Date of Award

5-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Committee Chair/Advisor

Co-Chair: Carlos Toxtli-Hernández

Committee Member

Co-Chair/Advisor: Vidya Samadi

Committee Member

Kunpeng Liu

Abstract

Crop modeling is essential for agricultural water management but often relies on simplified water balance routines that limit its ability to represent complex soil moisture dynamics. To address these limitations, the primary objective of this research was to develop a coupled system that integrates the one-dimensional Richards equation, solved using a finite difference method, into the FAO Crop Water Productivity Model (AquaCrop), to enable more physically realistic simulations of soil water movement. The coupled AquaCrop-Richards system was calibrated and validated using canopy cover, biomass, and yield observations from cotton field experiments in the southeastern United States.

Under rainfed conditions, the original AquaCrop model exhibited rapid, stepwise drainage, resulting in root-zone water content (RZWC) values that were 33-37% lower than those produced by the coupled system. This underestimation of soil moisture triggered earlier water stress responses, increased the harvest index, and ultimately led to yield overestimation by approximately 15.5%. In contrast, under fully irrigated conditions, both modeling approaches performed similarly, with comparable yields. With respect to computational efficiency, the AquaCrop-Richards system increased the average runtime from 1.56 seconds to 597 seconds per growing season, while maintaining numerical stability and negligible mass balance errors. These results demonstrate that incorporating a physically based soil water solver into AquaCrop substantially improves the accuracy of soil moisture and crop yield simulations, particularly under water-limited conditions.

The secondary objective of this research was to develop and evaluate deep reinforcement learning (DRL) algorithms for irrigation optimization. Several DRL approaches, including Double Deep Q-Network (DDQN), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC), were benchmarked against Differential Evolution (DE) and brute-force search to identify robust and efficient irrigation optimization strategies. Among the evaluated methods, PPO demonstrated the most consistent and reliable performance, achieving approximately 95% of the maximum attainable reward and performing comparably to or better than DE. The trained PPO model was then used to derive optimal irrigation policies for the coupled AquaCrop-Richards system. Overall, the results show that PPO provides stable and optimal irrigation strategies, while the AquaCrop-Richards system yields more physically consistent schedules compared to AquaCrop.

Author ORCID Identifier

0009-0004-4587-9503

Download

Available for download on Monday, May 31, 2027

Share

COinS