From Hyperspectral Signatures to Robotic LiDAR-Based Moisture Monitoring: A Machine Learning-Based Framework
This study presents a machine learning-based framework that transitions from high-resolution laboratory spectroscopy to a deployable robotic LiDAR system, successfully achieving non-contact, high-precision estimation of volumetric water content in soilless rockwool substrates for automated irrigation in controlled-environment agriculture.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the high-tech world of indoor farming, where crops grow under artificial lights in climate-controlled rooms, water is both the most vital resource and the most difficult to manage. Unlike traditional fields where rain and soil provide a natural buffer, these soilless systems rely on inert materials like rockwool—a fibrous, stone-based substrate that holds water for plant roots. The challenge is that this material does not behave like soil; it cannot be easily probed with standard sensors without disturbing the delicate root systems or requiring expensive, intrusive equipment. If the water level is too low, the plants dry out and die; if it is too high, they drown. For farmers to automate this delicate balance, they need a way to "see" the moisture inside the rockwool without ever touching it, using light and sensors to read the hidden state of the water.
A team of researchers at Toronto Metropolitan University has developed a new method to solve this problem, moving from detailed laboratory experiments to a practical, robot-mounted system. Their work begins by treating the rockwool like a unique fingerprint. They took small discs of the material and carefully adjusted their water content to eleven different levels, ranging from completely dry to fully saturated. Using a high-resolution spectrometer, they measured how these samples reflected light across a vast spectrum, from ultraviolet to near-infrared. The results showed a clear pattern: as the rockwool absorbed more water, its ability to reflect light changed in specific, predictable ways, particularly in the infrared regions where water molecules absorb energy.
The researchers realized that while the full spectrum of light contained all the necessary information, a simpler approach might be more practical for real-world use. They developed a custom formula, which they called the Rockwool Water Index, by identifying just three specific wavelengths of light that, when combined, could distinguish between the different moisture levels with high accuracy. This index acted as a compact summary of the material's condition. To test this, they trained a series of computer models to recognize these patterns. One model, a deep learning system designed to process the entire light spectrum, performed well. However, a more sophisticated ensemble model, which combined the predictions of several different learning algorithms using the custom index, proved to be the most effective. This combined approach correctly identified the moisture level in the rockwool with a precision exceeding 97 percent, significantly outperforming standard methods that rely on generic soil indices.
The final stage of the research asked whether these laboratory findings could survive the transition to a real, moving robot. The team mounted a compact LiDAR sensor—a device that uses laser pulses to measure distance and intensity—onto a robotic arm. They placed full-sized rockwool cubes, similar to those used in commercial greenhouses, in front of the robot. The robot captured data from multiple angles and distances, recording visible light images, active infrared signals, depth measurements, and ambient light levels. The goal was to see if the robot could detect moisture changes using only the limited sensors available on a standard agricultural machine, rather than the expensive laboratory spectrometer.
The results were promising. By feeding the robot's data into a machine learning model, the system successfully classified the moisture content of the rockwool cubes into three distinct categories: dry, moderately wet, and fully saturated. The model achieved an average precision of nearly 93 percent. This demonstrated that even without the ability to measure the exact three wavelengths identified in the lab, the combination of visible light, infrared return, and depth information was sufficient to reveal the water status of the substrate. The study confirms that moisture-related optical signals can be recovered in a practical, deployable setting, bridging the gap between complex spectral science and the everyday tools of agricultural robotics.
This work provides a clear pathway for the future of automated irrigation. By proving that non-contact sensors can accurately read the moisture levels in soilless media, the researchers have laid the groundwork for robots that can monitor entire greenhouses, identifying dry spots and triggering watering systems only where needed. The method moves beyond the limitations of physical probes, offering a way to maintain the perfect balance of water and air for plant roots without ever disturbing the crop. While the current system was tested on a specific type of rockwool under controlled conditions, the framework suggests that similar approaches could be adapted for other substrates and environments, bringing a new level of intelligence to how we grow food indoors.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.