| Winter wheat is one of the main food crops in China.It is of great significance to obtain spatial distribution information of winter wheat for crop yield estimation,agricultural policy formulation and crop spatial distribution optimization.In recent years,high-resolution remote sensing data has been increasing,which provides reliable data sources for obtaining spatial distribution information of winter wheat with high accuracy.The use of convolutional neural network for remote sensing image segmentation makes it possible to quickly and accurately extract the spatial distribution information of winter wheat in a large range.When the classical convolutional neural network model is directly applied to the extraction of spatial distribution information of winter wheat,the extraction accuracy is low,a multi-scale feature convolutional neural network（MSFCNN）was proposed to extract the spatial distribution information of winter wheat using GF-2 remote sensing image as the data source in Feicheng City,Shandong Province,which effectively improved the extraction accuracy of spatial distribution information of winter wheat.The main research contents of this dissertation are as follows:1.In order to improve the quality of remote sensing image,GF-2 remote sensing data is preprocessed.The classification system suitable for this study was determined,and the remote sensing images were labeled,the labeled data were verified by on-the-spot verification;Finally make a data set.2.In order to extract spatial distribution information of winter wheat,a multi-scale feature convolutional neural network（MSFCNN）based on code-decoding structure was proposed.In the coding stage,the dilated convolution is used to capture the features of different scales of the image,and the atrous space pyramid module is used to further extract and fuse the multi-scale features.In the decoding stage,the features of each level in the encoder are conveyed to the decoder by jump connection,and the shallow features are fused with the deep features by the way of feature superposition,so that the detailed information of the segmented target can be recovered better.3.In order to solve the problem of category imbalance caused by the uneven distribution of objects in remote sensing images,a cross-entropy loss function with weight is proposed,and the median frequency balance method is used to determine the weight of each category.In order to avoid the problem of model overfitting,a method of data augmentation is used to expand the training data set and make the training data more diversified.In this dissertation,SegNet and Deep Labv3+ models were chosen as the comparison models for comparison experiments.The experimental results show that the MSFCNN model has mean pixel accuracy,recall and Kappa of 90.3%,91.2% and 83.0%,which are higher than those of the comparison models.The superiority of MSFCNN model in extracting spatial distribution information of winter wheat from GF-2 remote sensing images was verified. |
