Abstract. Smallholder agroecological subzones (AEsZs) produce an array of crops occupying large areas throughout Africa but remain largely unmapped. We explored multisource satellite datasets to produce a seamless land-use and land-cover (LULC) and fragmentation dataset for upper midland (UM1 to UM4) AEsZs in central Kenya. Specifically, the utility of PlanetScope, Sentinel 2, and Landsat 8 images for mapping coffee-based landscape were tested using a random forest (RF) classifier. Vegetation indices, texture variables, and wavelength bands from all satellite data were used as inputs in generating four RF models. A LULC baseline map was produced that was further analyzed using FRAGSTAT to generate landscape metrics for each AEsZs. Wavelength bands model from Sentinel 2 had the highest overall accuracy with shortwave near-infrared and green bands as the most important variables. In UM1 and UM2, coffee was the dominant cover type, whereas annual and other perennial crops dominated the landscape in UM3 and UM4. The patch density for coffee was five times higher in UM4 than in UM1. Since Sentinel 2 is freely available, the approach used in our study can be adopted to support land-use planning in smallholder agroecosystems.

中文翻译：

---

肯尼亚中部咖啡农业生态分区的景观破碎化：一种多尺度遥感方法

摘要。小农生态农业分区 (AEsZs) 生产一系列作物，占据整个非洲的大片土地，但大部分地区仍未绘制地图。我们探索了多源卫星数据集，为肯尼亚中部的上中部地区（UM1 至 UM4）AEsZ 生成无缝的土地利用和土地覆盖 (LULC) 和碎片化数据集。具体而言，使用随机森林 (RF) 分类器测试了 PlanetScope、Sentinel 2 和 Landsat 8 图像用于绘制基于咖啡的景观的效用。来自所有卫星数据的植被指数、纹理变量和波段被用作生成四个 RF 模型的输入。生成的 LULC 基线图使用 FRAGSTAT 进行进一步分析，以生成每个 AEsZ 的景观指标。Sentinel 2 的波长带模型具有最高的整体精度，其中短波近红外和绿带是最重要的变量。在 UM1 和 UM2 中，咖啡是主要的覆盖类型，而一年生和其他多年生作物在 UM3 和 UM4 中占主导地位。UM4 中咖啡的斑块密度是 UM1 中的五倍。由于 Sentinel 2 是免费提供的，因此可以采用我们研究中使用的方法来支持小农农业生态系统的土地利用规划。