Comparative Evaluation of Spectral Preprocessing Techniques in Soil Clay Prediction: A Study of SG PLSR and SNV PLSR Models

Description

Page: 06-13

Arup Jyoti Pathak, Danish Tamuly, and Bijesh Thakur (Department of Soil Science, Assam Agricultural University, Jorhat, Assam)

Accurate estimation of soil clay content is crucial for sustainable land management and precision agriculture. Diffuse Reflectance Infrared Spectroscopy (DRIS), combined with machine learning models, offers a rapid and cost-effective alternative to traditional chemical analysis methods. This study evaluates the performance of two Partial Least Squares Regression (PLSR) models-Savitzky-Golay (SG) PLSR and Standard Normal Variate (SNV) PLSR-for predicting soil clay content using visible-near infrared (Vis-NIR) spectral data. Based on a dataset of 150 soil samples from Biswanath District, Assam, different spectral preprocessing techniques were applied to refine reflectance data before model development. Model validation metrics revealed that SG PLSR outperformed SNV PLSR, achieving an R² of 71.3%, an RPD of 1.81, an RPIQ of 2.56, and an RMSE of 1.25, indicating strong predictive accuracy. In contrast, SNV PLSR yielded lower performance, with an R² of 55.7%, an RPD of 1.34, an RPIQ of 0.71, and an RMSE of 2.14. The superior performance of SG PLSR is attributed to its effective noise reduction and enhanced feature extraction capabilities, which significantly improved model robustness. The study underscores the importance of preprocessing techniques in spectral modeling and highlights SG PLSR as the optimal approach for reliable soil clay estimation in tropical soils. These findings support the integration of advanced spectral techniques into routine soil analysis, offering an efficient and scalable alternative to conventional laboratory methods.