Abstract: To address low efficiency in cropland extraction and fragmentation evaluation in such hilly areas，this study took Shidong Town，Anju District，Suining City，Sichuan Province—a typical hilly region—as the test area，and adopted the U‑Net+SE model using Gaofen‑2 imagery for cropland patch identification and boundary extraction，respectively.Then，the fragmentation evaluation system was created and implemented from the dimension of parcel area，shape and spatial distribution.Results showed that patch identification achieved higher accuracy［overall accuracy（OA）：92.23%；precision：0.80］than boundary extraction（OA：90.77%；precision：0.77）.Fragmentation assessment presented higher performance using the result of cropland boundary mapping compared with visual interpretation as the reference，which displayed lower difference for the comprehensive index（11.76%），as well as the subindex of area（ 8.51%） and shape（ 13.51%），which was superior to the results derived from cropland patch extraction.

Key words:  , Patch identification；Deep learning；Fragmentation assessment；Hilly region；U?Net+SE；High?resolution remote sensing；Boundary extraction

摘要： 针对当前丘陵区耕地识别方法效率低、细碎化程度评估时效性不足的问题，选择典型丘陵区四川省遂宁市安居区石洞镇为试验区，以国产高分二号影像为数据源，采用U-Net+SE模型分别开展耕地田块的斑块识别、边界提取及精度评价；进而构建耕地面积、形状、分布等细碎化评价体系，实现丘陵区耕地细碎化程度评估。结果表明，耕地斑块提取精度更高，其总体精度（OA）为92.23%、准确率（Precision）达0.80，优于耕地边界提取结果（OA为90.77%、Precision为0.77）。利用耕地边界提取结果进行耕地细碎化程度评估更接近目视解译，其面积、形状细碎指数及综合评价指数与目视解译分别仅相差8.51%、13.51%和11.76%，优于耕地斑块提取结果。

关键词: 斑块识别, 深度学习, 耕地细碎化评价, 丘陵区, U-Net+SE模型, 高分辨率遥感, 边界提取