Pedotransfer functions for Peruvian soils: A web tool for dry bulk density estimation

Abstract

Dry bulk density measurements are crucial in soil science for calculating soil mass and absolute contents of compounds such as carbon, nutrients, or contaminants. Despite its importance, bulk density is often omitted in soil survey due to the specialized equipment and time required for direct measurement. Pedotransfer functions provide an accurately and cost-effective alternative for estimating bulk density from readily available soil data. However, these equations face two key limitations: they lack universal applicability, requiring country-specific production or recalibration to account for national soil conditions and laboratory protocols, and their implementation remains challenging for end-users (e.g., farmers and agronomists), who need simplified tools to implement functions in field settings. Here we developed dry bulk density pedotransfer functions for Peruvian soil conditions and an open-access web tool to facilitate their application. A total of 15 pedotransfer functions were developed, 4 traditional and 11 machine learning-based, the latter including 3 models based on tabular deep learning. Model performance was evaluated based on the root mean square error (RMSE), goodness of fit (R2), and training time (TT). Statistical comparisons between the model predictions were performed with the Friedman test. Our results show that eXtreme Gradient Boosting machine (RMSE = 0.2215 Mg·m−3, R2 = 0.56, TT = 0.24 s) achieve the highest predictive performance. However, Friedman test revealed no statistically differences among most models, suggesting that traditional approaches, like the multiple linear regression (RMSE = 0.2475 Mg·m−3, R2 = 0.45, TT = 0.02 s), retain practical advantages due to their simplicity and practicality. Among tabular deep learning, only the Feature Tokenizer Transformer demonstrated competitive performance (RMSE = 0.2278 Mg·m−3, R2 = 0.54, TT = 223 s), other models showed limited predictive capability, likely due to constraints imposed by our training dataset size. The pedotransfer functions web tool enables end-users to access and utilize the developed models, thereby reducing the knowledge and application gaps.