Evaluating Machine Learning Models for Key Soil Properties

Machine learning has become an essential tool in environmental sciences, offering faster and more cost-effective alternatives to traditional laboratory soil analysis. This study evaluates the potential of mid-infrared (MIR) and visible–near-infrared (Vis-NIR) soil spectroscopy combined with machine learning techniques to predict key soil attributes, such as clay, sand, soil organic carbon, calcium content, and cation exchange capacity. Three machine learning models were tested: partial least squares regression (PLSR), Cubist, and a one-dimensional convolutional neural network (1D CNN). Results show that models trained on MIR spectral data generally outperformed those based on Vis-NIR data, and that Cubist is the best-performing method on average. Out of 16 soil attributes that were analyzed, seven attributes were predicted with high accuracy (R² > 0.9), while only two attributes showed moderate performance (R² between 0.5 and 0.7). The findings demonstrate that combining soil spectroscopy with machine learning techniques provides a reliable and efficient alternative to conventional laboratory analysis for several key soil properties.