基于样本与识别算法差异的耕地地块遥感识别结果分析

Analysis of remote sensing identification results for arable land parcels based on sample and algorithm differences

关键词:  耕地  地块识别  样本  识别算法  识别精度

Key words:  cropland  parcel identification  samples  identification algorithms  identification accuracy

摘要:

【目的】耕地地块是农业生产与经营管理的基本单元。及时准确地获取耕地地块信息对农业资源监管、耕地利用监测等具有重要意义。【方法】本文运用2021-2022“MAP杯”大赛地块提取结果，采用地面真实样本，对26个地块识别项目结果分别计算混淆矩阵，验证耕地地块遥感提取精度，并从耕地地块识别训练样本获取、地块识别算法模型和识别精度差异三个视角，分析了耕地地块识别技术框架应用潜力和效果。同时，对同一研究区的各识别任务结果进行空间叠加分析和一致性评价，评估不同样本与识别算法组合方式对耕地地块识别结果影响。【结果】人工标注样本输入深度学习算法进行耕地地块识别是最主流的方式；边缘提取和语义分割结合的识别策略，应用人工标注样本进行训练获得的耕地地块精度普遍较高。【结论】使用人工标注和样本迁移相结合的方法可以提高采样效率，同时保证较高的准确性。针对地块地块破碎，景观异质性高等区域，应构建能有效利用上下文信息的识别模块；针对地块边界不清晰的场景，在边缘提取网络中融入边界增强模块，能够充分发挥高分卫星数据的空间和时间优势，提高地块提取精度，这将成为未来研究的重点发展方向。

Abstract:

[Purpose] Arable land parcels are the fundamental units of agricultural production and management. Timely and accurate acquisition of information on arable land parcels is of great significance for agricultural resource supervision and monitoring of arable land utilization. [Method] This study utilized the parcel extraction results from the 2021-2022 “MAP Cup” competition and ground-truth samples to compute confusion matrices for the results of 26 parcel identification projects, thereby verifying the accuracy of arable land parcel remote sensing extraction. The potential and effectiveness of the technical framework for arable land parcel identification were analyzed from three perspectives: the acquisition of training samples, the algorithm models employed for parcel identification, and the variability in identification accuracy. Additionally, spatial overlay analysis and consistency evaluation were conducted for the parcel identification results within the same study area, assessing the impact of different combinations of samples and algorithms on the identification results of arable land parcels. [Result] The findings revealed that the use of manually annotated samples as input for deep learning algorithms was the predominant approach in the field. When combined with edge extraction and semantic segmentation strategies, training with manually annotated samples generally achieved high precision in arable land parcel identification. [Conclusion] The combination of manual labeling and sample migration can improve sampling efficiency while ensuring high accuracy. For areas with fragmented parcels and high landscape heterogeneity, identification modules that effectively leverage contextual information should be developed. For scenarios with unclear parcel boundaries, integrating a boundary enhancement module into the edge extraction network can fully leverage the spatial and temporal advantages of Gaofen satellite data, improving the accuracy of parcel extraction. This represents a key direction for future research and development of arable land parcel identification technologies.