Remote sensing data are used to map the extent of croplands. They are especially useful in sub-Saharan Africa (SSA) where landscapes are complex and farms are small, i.e. less than two ha. In this study, a hierarchical remote sensing approach was developed to estimate field fractions at 30 m spatial resolution in a highly fragmented agricultural region of Ethiopia. The landscape was stratified into crop production system (CPS) zones with ten-day SPOT Proba-V 1 km normalized difference vegetation index (NDVI) composites. The CPS zones were used to disaggregate agricultural census statistics to 1 km field fractions and mask “wet” and “dry” seasons. Long-term average wet-dry season NDVI and topographic information derived from 30 m Landsat-8 (OLI) surface reflectance and the SRTM digital elevation model were combined with 1 km field fractions in a Generalized Additive Model (GAM) to produce the field fractions. Sample dot grids were manually interpreted from very high-resolution DigitalGlobe imagery on the Google Earth platform for training and testing. The model yielded an Area Under the Curve (AUC) of 0.71 and R² of 0.65 in the holdout sample set. The high AUC reveals the model was effective at classifying 30 m pixels as “crop” or “not crop” while the high R² indicated leveraging at the extremes (100 and 0% probability), meaning at 30 m resolution, subpixel variations were difficult to discern. The improved model skill compared to previous cropland mapping studies using GAMs can be attributed to the stratification and decomposition of the Landsat time series using CPS-defined phenology. Additional remote sensing model inputs, such as Sentinel-1 radar backscatter and Sentinel-2 red-edge reflectance, could provide additional explanatory power. Wall-to-wall national coverage for agricultural production estimation or other food security related application could be achieved by manually digitizing additional sample data in other regions of Ethiopia or using existing crowd-sourced databases, such as Geo-Wiki.

遥感数据用于绘制农田面积图。它们在撒哈拉以南非洲地区（SSA）尤其有用，那里的景观复杂而农场很小，即少于2公顷。在这项研究中，开发了一种分级遥感方法，以估计埃塞俄比亚高度零散的农业地区在30 m空间分辨率下的野外部分。用十天的SPOT Proba-V 1 km归一化植被指数（NDVI）复合材料将景观分层为作物生产系统（CPS）区域。CPS区用于将农业普查统计数据细分为1 km的田野部分，并掩盖了“潮湿”和“干燥”季节。从30 m Landsat-8（OLI）表面反射率和SRTM数字高程模型得出的长期平均干湿季NDVI和地形信息与1 km场分数在通用加性模型（GAM）中组合以产生场分数。样本点网格是通过Google Earth平台上非常高分辨率的DigitalGlobe图像手动解释的，以进行培训和测试。该模型得出的曲线下面积（AUC）为0.71和R保留样本集中的0.65中的²。高AUC显示该模型可有效地将30 m像素分类为“作物”或“非作物”，而高R ²表示利用了极限（100和0％的概率），这意味着在30 m分辨率下，很难分辨出亚像素变化。与以前使用GAM进行耕地制图研究相比，改进的模型技能可归因于使用CPS定义的物候对Landsat时间序列的分层和分解。其他遥感模型输入，例如Sentinel-1雷达后向散射和Sentinel-2红边反射率，可以提供额外的解释能力。可以通过手动数字化埃塞俄比亚其他地区的其他样本数据或使用现有的众包数据库（例如Geo-Wiki）来实现全国范围的农业生产估算或其他与食品安全相关的应用。