Mapping the Planting Area of Winter Wheat at 10-m Resolution Using Sentinel-2 Data and Multimodel Fusion Method

Accurately classifying and mapping winter wheat is important for agricultural development. It is difficult to meet the requirement of high accuracy when using single models to identify winter wheat; thus, model fusion methods have been used to improve classification accuracy. However, complex model fusion methods are challenging for winter wheat classification because of model complexity and data processing efficiency problems. Therefore, we propose an efficient and simple multimodel fusion (MMF) method to improve winter wheat classification accuracy at 10-m resolution and map the planting area of winter wheat. Two common supervised classification models, random forest (RF) and gradient boosting decision tree (GBDT), and a similarity matching algorithm, time-weighted dynamic time warping (TWDTW), were used to initially classify the main planting area of winter wheat in Shandong Province. Subsequently, the proposed MMF method with five fusion strategies based on the preliminary classification results was used to subclassify winter wheat. The evaluation results showed that the MMF method with each fusion strategy could effectively improve the overall accuracy (OA), with an optimal OA of 95.7%, compared to RF, GBDT, and TWDTW (OA =91.5%, 92.1%, and 93.3%, respectively). In addition, the coefficient of determination (R2) and the root mean square error (RMSE) of statistical vs. mapped areas obtained using the optimal MMF method were 0.92 and 62.72 km2 respectively, under county level area statistics, which are higher and lower than those obtained using RF, GBDT, and TWDTW (R $^{2} =0.81$ , 0.86, and 0.90, respectively; RMSE =107.1, 89.67, and 72.48 km2, respectively). The results of this study can serve as a scientific basis for improving the accuracy and method selection for classifying and mapping the fine resolution of winter wheat.