Abstract: To improve the weak generalization ability of traditional machine learning methods in remote sensing crop classification, winter wheat classification models that employs Sentinel-2 images with different feature selections and semantic segmentation networks are tested and evaluated in Jiyuan city, Henan province. The results show that compared to the spectral features, the model loss and IoU values of the DeepLab V3+and U-Net++ based on spectral and vegetation indices are reduced and improved by 13.30% and 7.83%, 7.80% and 5.54%, respectively. In addition, the overall accuracy of winter wheat classification results based on U-Net++from 2020 to 2023 is 93.47%～95.60%, which is 0.12%～2.29% and 4.84%～7.40% higher than that of DeepLab V3+ and random forest, respectively. Moreover, the landscape metrics values also indicate that the winter wheat classification results based on U-Net++network perform better patch integrity and compactness. Finally, the change data and spatial distribution of winter wheat based on U-Net++ from 2020 to 2023 are analyzed. It can provide methodological support for practical applications such as crop area monitoring under complex terrain conditions.

Key words: hilly areas, winter wheat, Sentinel-2, semantic segmentation network, random forest