Partial least square regression based machine learning models for soil organic carbon prediction using visible–near infrared spectroscopy

Monitoring and assessment of soil organic carbon (SOC) are critical for maintaining and enhancing the productivity of agricultural soils. The SOC is commonly determined through soil sampling and subsequent laboratory analysis using chemical methods. This method though very precise is time-consuming, labour-intensive and expensive. Contrarily, visible and near-infrared reflectance spectroscopy (VNIRS) may be utilised to estimate SOC in a quick, labour-saving, and cost-effective manner. In this study, 72 soil samples were collected for SOC estimation and spectra collection. This current work proposes to investigate the use of PLSR scores in place of raw spectral reflectance to increase both the computation and model efficiency by reducing the number of input variables while retaining the maximum information present in the original data. With the existing indices, ratio and normalized difference indices were calculated in all possible combinations and were regressed to SOC content to identify the best-performing indices. Ten different multivariate models were evaluated for SOC estimation using full-spectrum and partial least squares regression (PLSR) scores. The results revealed that reflectance gradually increased with increasing soil depth and decreasing SOC. The prediction models developed using existing indices were observed to be poor in predicting the SOC with the R2 values ranging from 0.009 to 0.34. The best spectral indices for estimating SOC were RI (R1888, R2015) and NDI (R1888, R2015) with R2 of 0.60, 0.61 and 0.39, 0.43 for calibration and validation datasets, respectively. The PLSR score-based multivariate models outperformed solo multivariate and optimized index-based models. Our study suggested that VNIRS with PLSR combined multivariate models can reliably be used for fast and non-invasive estimation of SOC.