Crop Classification For GF-6 WFV Image Based On SLIC-MRF And Swin Transformer

Fine classification and identification of crops based on remote sensing images is the basic link for growth monitoring,planting structure extraction and disaster assessment in agricultural remote sensing applications.The rapid development of remote sensing technology in recent years has provided convenient and fast technical support for the acquisition of remote sensing image data,and plays an important role in promoting the development of agricultural informatization.High-resolution remote sensing images can effectively reflect the differences in characteristics of different crops in different periods,providing a reliable data source for the development of crop monitoring and other fields.However,the features on the high-resolution remote sensing images are complex,and the crops are easily affected by the changes of weathering period and growth state and present large differences.Therefore,the traditional crop classification methods often cannot make full use of the rich high-dimensional features in the images,and it is difficult to obtain ideal classification results.Deep learning is a new algorithm built on the basis of artificial intelligence field,which has strong feature learning and expression ability and can get higher level abstract feature representation through autonomous learning,and has unique advantages in classification accuracy and generalization performance,which brings new ideas for high resolution remote sensing image classification.The GF-6 is the first satellite applied to precise agricultural observation in China,and the GF-6 WFV image is also the first medium and high-resolution remote sensing image with rededge band in China,which can better provide fine service for crop monitoring.As a sensitive characteristic spectral band for plant growth condition,the red-edge band can be used to effectively extract the crop information and conduct the classification by remote sensing images.Based on the above background,this paper applies deep learning algorithm to GF-6WFV satellite images for crop classification,and the main research contents are as follows:(1)To evaluate the capability of GF-6 WFV satellite image for crop classification applications,this study proposes a combined SLIC-MRF and deep learning crop classification method and a region of Shuangyashan City,Heilongjiang Province is the study area.Firstly,the SLIC-MRF segmentation algorithm is used to initially classify the study area to extract vegetation areas;then the Swin Transformer deep learning model is used to extract features and classify crops in vegetation areas with and without red-edge bands respectively;finally,the the results by the proposed method is compared with other four algorithms of Nearest Neighbor,Support Vector Machine,Random Forest and Convolutional Neural Network Resnet to verify.(2)The experimental results show that a small number of crops are mixed and missed in the classification results without the red-edge band,and the classification boundaries are blurred,while three crops can be identified more completely in the classification results with the red-edge band.Meanwhile,the overall classification accuracy of crops improves from95.68% to 98.05% using satellite images with red-edge bands,and the Kappa coefficient also improves from 0.94 to 0.97.This indicates that the introduction of red-edge bands can effectively reflect the differences in the characteristics of different crop categories and improve the classification accuracy significantly,which fully reflects the advantages of GF-6 WFV images in crop The advantages of GF-6 WFV images in crop classification recognition were fully demonstrated.(3)Different algorithms are applied for crop classification,and the results show that the classification accuracy of Resnet and Swin Transformer based on deep learning algorithms is higher than the traditional machine learning algorithms KNN,SVM and RF,indicating that deep learning algorithms have advantages in remote sensing image classification.In addition,the classification effect of Resnet model is better than RF model,and the classification effect of KNN model is worse.Compared with the four classification algorithms of KNN,SVM,RF and Resnet,the classification method proposed in this paper has the highest overall classification accuracy and Kappa coefficient,and the classification effect is better than other methods,which can provide important reference information for the application of GF-6 WFV images to precision agriculture monitoring...