Research On Application Of Agricultural Remote Sensing Image Processing Based On Deep Adversarial Learning

Rural settlements are the carriers of rural production and life.Rural housing is the main element of rural residential land.Therefore,optimizing the layout of rural settlements and intensifying land use play a vital role to reconstruct rural space and promote rural revitalization.With the rapid development of high spatial resolution remote sensing and machine learning technique,it is possible to accurately identify and quickly map rural residential buildings.This paper has investigated the segmentation of rural settlements based on sub-meter high spatial resolution remote sensing image and deep learning.According to analyzing the characteristics of high-resolution images and the existing problems of image segmentation,this paper has introduced adversarial mechanism to further improve the performance of deep learning segmentation model.The main research content of this paper mainly includes the following three parts:(1)For remote sensing image segmentation,this paper has studied three classical image s egmentation models,including FCN,Encoder Decoder and UNet,compared their characteristi cs of the three network structures,analyzed and verified the functional modules that could im prove the effect of image segmentation through experiments.Experimental results have show n that among the three classical segmentation models,unet could significantly improve the se gmentation accuracy by using cross connection and deconvolution operator.(2)Aiming at the problem of over-fitting and under-fitting of traditional convolution network based on the loss function of L1 norm,this paper has introduced the adversarial mechanism and proposed a deep adversarial network called UGAN,which has used U-Net as the generator and a full convolution network with five layers as a discriminator inspired by patchGAN.Experimental results have demonstrated that the proposed network can effectively distinguish buildings from background and obtain more accurate building segmentation results compared with traditional full convolution networks.It has shown that the adversarial mechanism can further optimize the representation ability of the objective function of generator.(3)Aiming at the instability problem in the training process of deep countermeasure network based on UGAN,this paper has further optimized the objective function of adversarial network,and proposed a WGAN-gp segmentation model based on the penalty regular term of gradient;Meanwhile,this paper has also proposed an improved WGAN-div model based on Wasserstein divergence to alleviate the Lipschitz continuity constraints in WGAN-gp model.Experimental results have illustrated that WGAN-div segmentation performance is better than WGAN-gp model.In summary,this paper has studied deep learning segmentation model for high-resolution agricultural remote sensing images.Experimental results have shown that the proposed methods in the paper can segment buildings from complex background more accurately.It has an important practical significance for agricultural informatization.