Kilimo Kikubwa

Dataset Overview

This dataset is associated with a pre-print article "Deep Learning for Monitoring Large-Scale Croplands in sub-Saharan Africa" and consists of land-use/land cover classifications of known Large-Scale Land Acquisitions (LSLA) in Tanzania and Ethiopia. In total, 12 LSLA sites were classified based on available high-resolution and Sentinel-2 imagery spanning 2006 to 2018. Where possible, classifications before and after the occurrence of LSLA are provided (see kilimo_kubwa_metadata.xlsx file). Additionally, land-use land cover classifications are provided for a set of "treatment" areas exposed to LSLAs and control areas with no LSLAs to provide a basis of comparison and counterfactual analysis. The land-use/ land cover include 12 classifications covering various cropland types, forest, natural land cover, built area and water bodies (see kilimo_kubwa_metadata.xlsx file). 

In addition to the land-use/land cover dataset, we provide the trained Random Forest (RF) and UNet model presented in "Deep Learning for Monitoring Large-Scale Croplands in sub-Saharan Africa". Find the trained models in the ml_models.zip folder in .pkl and .h5 format.

Article Abstract

Increasing commodity prices, rising food demand, and technological advances are changing the scale of global agriculture in the 21st century, with large-scale croplands emerging as primary driver of global environmental change. Understanding how small-scale versus large-scale cropland contributes and responds to global environmental change is increasingly important for designing effective agricultural, development, and land-use policies. However, current remote sensing methods are inadequate for differentiating small versus large cropland types across broad spatial scales. Existing methods to monitor large-scale cropland are designed at local or site-specific scales with unknown skill in generalizing to new regions. We address this gap by leveraging machine learning to differentiate small-scale versus large-scale cropland in Tanzania and Ethiopia from 2006-2018. We compare the ability of deep learning versus random forest to disaggregate cropland by size using both in-sample and out-of-sample datasets. We find that a random forest model performs better in-sample (88% overall accuracy) compared to a deep learning UNet model (81%). The deep learning UNet model, however, generalizes better out-of-sample with 72-74% accuracy compared to random forest (62-69%). Our findings suggest that deep learning models provide greater generalization because they are more robust to changing landscape patterns, although they are more sensitive to sensor noise. The acceleration of large-scale croplands is occurring across the Global South, demanding methods capable of accurately monitoring a diverse set of agricultural conditions. We anticipate our dataset and deep learning method to be a starting point for scaling identification of large-scale croplands that can be amended with additional hand-labeled data, leveraged for computer generated labels, or applied to more sophisticated model frameworks such as transfer learning. Time-series of large-scale croplands at country or continental scales will be important to improved understanding of shifting food systems, food-security, and impacts on global environmental change.