Pedotransfer functions and machine learning: Advancements and challenges in tropical soils

Soil hydraulic properties (SHP) are crucial information for several Earth system applications and their availability strongly relies on pedotransfer functions (PTFs). In this review, we aimed to bring to light what was done in the period 2012-2021 regarding the modeling of SHP by PTFs. Data were collected from 101 publications, yielding information from 1844 different PTFs across the world. PTF development was analyzed regarding coverage/domain, predominant climate, countries, target variable, methods to derive PTF, algorithms used (in case of machine learning), predictors, model sensitivity and uncertainty assessment, number of samples for models training and testing, error metrics and goodness-of-fit criteria, among more. A significant contribution from tropical regions was observed, which was seldom in past studies. The increased use of machine learning techniques to build PTFs was also a remarkable result, although conventional regression still prevailed. Given the growth of machine learning techniques in recent years, we delve deeper into this discussion. Generally, machine learning-based PTFs provided lower errors than conventional PTFs and should be considered for future studies. As a case study, we evaluated existing PTFs to predict water retention and hydraulic conductivity of soils from an important tropical agricultural center, the Mato Grosso state in Brazil. Results reinforce the low ability of PTFs based on temperate soils data to predict SHP of tropical soils, especially concerning fine-textured soils. Nonetheless, PTFs based on tropical soils also showed modest performances, indicating that the seek for greater reliability of PTFs is still pursued. The most outstanding results were observed for fine-textured soils, for which all PTFs had poor performance. Overall, results corroborate that there is a special demand for developing PTFs capable of well recognizing the hydraulic patterns of fine-textured tropical soils.

Automated summary

This review examines the modeling of soil hydraulic properties (SHP) using pedotransfer functions (PTFs) from 2012 to 2021. The analysis of 101 publications reveals a significant contribution from tropical regions and an increased use of machine learning techniques in PTF development. However, PTFs based on temperate soils data struggle to predict SHP of tropical soils, particularly fine-textured soils. Results highlight the need for improved PTFs capable of accurately predicting the hydraulic patterns of fine-textured tropical soils.