Project Description

In response to the 2020 East African Desert Locust crisis, we collaborated with PlantVillage (the world’s leading knowledge delivery system for African farmers) to develop PLAN, an end-to-end Deep Learning engine that can learn effective representations from sparse crowdsourced spatiotemporal data to provide early warnings to county governments in Kenya that are likely to be affected by Desert Locust swarms, so that preventative actions can be taken early. PLAN's deep learning architecture leverages crowdsourced reports of locust observations collected by PlantVillage, and utilizes multi-channel Convolutional Neural Nets (CNNs) to capture spatial dependency of locust presence/absence among nearby geographic regions. At the same, PLAN uses a multi-view Recurrent Neural Net (RNN) to capture temporal dependencies of locust presence/absence with meteorological factors. The output from these two networks is combined to create heatmap forecasts of locust absence/presence over large geographic regions.

In India, we have developed reinforcement learning algorithms to find optimal policies for spreading local news (about government injustices) from CGNetSwara (a local audio-based media platform) among low-literate farmers in a cost-effective manner, which is necessary to ensure CGNetSwara's financial sustainability \cite{yadavreferral}. Finally, we have also developed PECAD, a prediction engine for Indian agriculture produce based on sparse spatio-temporal pricing data.

Publications