Assessing the soil quality of Bansloi river basin, eastern India using soil-quality indices (SQIs) and Random Forest machine learning technique

In order to address agronomic sustainability issues for the densely populated country like India tools for assessing soil quality are the prime requisite. In the present study with the fundamental goal to assess soil quality we have intended to assemble a range of soil quality indicators and validate the result with respect to a Bansloi river basin (Area-1860 km2) of eastern India. In aim to approximate soil quality, field survey data of 296 surface samples (0–20 cm) are tested in laboratory settings and 12 soil quality indicators (namely- N, P, K, pH, EC, OC, Fe, Cu, Mn, Zn, B, and S) are integrated into the Soil Quality Indices (SQIs). Simple additive (SQISA), PCA based (SQIPCA) and Correlation (SQIr) based weighted index methods are used for the SQI estimation. SQIs results are then co-related with major crops (Oryza sativa, Triticum aestivum, Brassica juncea, Zea mays etc.) yield to verify their functional relationships. Spatial prediction and mapping of soil quality in data sparse environment is another aim of the study. In this regard Random Forest (RF) method is employed to calculate the predicted soil quality indices (PSOIs). Three error criteria (namely, R2; ME and RMSE) are used to appraise their relative credibility. PCA results show that, micro-nutrients Fe, Mn apart from N, P, K have significant control towards soil quality. Results also demonstrate that, SQIr and PSQIr are relatively more consistent in their correlation with the crops yield. Among the PSQIs, PSQIr counts the smallest error in the prediction of spatial soil quality. Overall results also confirm optimal performance of RF for SQI prediction and PSQIr is found to be the most competent tool to predict the soil quality at un-sampled locations. However, before generalizing, the authors recommend thorough validation of the PSQIr in other regions and long-term field experiments.