Ground-truthing training data for agricultural machine learning models in Northern India

The AgriFieldNet project was a collaboration between IDinsight’s DataDelta team and Radiant Earth Foundation that focused on creating high-quality, ground-truthed baseline training data for machine learning models in agriculture. The project involved collecting precise geospatial data on crop types and field boundaries across four states in northern India—Bihar, Odisha, Rajasthan, and Uttar Pradesh.

Decision maker’s challenge 

Machine learning (ML) models can transform agriculture by using satellite images to estimate crop yields, monitor plant health, and support better farming decisions. But these models are only as good as the data they’re trained on. To work well, ML models need accurate, ground-truthed data—real-world observations that tell the model what’s actually happening on the ground.

Without this kind of reliable data, it’s hard to track crop performance, manage food security, or respond to climate risks. For example, governments and development agencies struggle to monitor progress on goals like reducing hunger or improving climate resilience if they don’t have dependable agricultural information. Combining Earth observations with high-quality ground data is key to closing this gap.

Impact Opportunity

The AgriFieldNet project aims to improve how agricultural land is monitored by combining satellite imagery with high-quality field data. By making its training dataset and baseline machine learning model publicly available, the project helps farmers, researchers, and policymakers make better decisions. It supports more accurate crop classification and land-use mapping, especially in regions where reliable agricultural data is limited.

For DataDelta, this was an important opportunity to apply our inclusive AI approach to a practical challenge. The project needed accurate, representative ground-truth data to train machine learning models, and DataDelta was selected because of our strong focus on quality, speed, and infrastructure. Our localized field teams and rigorous data collection systems made us a trusted partner for this effort.

Our Approach

For this project, IDinsight used high-precision Garmin devices to collect geospatial data of approximately 10,000 agricultural plots, including data on field coordinates, crop types, agricultural inputs, and production. Radiant Earth Foundation used the data to generate a publicly available training dataset[footnote]https://source.coop/radiantearth/agrifieldnet-competition[/footnote] – the AgriFieldNet dataset – and to develop a baseline machine learning model for field boundary detection.