Development of land-use-specific pedotransfer functions for predicting bulk density of acidic topsoil in eastern Himalayas (India)

Abstract

Soil bulk density (BD), although a key input parameter in many soil functions is often overlooked in soil studies, including surveys. Direct measurements over a large area are slow and expensive, especially in resource-poor mountain ecosystems. For soils of eastern Himalayan ecosystems, cost-effective indirect methods such as pedotransfer functions (PTFs) are lacking. We developed seven landuse-specific PTFs for BD estimation using a soil particle size distribution (sand, and silt) and soil organic carbon database of 1206 sampling sites. Land uses were 3 uncultivated (dense forests-DF, open forests-OF, and grasslands-GL), 3 cultivated (shifting cultivation-SC, upland agriculture -UA, lowland paddy-LP), and a perennial plantation (PL). PTFs were developed using five machine learning algorithms (MLAs) and a multi-linear regression (MLR). The best PTFs were chosen based on cross-validation performance with the values of the greatest coefficient of determination (R2) and the smallest root mean square error (RMSE). In addition, the relative performance of thirteen (13) documented PTFs was evaluated on soils in the region. Among the MLAs, the artificial neural network (ANN) showed the best BD prediction for the whole dataset (R2: 0.58, RMSE: 0.09 Mg m−3) and four (e.g. GL, UA, SC, and LP: R2 of 0.65 to 0.69, RMSE of 0.06 to 0. 09 Mg m−3) of the seven land use classes. For the remaining three land uses (e.g., DF, OF and PL), based on the greatest R2 values (0.55 to 0.75), PTFs based on random forests (RF) made the best predictions. However, considering the smallest RMSE values (0.06–0.08 Mg m−3), ANN-PTFs were another option for these three land uses. Of the land uses, the best prediction for BD was obtained for PL (R2: 0.75), while the DF with variable soil properties had the lowest predictive performance (R2: 0.55). The performance of MLA-based PTFs such as the support vector machine (SVM), ridge regression (RR), and extreme gradient boosting (XGB) and MLR-PTFs were lower than the ANN-PTFs and RF-PTFs. The precision of BD prediction for 13 published PTFs was inconsistent and inferior (R2: 0.13 to 0.39) to the ANN-PTFs and RF-PTFs (R2: 0.55 to 0.75) developed in this study. Since PTFs are very region-specific, we suggest ANN-PTFs to estimate the BD of grasslands and cultivated soils while the RF-PTFs (based on the greatest R2) for dense and open forests and plantation soils in the eastern Himalayas and similar mountain ecosystems.