ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Reinforcement Learning-Based Forward Prediction of Greenhouse Lettuce Growth with Multimodal Sensing

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2601144.(doi:10.13031/aim.202601144)
Authors:   Yibin Wang, Bashir Al, Partho Ghose, Mike Ojo, Azlan Zahid
Keywords:   Controlled Environment Agriculture; Forward Growth Forecasting; Reinforcement Learning; Multimodal Data Fusion.

Abstract. Accurate forward prediction of crop growth is a prerequisite for closed-loop, AI-driven control in controlled environment agriculture (CEA). Management actions, including lighting, nutrients, and microclimate, influence both short-term physiology and long-term morphology. Building on our prior multimodal plant model (MPM) for day-ahead forecasting, this study proposes a reinforcement learning (RL)-based growth-forward prediction framework that models the impact of control actions on future phenotypic trajectories using multimodal sensing. We formulate greenhouse crop growth as a decision process in which the state is represented by fused multimodal observations, time-series microclimate, and nutrient signals, together with vision-derived canopy embeddings, and the action space consists of feasible adjustments. A learned dynamics module predicts the multi-step evolution of key traits, including leaf fresh weight and leaf area. A temporal-difference model predictive control (TD-MPC) RL policy is trained in a latent state space to select control actions that maximize a reward function balancing predicted crop growth with resource-use penalties. This model-based formulation improves sample efficiency and supports multi-step forward forecasting, counterfactual “what-if” simulations, and sensitivity analysis of control variables for decision support. Experiments in a greenhouse lettuce setting demonstrate that the proposed RL-based predictor produces more accurate and temporally consistent growth trajectories than supervised forecasting baselines with 8.7% and 8.3% MAPE for fresh weight and leaf area, respectively. Overall, the proposed approach advances plant growth prediction from passive observation to action-conditioned forecasting, allowing autonomous and resource-efficient greenhouse adaptive management.

(Download PDF)    (Export to EndNotes)