Synthetic Agricultural Training Data for Satellite Observations

Hamed Alemohammad of Open Imagery Network Inc. in the U.S. and Ernest Mwebaze of Google AI Research Center in Ghana will generate synthetic imaging data to train machine learning algorithms to better interpret satellite images in low-resource settings to monitor crops and increase food security. The increase in global satellite observations at different spatial and temporal scales has led to the development of sophisticated analytical methods such as machine learning for a variety of applications. For agricultural applications, the optimal performance of these methods requires ground reference data from field visits, which is time-consuming, expensive, and challenging in remote areas. To circumvent this need, they will generate a time series of realistic, fully synthetic images for around 8000 plots of major crops in Kenya using a generative adversarial networks approach, which involves developing two neural network models that compete against each other. They will then compare the synthetic imaging data with their existing ground reference data to see how well they can improve the classification of crop types using machine learning methods.