Research On The Land Cover Identification Of Remote Sensing

Remote sensing has been increasingly applied to national defense,territory planning,agriculture and financial sector with its rapid development.Therefore,land cover identification on remote sensing images is a fundamental pre-processing step in many practical tasks.However,there are several drawbacks in conventional image processing by feature extraction and analysis,such as complicated feature extraction for complex images,lack of generalization and highly computational cost.Recently,deep learning-based semantic segmentation technology provides a new perspective on this field and achieves desirable performances.On the other hand,models with massive parameters are difficult to migrate in many scenes.Besides,existed approaches lack specificity in land cover identification task,which become less effective on small classes of imbalance dataset.Therefore,this thesis works on land cover identification based on deep learning and attempts to alleviate the above problems.The main contributions are listed as follows:(1)A new training and recognition framework is designed for land cover identification based on deep learning.It builds specific steps in image preprocessing,model training and image postprocessing for network prediction.In addition,a suitable cost function for land cover identification task is selected by experimental comparison.As a whole,the proposed framework become a foundation for improving the network model.(2)A novel lightweight model with less parameter is proposed to reduce the number of parameters of Linknet in semantic segmentation network.Firstly,Res Net is replaced by Res Ne Xt which acts as the encoder part in Link Net.Then it introduces depthwise separable convolution as a substitute for normal convolution inside building block’s group convolution of Res Ne Xt.It results in a new model called Sep-Link Net that has fewer parameters.Experiments show that the proposed Sep-Link Net reduces the amount of network parameters by 5.9% compared with Link Net.Meanwhile,it maintains the accuracy and improves MIo U and MPA by 2.89% and 3.01%,respectively.(3)An improved model is proposed to deal with the shortcoming in PSPNet,which shows poor performance in small categories of unbalance land cover dataset.This thesis introduces the encoder-decoder structure to conventional PSPNet framework,and optimizes the original deconvolution network in the decoder part.The proposed ED-PSPNet model improves the recognition effect of the model in small categories.Experimental results show that,compared with PSPNet,MIo U and MPA in ED-PSPNet is improved 1.68% and 2.19% respectively.Besides,in five categories with the smallest amount of data in the dataset,EDPSPNet has a better performance,which verifies its effectiveness to promote the identification performance in small categories of land cover imbalance dataset.