Mapping Soybean Planting Areas in Regions with Complex Planting Structures Using Machine Learning Models and Chinese GF-6 WFV Data

To grasp the spatial distribution of soybean planting areas in time is the prerequisite for the work of growth monitoring, crop damage assessment and yield estimation. The research on remote sensing identification of soybean conducted in China mainly focuses on the major producing areas in Northeast China, while paying little attention to the Huang-Huai-Hai region and the Yangtze River Basin, where the complex planting structures and fragmented farmland landscape bring great challenges to soybean mapping in these areas. This study used Chinese GF-6 WFV imagery acquired during the pod-setting stage of soybean in the 2019 growing season, and two counties i.e., Guoyang situated in the northern plain of Anhui Province and Mingguang located in the Jianghuai hilly regionwere selected as the study areas. Three machine learning algorithms were employed to establish soybean identification models, and the distribution of soybean planting areas in the two study areas was separately extracted. This study adopted a stepwise hierarchical extraction strategy. First, a set of filtering rules was established to eliminate non-cropland objects, so the targets of subsequent work could thereby focus on field vegetation. The focal task of this study involved the selection of well-behaved features and classifier. In addition to the 8 spectral bands, a variety of texture features, color space components, and vegetation indices were employed, and the ReliefF algorithm was applied to evaluate the importance of each candidate feature. Then, a SFS (Sequential Forward Selection) method was applied to conduct feature selection, which was performed coupled with three candidate classifiers, i.e., SVM, RF and BPNN to screen out the features conductive to soybean mapping. The accuracy evaluation results showed that, the soybean identification model generated from SVM algorithm and corresponding feature subset outperformed RF and BPNN in both two study areas. The Kappa coefficients of the ground samples in Guoyang ranged from 0.69 to 0.80, while those in Mingguang fell within the range of 0.71 to 0.76. The near-infrared band (B4) and red edge bands (B5 and B6), the ‘Mean’ texture feature and the vegetation indices, i.e., EVI, SAVI and CI green , demonstrated advantages in soybean identification. The feature selection operation achieved a balance between extraction accuracy and data volume, and the accuracy level could also meet practical requirements, showing a good application prospect. This method and findings of this study may serve as a reference for research on soybean identification in areas with similar planting structures, and the detailed soybean map can provide an objective and reliable basis for local agricultural departments to carry out agricultural production management and policy formulation.