Date of Award

Spring 5-15-2025

Document Type

Thesis (Master's)

Department or Program

Computer Science

First Advisor

Soroush Vosoughi

Second Advisor

Timothy J. Pierson

Third Advisor

Christophe Hauser

Abstract

Modern agriculture generates vast volumes of satellite and environmental data, yet the tools used to forecast crop outcomes often fail to integrate these inputs into accurate, scalable predictions. In soybean production—a sector where yield forecasts underpin hundreds of billions of dollars in commodity trading, insurance, and supply chain activity—current methods typically rely on delayed government reports, aggregated seasonal summaries, or localized proprietary surveys. These limitations restrict both the timeliness and the geographic scalability of traditional forecasting systems.

This thesis presents a machine learning framework designed to predict county-level soybean yields by leveraging satellite imagery and environmental data collected over daily timescales. The system processes high-frequency time series inputs through a learnable model architecture capable of recognizing dynamic crop development pat terns across time and space. By directly modeling environmental dynamics at fine temporal resolution, the framework enables automated, high-accuracy forecasts with out the need for ground surveys or region-specific calibration.

Experimental results demonstrate test R-squared values between 0.87 and 0.91, with root mean squared error (RMSE) values between 3.02 and 3.37 bushels per acre on held out years—substantially outperforming published academic benchmarks. These out comes demonstrate that scalable, real-time yield forecasting can be achieved through advanced deep learning approaches, offering significant potential for earlier, more geographically resolved agricultural intelligence.

Available for download on Saturday, May 15, 2027

Share

COinS