Developing and Scaling Up the Mapping Africa Active Learning Platform

This need for both growth and reform of agriculture is particularly urgent in Sub-Saharan Africa (SSA), where populations are expected to double and economies quintuple by 2050, leading to a tripling of food demand. Existing agricultural maps for SSA fail to quantify even the most basic agricultural characteristics (where and how much cropland there is), and must become much more accurate at much finer resolutions if we are to adequately solve agriculture’s challenges.

This project refines and tests a methodology for a scalable, fast, and cost-effective land cover mapping platform based on active learning, a next generation computer vision/machine learning algorithm that directs human mappers (based in SSA) to collect training data over the most difficult to classify locations, iterating until maximal accuracy is achieved. Active learning produces maps that are more accurate across a broader range of agricultural types than conventional classification methods. The maps will not only distinguish agricultural from non-agricultural areas with unprecedented accuracy, but will go beyond pixel-based classifications to map individual fields. The platform will be tested in Ghana.