2000-2023年中国30m空间分辨率冬小麦识别数据集

大尺度、长时序的作物空间分布信息对于作物种植规划和粮食安全评估至关重要。现有作物遥感制图方法因缺乏地面标签和影像数据，时空拓展性受限。本研究提出一种知识引导的机器学习（KGML）方法，以作物物候知识为基础，通过进一步整合多源遥感和环境数据，逐年自动提取训练样本，从而生成2000-2023年30米空间分辨率中国冬小麦空间分布产品（ChinaWheat30L）。独立验证表明，ChinaWheat30L总体精度为0.929，F1分数为0.906，其估算的冬小麦种植面积与农业统计数据吻合度高（省级R²=0.93，市级R²=0.84），较其他产品精度提高4-10%。此外，ChinaWheat30L表明，全国冬小麦种植面积总体保持稳定，但山地丘陵、干旱半干旱以及快速城镇化地区冬小麦种植面积呈现减少趋势，而黄淮海平原略有增加。该方法无需地面标签，为作物种植布局和粮食生产力监测预警提供重要的数据支撑。

Large-scale, long-time-series spatial distribution data of crops is critical for crop planting planning and food security assessment. However, existing crop remote sensing mapping methods have limited spatial-temporal scalability due to the lack of ground reference labels and sufficient remote sensing image data. This study proposes a knowledge-guided machine learning (KGML) method grounded in crop phenological knowledge. By further integrating multi-source remote sensing and environmental datasets, it automatically extracts training samples annually, thereby generating the 30-meter spatial resolution spatial distribution product of winter wheat in China for the period 2000–2023, termed ChinaWheat30L. Independent validation results show that ChinaWheat30L achieves an overall accuracy of 0.929 and an F1-score of 0.906. The estimated winter wheat planting area aligns closely with agricultural statistical data, with a coefficient of determination (R²) of 0.93 at the provincial level and 0.84 at the municipal level. Compared with other existing products, its accuracy is improved by 4–10%. Furthermore, ChinaWheat30L reveals that the national total winter wheat planting area remains generally stable, while regions with decreasing winter wheat cultivation include mountainous and hilly areas, arid and semi-arid zones, and rapidly urbanizing areas, and the Huang-Huai-Hai Plain sees a slight increase. This method requires no ground reference labels, providing vital data support for crop planting layout monitoring and early warning of food productivity.

搜集汇总

背景与挑战

背景概述

该数据集是2000-2023年中国30米空间分辨率冬小麦识别数据集，采用知识引导的机器学习方法生成，无需地面标签，整合多源遥感和环境数据自动提取训练样本。数据集精度高（总体精度0.929），与农业统计数据吻合度高，并揭示了全国冬小麦种植面积总体稳定但区域变化明显的时空动态特征。