From spectral signatures to LiDAR-based moisture monitoring: a machine learning-based framework

Early, non-invasive detection of moisture deficits in soilless substrates is essential for precision irrigation and condition monitoring in controlled-environment agriculture. However, widely used substrates such as rockwool remain under-characterized spectrally, while existing sensing approaches are often contact-based or limited to laboratory conditions. This study presents a laboratory-to-robotic-deployment framework for non-contact estimation of volumetric water content (VWC) in horticultural rockwool. High-resolution UV–Vis–NIR spectroscopy spanning 200–2500 nm was first used to characterize moisture-dependent reflectance across eleven VWC levels. A one-dimensional convolutional neural network (1D-CNN) was developed for full-spectrum classification. A material-optimized rockwool water index (RWI) was then derived by identifying the wavelength triplet that maximized classification performance. Further, a multi-learner spectral drought (MLSD) model was developed and optimized for reduced-feature classification, improving the baseline class-average F1-score by 15.9% and achieving precision above 97% for nine VWC levels. Finally, the laboratory findings were translated into a practical sensing experiment using a LiDAR mounted on a robotic platform. Visible imagery, active near-infrared return, depth, and signal-quality measurements were acquired from full-size rockwool cubes under variations in moisture content, sensor distance, and ambient illumination. Gravimetric measurements provided reference VWC, while validation was performed by holding out complete cubes and acquisition sessions. The multimodal model achieved an average precision of \(\:92.9\%\) across VWC classes. These results connect laboratory spectral characterization with deployable active optical sensing, demonstrating a practical framework toward non-contact robotic root-zone monitoring and automated irrigation.