Spatial Assessments of Soil Properties in Panrui Gram Panchayet, West Bengal, India

Abstract

Soil quality varies due to natural and anthropogenic factors, making accurate spatial prediction important for sustainable agriculture. Panrui Gram Panchayat in Birbhum, West Bengal, is a monsoon-dependent agricultural area with diverse cropping patterns, soil constraints, and groundwater issues, making it essential for spatial analysis. This study evaluated the spatial distribution of soil properties, i.e., sand, silt, clay, pH, EC, OC, N, P and K, using geostatistical and machine learning algorithms (MLA). Four MLA viz. Random Forest Regression (RFR), Support Vector Regression (SVR), Gradient Boosting Regression (GBR) and eXtreme Gradient Boosting Regression (XGBR) were compared with kriging, a geostatistical interpolation method, using 66 Ground Truth (GTs) points (calibration 45 and validation 21). Different semi-variogram models were assigned in kriging: K-Bessel for sand, Rational Quadratic for silt, pH, K; Exponential for clay, EC, OC, P and Circular for N. Model performances were evaluated using R2 and root-mean-square error (RMSE). pH, EC, and OC performed well with low RMSE values under geostatistical approaches. Based on the results, SVR showed the lowest performance for most properties. RFR and GBR attained the highest R2 values for soil nitrogen (N). Considering all soil properties during the calibration phase RFR and XGBR performed best. Kriging outperformed MLA in predicting soil pH, EC, N, P, and K, and RFR provided a good result for OC (R² = 0.443). Kriging proved more reliable for spatial analysis, accurately estimating EC, pH, and OC. Overall, in the present study, under the constraint of data scarcity and very limited number of Ground Truth, geostatistical interpolation proved more robust and spatially consistent than MLA regressors.