Predicting Soil Moisture Content Based on Laser-Induced Breakdown Spectroscopy-Informed Machine Learning

This study presents a pioneering approach that combines artificial intelligence and laser-induced breakdown spectroscopy (LIBS) to predict soil moisture content (MC). The traditional laboratory-based method of MC measurement, involving soil weight comparison before and after heating, is time-consuming, labor-intensive, and prone to low accuracy. In this work, we propose a non-destructive soil MC measurement technique utilizing robust nonlinear models based on LIBS-derived elemental intensities. Support vector regression (SVR) and AdaBoost-based SVR models (SVR-ADB), employing Gaussian Kernel and input features from LIBS data, were employed for MC prediction. Model performance was assessed using standard metrics such as root mean square error, mean absolute error, Nash–Sutcliffe efficiency (NSE), and correlation coefficient (CC) between predicted and actual moisture content. The study employed 485 datapoints generated in our laboratory. An advanced feature optimization technique based on the correlation between the soil MC and the descriptors was employed to select relevant mineral elements as input features. Three feature combinations (Combo-1, Combo-2, and Combo-3) were evaluated to identify the most effective configurations for accurate soil MC predictions. SVR-ADB-3 (Combo-3) demonstrated the highest prediction efficiency in the testing phase, achieving an impressive CC of 0.9998 and NSE of 0.9997. Consistently, Combo-3 outperformed other configurations, emphasizing the importance of the selected features. Validation of the developed models on soils treated with cement and lime stabilizers, whose data were not used during model calibration and verification, confirmed the generalization capability of the models. This study provides valuable insights for policymakers and industry stakeholders, facilitating optimized soil moisture management practices.