Prediction of soil nutrients through PLSR and SVMR models by VIs-NIR reflectance spectroscopy

Though soil nutrients play important roles in maintaining soil fertility and crop growth, their estimation requires direct soil sampling followed by laboratory analysis incurring huge cost and time. In this research work, soil nutrients were predicted using VIs-NIR reflectance spectroscopy (range 350–2500 nm) with Partial Least Squares Regression (PLSR) and Support Vector Machine Regression Model (SVMR) model through principal component analysis. Two hundred soil samples were collected from Tarekswar, Hooghly, West Bengal, India to predict eight selected soil nutrients, such as soil organic carbon (OC), pH, available nitrogen (N), available phosphorus (P), available potassium(K), electric conductivity (EC), zinc (Zn) and soil texture (sand, silt, and clay) levels. The OC content was predicted with sound accuracy (R²: 0.82, RPD: 2.28, RMSE: 0.13, RPIQ: 4.15 FD-SG), followed by P (R²: 0.71, RPD: 1.83, RMSE: 4575, RPIQ: 3.44 1st derivative). The soil parameters sensitive to the particular band of visible spectrum were also identified viz. wavelengths of 409, 444, 591 and 592 nm for OC, 430 and 505 nm for P, 464 nm for K; 580 nm for Zn, 492,511,596 and 698 nm for N; 493, 569 and 665 nm for EC; 492,567 and 652 nm for pH; 457 nm for sand and 515 nm for clay.

The soil nutrient levels were predicted by PLSR and SVMR models through PCA and Sentinel 2 imagery and soil suitability map were also generated for seven soil parameters such as OC, pH, EC, N, P, K and clay content. Through map query tool in ArcGIS software environment the PLSR and SVMR model successfully identify the suitability class with level of accuracy of 87.2% and 88.9%, respectively, against the direct soil analysis based suitability mapping.

The machine learning technique based soil nutrient and soil suitability prediction can be easily adopted in different regions. This will reduce the cost of laboratory soil analysis and optimize the total time requirement.