CONTEXT

Mapping crop types from satellite images is a promising application in agricultural systems. However, it is a challenge to automate in-season crop type mapping over a large area because of the insufficiency of ground truth and issues of scalability, reusability, and accessibility of the classification model. This study introduces a framework for automatic crop type mapping using spatiotemporal crop information and Sentinel-2 data based on Google Earth Engine (GEE). The main advantage of the framework is using the trusted pixels extracted from the historical Cropland Data Layer (CDL) to replace ground truth and label training samples in satellite images.

OBJECTIVE

This paper will achieve three objectives: (1) assessing spatiotemporal crop information derived from the historical crop cover maps; (2) mapping crop cover, mainly crop fields without regular historical crop rotation patterns, from remote sensing data using supervised learning classification and validating mapping results; and (3) automating in-season crop mapping and exploring the scalability of the framework.

METHODS

The proposed crop mapping workflow consists of four stages. The data preparation stage preprocesses CDL and Sentinel-2 data into the required structure. The spatiotemporal crop information sampling stage extracts trusted pixels from the historical CDL time series and labels Sentinel-2 data. Then a crop type classification model can be trained using the supervised learning classifier in the model training stage. In the mapping/validation stage, an in-season crop cover map over the full Sentinel-2 tile will be produced using the trained model and the classification performance will be validated using CDL or other ground truth data.

RESULTS AND CONCLUSIONS

We systematically perform a group of experiments for in-season mapping of five major crop types (corn, cotton, rice, soybeans, and soybeans-wheat double cropping) over the Mississippi Delta region. The result indicates that the crop cover map of the study area is expected to reach 80%–90% agreement with CDL within the growing season. To further facilitate the use of the framework, we also develop a GEE-enabled online prototype, In-season Crop Mapping Kit, and explore its scalability over agricultural fields in various ecoregions including California, Idaho, Kansas, and Illinois.

SIGNIFICANCE

The mapping-without-ground-truth approach described in this paper can significantly reduce ground truthing process and save substantial resource needs and labor costs, which is applicable to the production of in-season CDL-like data for the entire United States. The findings and outputs will benefit the agriculture community and other agricultural sectors ranging from government, academia, and companies.

中文翻译：

---

利用时空作物信息和遥感数据实现当季作物类型映射的自动化

语境

从卫星图像映射作物类型是农业系统中一个很有前景的应用。然而，由于基本事实的不足以及分类模型的可扩展性、可重用性和可访问性问题，在大面积上自动化当季作物类型映射是一个挑战。本研究介绍了一个使用时空作物信息和基于谷歌地球引擎 (GEE) 的 Sentinel-2 数据的自动作物类型映射框架。该框架的主要优点是使用从历史农田数据层 (CDL) 中提取的可信像素来替换卫星图像中的地面实况和标签训练样本。

客观的

本文将实现三个目标：（1）评估来自历史作物覆盖图的时空作物信息；(2) 使用监督学习分类从遥感数据中绘制作物覆盖图，主要是没有规律的历史作物轮作模式的农田，并验证绘图结果；(3) 自动化当季作物映射并探索框架的可扩展性。

方法

建议的作物映射工作流程包括四个阶段。数据准备阶段将 CDL 和 Sentinel-2 数据预处理为所需的结构。时空作物信息采样阶段从历史 CDL 时间序列中提取可信像素并标记 Sentinel-2 数据。然后可以在模型训练阶段使用监督学习分类器训练作物类型分类模型。在映射/验证阶段，将使用经过训练的模型生成整个 Sentinel-2 切片上的当季作物覆盖图，并使用 CDL 或其他地面实况数据验证分类性能。

结果和结论

我们系统地进行了一组实验，对密西西比三角洲地区的五种主要作物类型（玉米、棉花、水稻、大豆和大豆-小麦双季作物）进行季节映射。结果表明，研究区的作物覆盖图预计在生长季节内与CDL的一致性达到80%~90%。为了进一步促进该框架的使用，我们还开发了一个支持 GEE 的在线原型，即当季作物绘图工具包，并探索其在包括加利福尼亚、爱达荷、堪萨斯和伊利诺伊在内的各种生态区的农业领域的可扩展性。

意义

本文描述的无地面实况制图方法可以显着减少地面实况处理过程，节省大量资源需求和劳动力成本，适用于整个美国的季节性 CDL 类数据的生产。研究结果和产出将使农业界和政府、学术界和公司等其他农业部门受益。