Research And Implementation Of High-resolution Remote Sensing Imagery Land Cover Classification Based On Semantic Segmentation

Satellite remote sensing is an important method for researchers to observe the land cover conditions.With the continuous improvement of high-resolution satellite observation capability in China,the amount of high-resolution remote sensing image data has increased dramatically.How to extract the surface cover accurately and quickly has become a research hotspot.Most land cover classification datasets are of few categories.The research work about complex subdivision categories is lack.Deep learning is one of the most popular research interests in computer vision(CV).The neural network in deep learning has a strong nonlinear fitting ability.They can learn and extract high-dimensional features from images in order to match different CV tasks,which greatly improves the accuracy of these tasks.At present,convolutional neural network is introduced to study high-resolution remote sensing imagery,including scene classification,pixel classification,target recognition and detection.In this paper,the task of pixel classification in high-resolution remote sensing image is studied,so as to achieve regional segmentation of land cover.The 4 meter multispectral data by gaofen-2 satellite are selected in the experiment,which is divided into 15 subdivision classifications.On the basis of many semantic segmentation networks,the feature extraction ability of specific model is improved by modifying the backbone network.Combined with the experimental results,the pretraining initialization scheme is further proposed.U-Net is selected to optimize the features of remote sensing imagery,which improves the accuracy of land cover extraction.The work provides technical basic research for fine processing of high-resolution remote sensing imagery.In terms of network structure and training method,the improvements are as follows:(1)We optimize the backbone network of U-Net structure with SE-ResNeXt-50 as the encoder to enhance the ability of feature extraction.(2)Combined with transfer learning,a scene classification dataset is used to pretrain the classification network.Then the weights of classification network are loading to the backbone as initialization.It improves the overall performance of the segmentation model by the sharing of weights.(3)Bisupervised network with Pyramid Pooling Module is proposed,which provides an auxiliary output of backbone.We combine the customized loss function to effectively improve the performance of high-resolution remote sensing imagery land cover classification.The experiments shown that the pipeline proposed in this paper achieved better performance in the GID-15 dataset.Our bisupervised network obtains F1-score of 0.804,FEIoU of 0.6798,kappa coefficient of 0.7352.