented by convolutional neural networks and attention mechanisms have been introduced into the field of land use classification using remote sensing images,not only achieving rapid development in computer vision but also advancing the field of land use classification.Meanwhile,high-resolution land use datasets based on remote sensing images have also emerged.However,in these datasets,the problem of unbalanced sample quantities of land-use categories exists due to realistic reasons,leading to poor classification accuracy of small-sample land-use categories.Furthermore,the classification method using convolutional neural networks lacks sufficient learning of contextual semantic information,which also leads to poor classification results.Therefore,this article conducts in-depth research on these two key issues,including the following aspects:(1)In order to solve the problem of unbalanced land use data,increase the number of small sample land-use categories,and improve the diversity and reliability of the dataset.This article first improves Gau GAN by comparing it with two GAN image enhancement methods,pix2 pix HD and Gau GAN.The improved Gau GAN generates land use data with richer detail information,higher spatial resolution,relative stability in the synthesis process,and conforms to the distribution of real samples.Secondly,compared with conventional image enhancement methods,although the accuracy of individual categories is not high,the accuracy of smallsample land use categories is significantly improved,and the overall accuracy is the best.Finally,the improved Gau GAN has good adaptability to different deep learning classification networks,and the accuracy of small-sample land-use categories and overall accuracy in all networks have been significantly improved.Therefore,in the deep learning framework,the method proposed in this article can be used as an effective solution to the problem of unbalanced land use data.(2)General machine learning methods and image semantic segmentation models are difficult to meet high-precision land use classification requirements due to their inability to fully utilize spatial contextual semantic information.To overcome this difficulty,this article proposes a land use classification model that combines conditional random fields and attention mechanisms on the basis of data enhancement,named DADNet-CRFs.First,the convolution method in the U-Net network is modified to a dense convolution module,and hole pyramid pooling module,spatial position attention module,and channel attention module are merged at appropriate positions in the network to form DADNet.Then,DADNet segmentation results are used as prior conditions to guide the training of conditional random fields.Finally,compared with FCN-8s and Bi Se Net networks,the land use classification accuracy of the network model proposed in this article is higher,and the visualization of classification results is also better.Therefore,the land use classification model proposed in this article is more conducive to achieving high-precision land use classification.