Large-scale agricultural field delineation in India using deep learning and weak supervision

Digital field boundaries are a foundational dataset for digital agriculture. They enable analytics to be delivered to farmers at the field level, with applications ranging from field zoning to crop insurance, and allow researchers to study management practices, determinants of productivity, pest and disease spread, and species diversity. Despite their usefulness, few countries around the world have public field boundary datasets available. Meanwhile, in recent years, newly-accessible high-resolution satellite imagery and advances in computer vision offer opportunities for automated field boundary delineation at low cost.

India is a country where agriculture employs half of the work force and contributes nearly a fifth of GDP. However, to date, there is no public field boundary dataset available in India. In this work, we collect 10,000 sparse field labels across the country and use very high resolution satellite imagery with weak supervision to train a neural network to delineate fields across India. We then apply the trained model to extract fields in three states totaling an area of 500,000 km2. We report summary statistics and the spatial distribution of field size and shape. Finally, we quantify the uncertainty in each delineated field, examine sources of error, and provide a roadmap for bringing field delineation in India to national and operational maturity.