Small farms (<2ha) produce about 35% of the world’s food, and are mostly found in low- and middle-income countries. Reliable information about these farms is limited, and without accurate agriculture data, farmers aren’t able to estimate their yields, which affects the supply and demand of nutritious food locally, regionally, and globally.
When farmers are better able to monitor their crops, they can adjust their practices as weather patterns change and other exterior forces impact their yield. Machine Learning techniques are increasingly being applied to satellite imagery (or Earth observations, EO), to detect issues in crops and crop yields. To create the algorithms that will detect patterns and lead to recommendations for farmers, crop-related data (aka labels) must be collected, curated, and prepared.
Earlier this year, IDinsight’s Data on Demand team collaborated with Radiant Earth Foundation to create ground-truthed baseline training data for Machine Learning models in agriculture through high-quality data collection of crop types and field boundaries of agricultural plots across four states (Bihar, Odisha, Rajasthan, and Uttar Pradesh) in India. Our team collected ground reference data from approximately 9,000 plots in May this year using high-powered Garmin devices. Based on the data collected, we have collaboratively released a dataset.
To promote the use of this dataset, we have organized a challenge where the goal is to classify crop types in agricultural fields across Northern India using multispectral observations from the Sentinel-2 satellite. Fields are located in various districts in states of Uttar Pradesh, Rajasthan, Odisha and Bihar.
To be eligible for cash prizes, winners will release their top solutions under an open source license for ongoing use and learning. The winners’ solutions will be published on GitHub following Radiant Earth’s model repository template with proper documentation, and added to the Radiant MLHub model repository.