An LED-based multispectral discrete spectrometer for in-situ soil moisture sensing using machine learning algorithms

Abstract

Accurate, low-cost, and field-deployable soil moisture content (SMC) sensing remains a key challenge in precision agriculture and environmental monitoring. This study presents a compact and energy-efficient approach for in-situ SMC estimation using discrete near-infrared (NIR) spectroscopy combined with machine-learning (ML) algorithms. For this purpose, a portable transfer-standard sensor was developed using seven selected NIR light-emitting diodes (LEDs), targeting strong water absorption bands (970, 1150, 1450, and 1900 nm) and weak-absorption reference wavelengths (1100, 1300, and 1650 nm) for soil reflectance measurements. The optomechanical design complies with international soil reflectance spectroscopy standards, employing nadir illumination and 45 degrees detection geometry, while enabling direct comparability between laboratory and field measurements. SI-traceable calibration was performed using thermogravimetric methods over an SMC range from dry conditions to approximately 25% below saturation, utilizing more than ten soil types with varying composition and texture. Linear physical models based on relative absorption depth showed good performance under soil-specific calibration but degraded under heterogeneous field conditions, highlighting sensitivity to soil texture and composition properties. To address this limitation, six ML regression models were evaluated using stratified K-fold cross-validation, feature selection, and normalization. All ML models outperformed the physical approach, with Gaussian Process Regression achieving the highest accuracy, yielding errors below 0.5% for soil-specific calibration and below 2% under generalized field conditions, in agreement with gravimetric references. Wavelength subset analysis revealed diminishing returns beyond five LEDs, with a combination of strong and weak absorption bands providing the most informative feature set. Overall, the results demonstrate that LED-based discrete NIR spectroscopy integrated with ML offers a robust, scalable, and costand energy-efficient alternative to conventional spectroscopic systems for in-situ soil moisture monitoring.