基于遥感与DeepLabV3 + 的丘陵山区耕地地块精准识别方法

Precision Extraction of Cultivated Land Parcels in Hilly Areas Using Remote Sensing and DeepLabV3 +

摘要:

丘陵地区耕地细碎化问题,阻碍了农业机械装备应用,亟需在作业条件科学评估的基础上,加快开展耕地的宜机化改造。以四川省广汉市为研究区域,利用高分6号卫星(GF-6)的空间分辨率为2 m的遥感影像制作数据集,基于DeepLabV3 + 语义分割网络模型,融合了空洞卷积技术,采用ResNet34作为其骨干网络,有效提取丘陵山区农田的尺寸、形态和方向等多尺度特征,适应丘陵山区农田不规则边缘信息条件。开展了不同骨干网络、地物覆盖、地块面积等条件下的耕地提取及相关对比实验。研究结果表明,采用DeepLabV3 + 模型与ResNet34作为骨干网络的组合,地块边界识别与实际值的吻合度较高,精确率和召回率分别达到了93%和86%;样本区在作物全覆盖状态下,地块识别准确率分别达到71%和94%,平均准确率为82.5%。耕地分布离散程度对地块识别准确率无显著影响;地块面积越大,模型识别精度越高,地块面积超过100 m2时,识别精度提升至80%以上,但地块面积小于40 m2时,识别准确率不足50%。

Abstract:

The fragmented nature of farmland in hilly regions impedes the adoption of agricultural machinery, highlighting the urgent need for machine-accessible land transformation based on scientific operational assessments. Focusing on Guanghan, Sichuan as the study area, a dataset derived from GF-6 satellite imagery (2 m resolution) was utilized. The DeepLabV3 + semantic segmentation model, enhanced with dilated convolution and employing ResNet34 as its backbone network, effectively extracted multi-scale features, including size, shape, and orientation of hilly farmland, accommodating its irregular boundaries. Comparative experiments evaluated cropland extraction under varying backbone networks, land cover conditions, and plot sizes. Results demonstrated that the DeepLabV3 + / ResNet34 combination achieved high boundary alignment with actual values, yielding 93% precision and 86% recall. Under full crop coverage, plot identification accuracy reached 71% and 94% for specific metrics, averaging 82.5%. In the absence of crop coverage, the average accuracy was only 44.5%. The full crop coverage state can provide the model with rich spectral and textural information, which helped improve recognition accuracy. When cropped and non-cropped farmland were interspersed, the model's recognition accuracy was increased significantly, indicating that the boundary features of farmland under mixed coverage conditions were easier for the model to identify. Cropland dispersion showed no significant impact on accuracy, while larger plots correlated strongly with higher precision: accuracy exceeded 80% for plots more than 100 m2 but dropped below 50% for those less than 40 m2.