Benchmark Agricultural Training Datasets to Create Regional Crop Type Classification Models from Earth Observations

Machine learning techniques provide a unique opportunity to advance applications of Earth observations for agricultural monitoring. However, representative training datasets are needed in order to build agricultural models that are able to accurately predict outputs for a diverse range of input locales. As such, it is critically important that globally representative geospatial training data is available to researchers. These public datasets will also enable benchmarking of models' performance.

Radiant Earth Foundation has created MLHub, a cloud-based repository, to host and share open training datasets for Earth observations. The first group of datasets listed on Radiant MLHub is crop type labels in Africa including data from Tanzania, Kenya, Uganda and Nigeria. In this presentation, I will review these datasets and illustrate how to use them to build crop type classification models across different regions in Radiant MLHub training data.