Land Cover Classification Based On Deep Convolution Neural Networks For Optical Remote Sensing Image

Land Cover Classification is one of the important applications of remote sensing images.With the country’s vigorous exploration and development in the field of remote sensing,the imaging quality of remote sensing images has been greatly improved,the resolution of images has become smaller and smaller,and the information acquired has become more and more abundant,which brings advantages to the task of land cover classification as while as higher requirements and challenges.Traditional classification algorithms can distinguish fewer categories due to the limited image resolution.As the resolution increases,more and more categories are clearly visible,and the number of categories also increases.In order to better distinguish these categories,the features of these categories must first be extracted.With the widespread application of convolutional neural networks in the image field in recent years,a new method for land cover classification based on convolutional neural networks is expected to emerge.Based on the semantic segmentation algorithm of deep convolutional neural networks in natural images,with analyzing the difference and connection between remote sensing images and natural images,this paper proposes a series of improved methods for land cover classification,and obtains satisfactory segmentation results.The specific content of this paper is as follows:1.This paper proposes a land cover classification method based on attention mechanism and multi-scale features.This method is based on the Deep Labv3+ network structure,which is the state-of-art algorithm for semantic segmentation of natural images at present.Aiming at the complex scenes of remote sensing image,using the attention mechanism can make the feature extraction structure pay more attention to the typical elements in the complex background.For the problem of the differences in size among different typical element is large,a new multi-scale feature extraction module has been redesigned to adapt to different size elements.In order to obtain better model parameters even in the case of uneven data distribution,a new method to calculate the category weight is used,and effectively solves the problem of category imbalance.2.This paper proposes a land cover classification method based on high-resolution features and dense upsampling convolution.This method draws on the novel structure of the HRNet network,which always maintain a high feature resolution during the feature extraction process,which is very helpful for feature extraction and feature recovery of small target categories.In order to improve the segmentation results of the algorithm,after the HRNet network,a decoding structure,dense upsampling convolution,that completely relies on parameter learning is connected.At the same time,in order to better utilize the extracted high-resolution features,edge segmentation loss is used to weight fusion the Focal Loss to further improves the accuracy of edge segmentation.3.This paper proposes a remote sensing image generation method based on generative adversarial network.This method uses generative adversarial network to realize the migration process of generating real remote sensing images from image segmentation labels,and can effectively solve the problem of insufficient data in land cover classification task,thereby improving the accuracy and robustness of land cover classification.