Winter wheat identification and area extraction based on hyperspectral/multispectral image fusion*

Food is God for the people. Grain production has a bearing on the national economy and people's livelihood. This study takes the Daiyue District of Tai 'an City in Shandong Province as the research area. This paper uses multi-source remote sensing images to fuse Zhuhai-1 hyperspectral images ("OHS hyperspectral images") and GF-1 high-resolution images using a hyperspherical color spatial resolution merging algorithm. The feasibility of multi-source optical image fusion in winter wheat identification is explored. Meanwhile, this paper compares OHS hyperspectral image with the Landsat8 OLI image to explore the differences between hyperspectral and multispectral in extracting winter wheat areas. Firstly, this study combines land use data to select the optimal time phase and the best wave combination to extract the range of cultivated land in the study area. Secondly, the study extracts image features (texture features, spectral features, and index features). It builds Machine Learning models (Random Forest classification, Support Vector Machine classification, and Maximum Likelihood classification) to decode and identify winter wheat and calculates its planted area using the Dimidiate Pixel Model. Finally, the extraction results of fused images are compared and analyzed with those of OHS and OLI images. The results show that the overall classification accuracy of the fusion images is the highest, and the accuracy of the OHS images is higher than that of the OLI images. The Random Forest classification accuracy of HCS fused images is as high as 98.15%, and the Maximum Likelihood classification is 96.13%. The Random Forest classification accuracy of OHS images is 94.62%, the Support Vector Machine accuracy is 93.87%, and the Maximum Likelihood classification is 91.49%. Combining the statistical data and calculating the relative error, it can be seen that the Random Forest classification of the fused images has the lowest relative error of 2.20%, and the Maximum Likelihood classification has a slightly larger relative error of 7.74%. This study shows that the fusion of hyperspectral and high-resolution images can significantly improve the accuracy of area extraction in winter wheat. In winter wheat identification, OHS hyperspectral image has significant advantages over Landsat8 OLI multispectral image. Random Forest classification is more suitable for winter wheat identification and area extraction. The results of this study could provide a reference for hyperspectral remote sensing identification of winter wheat and contribute to the scientific regulation and precise management of agricultural production.