| Winter wheat is one of the major food crops in China,and it is of great significance for China’s agricultural development to have accurate information on the spatial distribution of winter wheat.In recent years,the increasing number of high-resolution remote sensing data has provided sufficient data sources for obtaining the spatial distribution of winter wheat.However,high-resolution remote sensing images are complex and difficult to classify,and the amount of data required to extract the spatial distribution of winter wheat over a large area is large,which requires a lot of computational resources and makes it difficult to improve the speed of the algorithm;at the same time,due to the limited coverage of single high-resolution data,it is difficult to obtain the original information covering the whole area over a large area,so it is usually necessary to use medium-resolution remote sensing images to supplement,but mediumresolution data Due to the constraints of spatial resolution and mixed image elements,there are often more errors in the extraction results,making it difficult to obtain high-resolution and highprecision results.To address the above problems,this paper uses multi-source remote sensing images as the data source and proposes two types of convolutional neural network extraction models for different resolution remote sensing images respectively.The main research contents and methods of this paper are as follows.（1）Production of multi-source remote sensing image datasets.The GF2,GF6 PMS and GF6 WFV remote sensing images collected from Shandong Province were pre-processed using remote sensing image processing software and the pre-processed remote sensing images were selected for winter wheat tagging.2970 GF2 datasets,1520 GF6 PMS datasets and 1729 GF6PMS-GF6 WFV datasets from Shandong Province were produced respectively.（2）To address the problem of winter wheat extraction from high-resolution remote sensing images,this paper improves and optimizes the ERFNet model to obtain an efficient semantic segmentation network,ERFNet-EX.The model helps the network to make full use of deep features to capture more contextual information by introducing asymmetric residuals and inverted residuals modules,so that the model can shorten the time required for segmentation while improving accuracy.time required for segmentation with improved accuracy.The experimental results show that the ERFNet-EX model achieves an accuracy of 92.35% on GF2 data,which is an average improvement of 5.97% compared with the comparison model chosen for the experiment;the accuracy on GF6 PMS data achieves 91.04%,and the average time taken to process an image with a resolution of 512×512 pixels is 25 ms.（3）For the extraction of winter wheat from medium resolution remote sensing images,a Subpixel segmentation model based on convolutional neural network was constructed.The model enhances the simulation ability of the model by fusing shallow features and deep features of the medium-resolution remote sensing image through a feature extraction module based on a multilayer residual network,and then up-samples the medium-resolution feature map to high resolution through the sub-pixel convolution method in the up-sampling module,and classifies it by a classifier in combination with the high-resolution marker map.The experimental results show that the accuracy of the Subpixel segmentation model reaches 87.58%,which is better than the experimentally selected pixel-level segmentation model and improves the accuracy of the medium-resolution remote sensing images as supplementary data. |
