Automatic Interpretation Of Winter Wheat From Remote Sensing Image Based On Deformable Full Convolutional Neural Network

Winter wheat is the most important agricultural product in the north of China.To obtain the growth and distribution of winter wheat timely and accurately is conducive to the estimation of agricultural production,and also can provide a strong basis for the agricultural policy and distribution of agricultural products in China.GF-2 Remote Sensing Image Auto-Interpretation Technology can obtain the spatial distribution of Winter Wheat in real time.The essence of remote sensing image interpretation and semantic segmentation is pixel-based classification.Therefore,this paper uses the semantic segmentation method in deep learning to improve the classical semantic segmentation network(U-Net),and designs a set of automatic interpretation scheme for winter wheat remote sensing image based on the deformable full convolution network model(DFCN).First,in order to solve the problem of image errors and unclear characteristics of remote sensing images,pre-processing and feature enhancement of the original images were made,2330 training set images of 256 ×256 size and 582 test set images of 256 ×256 size were made.Second,in order to solve the problem of missing image dimensions caused by sliding convolution,a filling operation was added to the convolution process;then,a network was added.Layer depth is used to characterize the features of remote sensing images.Finally,a trainable offset is added to the convolution process to deform the convolution,and an adaptive sensing field is obtained to extract the geometric deformation features of winter wheat.After training and adjusting the parameters of DFCN,the optimal DCCN network is compared with the FCN network,U-Net network and RF(Random Forest)algorithm.The result shows that the auto-interpretation accuracy on the GF-2 image test set is 98.1%,89.3%,93.9%,90.0%,respectively.The results of this study show that the decoding scheme based on deformable full convolution neural network can significantly improve the performance of winter wheat interpretation.Compared with traditional decoding algorithm,there are two main innovations in this algorithm: 1)The DFCN network model is improved on the basis of U-Net network,making full use of spectral,texture,edge and semantic features,and achieving end-to-end output;2)The DFCN networkDeformable convolution modules in the model make it more efficient to extract geometric deformation features.