Country-Scale Field Boundary Mapping Using Advanced AI Models Applied to Multi-Source High-Resolution Satellite Imagery

Agriculture, one of the largest drivers of terrestrial habitat loss and climate change, is fundamental to human health and well-being, and plays a critical role in socio-economic development. However, tracking agricultural change is difficult because of the uneven availability and varying quality of data, particularly in Africa. To understand how agricultural systems are changing, it is necessary to map field boundaries at national to regional scales on an annual basis. This task requires remote sensing, which has recently seen tremendous gains in the ability to map individual fields, due to increasing availability of high-resolution imagery and advances in artificial intelligence. To improve the ability to map small fields over large areas, this research will: 1) examine whether field boundary labels developed on VHR (very high resolution) imagery improves a boundary-aware model’s ability to delineate in HR (high resolution) imagery, 2) quantify how many VHR-based labels are needed to optimize HR-based field boundary delineation, and 3) to demonstrate the ability of VHR-improved models to generate a seven-year series of country-scale field boundary maps in Ghana, Zambia, and Tanzania, and use them to analyze agricultural change. This project will improve methods for tracking and understanding the nature and impacts of widespread agricultural change.