Key points are not available for this paper at this time.
Soil properties prediction is a critical determinant of agricultural productivity, directly affecting crop quality and yield. Consequently, regular monitoring of soil parameters is essential for optimizing agricultural practices. In this study, we introduce a novel wrapper-based approach utilizing Near Infrared (NIR) spectroscopy, combined with Savitzky-Golay (SG) smoothing and Multiplicative Scatter Correction (MSC) techniques, to accurately identify 11 key soil parameters. The conventional machine learning (ML) models and wrapper-based regressors – Random Forest Regressors (W-RFR), K-Nearest Neighbor (W-KNR), and Extra-Tree Regressors (W-ETR) were applied to perform a comprehensive spectral analysis of soil profiles. Among these models, the W-ETR demonstrated superior performance for all soil parameters by achieving an R2 in the range of 0.65–0.93, root mean square error (RMSE) of 0.18–64.80, standard error of prediction (SEP) of 0.01–4.94, and Ratio of performance to deviation (RPD) of 1.70–3.76, while best confidence interval (CI) was 88–98%. Moreover, the 10-fold cross-validation was employed, and again, W-ETR performed well. These results highlight the potential of enhancing soil composition analysis by integrating NIR spectral data with advanced wrapper-based ML models and preprocessing techniques. This integrated modelling technique introduces a new paradigm in soil spectroscopy by addressing the gap between preprocessing and ML compatibility. It offers a low-cost, reliable, adaptable, and field-deployable solution for monitoring real-time soil properties.
Panchbhai et al. (Sun,) studied this question.