Research On The Application Of Multi-level Encoder-decoder Networks In Land Cover Classification

With the rapid expansion of cities,the landforms of remote sensing images obtained by resource satellites have undergone earth-shaking changes in a relatively short macro time.Nowadays,land cover classification methods are faced with the following problems: on the one hand,with the rapid development of remote sensing technology,resource satellites have been able to capture high-resolution land cover remote sensing images,which contain a large number of rich details that can not be handled by traditional remote sensing classification technology;On the other hand,the imaging conditions of remote sensing images taken in different time and space vary greatly,and the interference will be very strong,while the traditional machine learning classification model lacks the ability of anti-noise and generalization.To solve the above problems,this paper introduces semantic segmentation technology in deep learning to achieve land cover classification.The specific work is as follows:First,the traditional recognition methods of high resolution remote sensing images often appear problems such as low detection rate of small targets and poor edge detail.This paper first puts forward the multi-level context information fusion network which improves the U-Net semantic segmentation network.Considering insufficient utilization of feature information,my model also proposes dense skip connections and attention modules,which effectively improve the accuracy of the model.The metrics PA and MIOU of the model have reached 99.29% and 98.88%.Second,to solve the problem of more categories and small samples in remote sensing dataset,a structural re-parametrization segmentation network is proposed.To capture the abstract texture features for locating the spatial information of small sample categories,this paper adds residual structures to the backbone network and improved it.By reparameterizing the structure,the training model and the inference model are divided to speed up the inference speed.In view of the interference between complex spectral information and the surrounding environment,this paper also uses asymmetric convolution to weight convolution kernels to extract semantic information from encoders’ output to solve the classification of multiple categories,and improves the attention module to divide the full-scale features into two branches,then enhance the information flow.The model metrics MPA and MIOU have both reached 91.36% and 85.23%,both superior to other comparison models.Meanwhile,the single picture of the whole model inference of the improved Res Net34 main trunk network is 32 ms,second only to ResNet18.