Project

DESCRIPTION:

Farmers possess significant knowledge about crops, their distress, and diseases. However, it is labor intensive for farmers to notice all the plants at all times. There needs to be an automated approach to discover crop distress and diseases to reduce farm labor requirements. We propose a computational approach to learn about crop distress and diseases from existing datasets. Moreover, we propose to use a continual learning approach to learn about novel symptoms or infections directly from end users like farmers. Our computational approach will take multi-modal inputs such as color images, hyper-spectral data and natural language to learn about anomalies in crop growth. We pose the problem of crop disease and distress detection as an anomaly detection problem. Anomalies are deviations from expected growth patterns and will allow our approach to query an expert when anomalous crop observations are made. This querying will allow our approach to collect data from farmers in a continual fashion enabling perpetual learning of unseen crop conditions good or bad. We aim to develop algorithms and user focused data collection approaches. Such user focused learning approaches are important for machine learning algorithms to cover the last mile of deployment.