ABSTRACT Accurate regional identification of cropland quantities and spatial distributions is important for cropland monitoring, food security, and sustainable regional development. Various countries and organizations have produced series of land-cover products. However, variability among remote sensors, land-cover classification schemes, and classification methods has resulted in discrepancies. In this study, we develop a novel method to improve cropland data accuracy for the Belt and Road (B&R) region, by fusing and correcting four cropland products: CCI-LC, GFSAD30, MCD12Q1, and FROM-GLC. Spatial analysis techniques are implemented, including climate stratification, consistency assessment, and statistical filtering, to develop training samples for model correction. The Google Earth Engine (GEE) platform and random forest (RF) algorithm are executed with these training samples to correct fused multi-data product and generate a corrected 2015 cropland product. The corrected product indicates that cropland accounts for 14.94% of the B&R region, which is closer to the results found via FAO statistics than the results from any of the four individual land-cover products. On the national scale, the root mean square error between the corrected cropland product quantities and FAO statistics is 11.39% and the correlation coefficient value is 0.77. This indicates that the method exhibits better fitting characteristics. The accuracies of the areas of inconsistency among the four cropland products and our corrected product are assessed using 3112 visually interpreted samples and Google Earth. The overall accuracy of the corrected cropland product is 77.54% in inconsistent areas. The highest accuracy produced by the corrected cropland product indicates the effectiveness of our method, which can rapidly improve cropland data accuracy in heterogeneous regions. Combining the training samples produced by fusing existing cropland products and updating techniques with multi-source remote sensing data from the GEE platform, we foresee potential applications to update global cropland product.

中文翻译：

---

使用空间分析技术和谷歌地球引擎的农田数据融合和校正

摘要 农田数量和空间分布的准确区域识别对于农田监测、粮食安全和可持续区域发展非常重要。各个国家和组织都生产了系列土地覆盖产品。然而，遥感器、土地覆盖分类方案和分类方法之间的差异导致了差异。在这项研究中，我们通过融合和校正四种农田产品：CCI-LC、GFSAD30、MCD12Q1 和 FROM-GLC，开发了一种提高“一带一路”（B&R）地区农田数据准确性的新方法。实施空间分析技术，包括气候分层、一致性评估和统计过滤，以开发用于模型校正的训练样本。使用这些训练样本执行 Google Earth Engine (GEE) 平台和随机森林 (RF) 算法以校正融合的多数据产品并生成校正后的 2015 农田产品。修正后的产品表明，耕地占“一带一路”区域的 14.94%，与四种单独土地覆盖产品中任何一种的结果相比，这更接近于通过粮农组织统计得出的结果。在全国范围内，修正后的耕地产品数量与FAO统计数据的均方根误差为11.39%，相关系数值为0.77。这表明该方法表现出更好的拟合特性。使用 3112 个视觉解释样本和谷歌地球评估四种农田产品和我们校正产品之间不一致区域的准确性。在不一致的区域，校正后的耕地产品的总体准确率为 77.54%。校正后的耕地产品产生的最高精度表明我们方法的有效性，它可以快速提高异构区域中耕地数据的准确性。将通过融合现有农田产品和更新技术产生的训练样本与来自 GEE 平台的多源遥感数据相结合，我们预见了更新全球农田产品的潜在应用。