Crop Yield and Condition Monitoring for U.S.

Image of a corn field in Iowa

Project Summary

Earth observation (EO) based and machine learning (ML) powered in-season yield forecasting is a well-studied problem, however, a gap remains between the models published in literature, and their uptake in an operational setting, either directly by farmers or by public sector entities. This is an example of an ‘over-the-wall’ problem. A lot of researchers develop models, throw them over the wall, hoping someone on the other side will catch them and utilize them. In this project, we aim to bridge this gap by refining an existing crop yield forecasting model (GEOCIF) to run at county scale for the coterminous U.S. To address the ‘over-the-wall’ problem, we will focus on: a) model availability (as a publicly available library, and potentially also as an API), b) result availability (by publicly sharing model outputs), c) model explainability (through ML techniques as well as the provision of AgMet graphics) and d) evaluating state of the art techniques like bayesian model averaging and physics informed machine learning. As part of this process, we will develop a suite of > 50 climate indices at county scale, both for ingestion into the GEOCIF model, as well as potential future studies on climate change impacts on U.S. crop yields. I am also planning to investigate at a later stage, the integration of long-term outlooks (6-12 month climate outlooks) into U.S. yield forecasts.

Study Area:

United States

Earth Observations Used

MODIS | VIIRS

Lead Institution

Project Lead

Ritvik Sahajpal

University of Maryland

Project Team:

Guanyuan Shuai

University of Maryland