Abstract

Accurate, non-destructive estimation of crop biomass is essential for automated high-frequency monitoring and optimization in controlled-environment agriculture, yet standardized approaches remain limited for short-cycle hydroponic systems. This study introduces a reproducible and fully automated method for estimating the biomass of hydroponically grown wheat sprouts (HWSs) using high-frequency RGB imaging. The workflow integrates image preprocessing, tray segmentation, and canopy feature extraction with synchronized load-cell measurements to enable continuous, non-invasive growth tracking. To account for irrigation events and associated weight fluctuations, raw mass signals were processed using a second-order low-pass Bessel filter, preserving underlying biomass trends while removing short-term oscillations. Across 3024 paired image–mass observations collected under commercial cultivation conditions, several canopy coverage, color-based indices (AGI, Proxy NDVI), and texture features exhibited strong predictive relationships with biomass. Features reflecting greenness, canopy density, and color uniformity were positively associated with plant mass, whereas brightness- and red-channel features showed consistent negative relationships. Feature selection using an elastic-net approach identified a compact subset of informative predictors, improving model stability and interpretability. Under a nested cross-validation framework based on contiguous interval splits within sprout-growth cohorts, support vector regression (SVR) achieved the best predictive performance, with an sMAPE of 3.64% and an RMSE of 0.16 kg. Additional experiments under altered illumination conditions showed that including light intensity as an explicit covariate improved model robustness across lighting regimes. These results demonstrate that combining elastic-net feature selection with environmental covariates provides a robust and transferable framework for visual biomass estimation in hydroponic HWS. More broadly, the proposed pipeline enables non-destructive crop monitoring and supports the development of intelligent, feedback-driven control strategies for hydroponic production systems.

IPC Classification

Keywords