Spatial structure, parameter nonlinearity, and intelligent algorithms in constructing pedotransfer functions from large-scale soil legacy data

Poulamee Chakraborty,Bhabani S. Das,Priyabrata Santra,Niranjan Panigrahi,Hitesh B. Vasava

doi:10.1038/s41598-020-72018-2

Poulamee Chakraborty, Bhabani S. Das + Show 3 more

Open Access

https://doi.org/10.1038/s41598-020-72018-2

Copy DOI

Abstract

Pedotransfer function (PTF) approach is a convenient way for estimating difficult-to-measure soil properties from basic soil data. Typically, PTFs are developed using a large number of samples collected from small (regional) areas for training and testing a predictive model. National soil legacy databases offer an opportunity to provide soil data for developing PTFs although legacy data are sparsely distributed covering large areas. Here, we examined the Indian soil legacy (ISL) database to select a comprehensive training dataset for estimating cation exchange capacity (CEC) as a test case in the PTF approach. Geostatistical and correlation analyses showed that legacy data entail diverse spatial and correlation structure needed in building robust PTFs. Through non-linear correlation measures and intelligent predictive algorithms, we developed a methodology to extract an efficient training dataset from the ISL data for estimating CEC with high prediction accuracy. The selected training data had comparable spatial variation and nonlinearity in parameters for training and test datasets. Thus, we identified specific indicators for constructing robust PTFs from legacy data. Our results open a new avenue to use large volume of existing soil legacy data for developing region-specific PTFs without the need for collecting new soil data.

Highlights

Pedotransfer function (PTF) approach is a convenient way for estimating difficult-to-measure soil properties from basic soil data
With a series of modelling studies, we observed that the best training dataset had three critical attributes: (a) locational similarity between the training and test datasets, (b) the presence of spatial correlations for each of the predictor and response soil properties, and (c) the presence of a strong correlation between the predictor and response soil property
PTFs developed with the legacy soil data belonging to the same location as that of the test data alone failed to predict the cation exchange capacity (CEC) values calling for additional features in training datasets

Summary

Introduction

Pedotransfer function (PTF) approach is a convenient way for estimating difficult-to-measure soil properties from basic soil data. Geostatistical and correlation analyses showed that legacy data entail diverse spatial and correlation structure needed in building robust PTFs. Through non-linear correlation measures and intelligent predictive algorithms, we developed a methodology to extract an efficient training dataset from the ISL data for estimating CEC with high prediction accuracy. The similarities or differences between the calibration and validation data and the underlying correlation structure should be considered as key determinants for the efficacy of a developed PTF20,21 rather than their geographical origin Such a hypothesis has not been tested with experimental data to Scientific Reports | (2020) 10:15050.

Objectives

Methods

Findings

Conclusion