Assessing the Information Potential of MIR Spectral Signatures for Prediction of Multiple Soil Properties Based on Data from the AfSIS Phase I Project

Summary

This study evaluated the information content of mid-infrared spectral signatures for predicting 60 soil properties across spatially variable sub-Saharan African environments using data from the AfSIS Phase I project. Three regression algorithms (PLSR, 1DCNN, GRNN) were compared, revealing that no single model optimally predicts all soil variables. Total and organic carbon, pH, and several elemental concentrations were predicted with high accuracy, whilst many other properties including texture and bioavailable cation content showed accuracy sufficient only for less demanding applications.

UK applicability

The methodology and algorithm performance may inform UK soil spectroscopy initiatives, though results are specifically calibrated to sub-Saharan African soil types and mineralogy; local recalibration would be required for UK conditions.

Key measures

Prediction accuracy (implicit R² or RMSE metrics) for 60 soil properties including texture, bulk density, moisture, carbon and nitrogen content, elemental composition (total and bioavailable), pH, electrical conductivity, and phosphorus sorption index

Outcomes reported

The study assessed the predictive accuracy of three machine-learning algorithms (PLSR, 1DCNN, GRNN) in estimating 60 soil properties using MIR spectral data from 18,250 samples across 19 sub-Saharan African countries. Results showed variable prediction accuracy across soil properties, with optimal performance for total and organic carbon, total iron and aluminium, and pH.

Topic tags