2023年中国10米空间分辨率冬小麦识别数据集

及时准确的冬小麦空间分布信息对于粮食安全和作物生产管理至关重要。由于训练数据的获取成本高、效率低，大尺度、高质量、高空间分辨率的冬小麦分类产品依然匮乏。因此，我们提出了一种融合冬小麦物候、光谱和极化信息的训练数据自动化提取方法（Automated training data generation, ATDG），用于生成高质量的冬小麦训练样本，从而开展基于机器学习方法的冬小麦遥感制图。此外，基于生成的训练数据预训练分类模型，然后结合模型年际迁移方法（Model transfer, MT）实现了冬小麦生长季内制图。通过联合ATDG和MT，结合光学和雷达影像，我们生产了2020-2023年10米空间分辨率的中国冬小麦制图数据集（ChinaWheat10）。其中，2020、2021及2023年冬小麦分布图由ATDG方法生成，2022年分布图由MT方法生成。实地调查数据表明ChinaWheat10的总体精度在94%以上，在省市两级与统计数据的相关性（R2)分别在0.95和0.91以上，而且ChinaWheat10数据集中冬小麦田块细节明显。ATDG和MT方法在大尺度高分辨率高精度冬小麦自动化制图中具有巨大潜力。

Timely and accurate spatial distribution information of winter wheat is critical for food security and crop production management. However, large-scale, high-quality, high-spatial-resolution winter wheat classification products remain scarce due to the high cost and low efficiency of acquiring training data. Therefore, we propose an automated training data generation (ATDG) method that integrates winter wheat phenology, spectral, and polarization information to generate high-quality winter wheat training samples, enabling remote sensing mapping of winter wheat using machine learning approaches. Furthermore, by pre-training a classification model with the generated training data and adopting the model transfer (MT) method for inter-annual migration, we achieved in-season mapping of winter wheat. By jointly applying ATDG and MT and integrating optical and radar imagery, we developed the ChinaWheat10 dataset, a 10-meter spatial resolution winter wheat mapping product for China spanning 2020 to 2023. Specifically, the winter wheat distribution maps for 2020, 2021, and 2023 were generated using the ATDG method, while the 2022 map was produced via the MT approach. Field survey data show that the overall accuracy of ChinaWheat10 exceeds 94%, with its correlation coefficients (R²) against statistical data reaching above 0.95 at the provincial level and 0.91 at the municipal level, respectively. Moreover, the ChinaWheat10 dataset clearly preserves fine details of winter wheat field parcels. Both the ATDG and MT methods hold great potential for large-scale, high-spatial-resolution, high-precision automated winter wheat mapping.

搜集汇总

背景与挑战

背景概述

该数据集是2023年中国10米空间分辨率冬小麦识别数据集，采用自动化训练数据生成（ATDG）和模型迁移（MT）方法，结合光学与雷达影像，生产了2020-2023年中国冬小麦分布图，总体精度超过94%。数据集存储量为1.23GB，覆盖中国范围，适用于农学和遥感技术领域，具有高空间分辨率和高精度的特点，支持粮食安全和作物管理研究。