Accounting for spatial structure and surface conditions through in situ soil spectral information: application to soil organic carbon modelling
Abstract
Soil organic carbon (SOC) prediction from spectroscopy has been widely explored, yet model evaluation under spatially clustered sampling designs and the role of surface-condition variables in situ modelling remain insufficiently documented. This study evaluated how internal validation strategies affect the interpretation of local spectrally based SOC models and assessed the contribution of moisture and roughness co-variables. SOC prediction models were developed using laboratory and in situ spectroscopy combined with machine learning algorithms (PLSR, RF, and SVM). Spectra from 76 in situ and 71 laboratory Luvisol samples collected across 16 fields in the Versailles Plain (France) were used to construct spectral models. Model performance was evaluated using Random CV and two spatial validation strategies: leave-one-field-out (LOFO) and buffered leave-one-out (B-LOO). Prediction performance varied among validation strategies, suggesting that they characterize distinct prediction contexts rather than a single estimate of model accuracy. B-LOO revealed distance-dependent changes in model performance within the sampled domain, while LOFO assessed field-to-field transferability. Laboratory spectra appeared more sensitive to extrapolation, exhibiting a larger decline in predictive ability under spatial validation, whereas in situ spectra retained comparatively more stable performances across validation scenarios. Moisture and roughness descriptors generally produced modest and model-dependent improvements; SHAP analyses indicated that some co-variables could be used by models without necessarily translating into substantial predictive gains. These findings, obtained under internal validation within a local clustered dataset, emphasize that validation strategies should be aligned with the intended prediction objective and that a single performance metric may inadequately describe model applicability.
Related
- Offset or not: guidance on accounting for sampling effort when modelling abundance data
- Accounting for heterogeneity in mosquito exposure to forecast chikungunya outbreaks in Reunion Island and Paraguay: a mathematical modelling study
- Accounting for the life history structure of fitness in tests for adaptive reproductive acceleration
- The Apparent Urban Advance in Juvenile Vulpes vulpes Observations Diminishes after Accounting for Switzerland's Elevational Structure
- Storage-Dependent Responses of KDML 105 Rice (Oryza sativa L.) Seeds Across Dynamic Magnetic-Field Treatment Conditions: Germination, Seedling Growth, and Dual-Phase Hydration Modelling
BibTeX
@article{merlet2026accounting,
title = {{Accounting for spatial structure and surface conditions through in situ soil spectral information: application to soil organic carbon modelling}},
author = {Hugues Merlet and Youssef Fouad and Didier Michot and Jean-Marc Gilliot and Arthur Scriban and Valentin Serre and Emmanuelle Vaudour},
journal = {Science of Remote Sensing},
year = {2026},
month = oct,
doi = {10.1016/j.srs.2026.100517},
url = {https://institut-agro-rennes-angers.hal.science/hal-05794004v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Accounting for spatial structure and surface conditions through in situ soil spectral information: application to soil organic carbon modelling
AU - Hugues Merlet
AU - Youssef Fouad
AU - Didier Michot
AU - Jean-Marc Gilliot
AU - Arthur Scriban
AU - Valentin Serre
AU - Emmanuelle Vaudour
PY - 2026
DA - 2026/10/12
JO - Science of Remote Sensing
DO - 10.1016/j.srs.2026.100517
UR - https://institut-agro-rennes-angers.hal.science/hal-05794004v1
AB - Soil organic carbon (SOC) prediction from spectroscopy has been widely explored, yet model evaluation under spatially clustered sampling designs and the role of surface-condition variables in situ modelling remain insufficiently documented. This study evaluated how internal validation strategies affect the interpretation of local spectrally based SOC models and assessed the contribution of moisture and roughness co-variables. SOC prediction models were developed using laboratory and in situ spectroscopy combined with machine learning algorithms (PLSR, RF, and SVM). Spectra from 76 in situ and 71 laboratory Luvisol samples collected across 16 fields in the Versailles Plain (France) were used to construct spectral models. Model performance was evaluated using Random CV and two spatial validation strategies: leave-one-field-out (LOFO) and buffered leave-one-out (B-LOO). Prediction performance varied among validation strategies, suggesting that they characterize distinct prediction contexts rather than a single estimate of model accuracy. B-LOO revealed distance-dependent changes in model performance within the sampled domain, while LOFO assessed field-to-field transferability. Laboratory spectra appeared more sensitive to extrapolation, exhibiting a larger decline in predictive ability under spatial validation, whereas in situ spectra retained comparatively more stable performances across validation scenarios. Moisture and roughness descriptors generally produced modest and model-dependent improvements; SHAP analyses indicated that some co-variables could be used by models without necessarily translating into substantial predictive gains. These findings, obtained under internal validation within a local clustered dataset, emphasize that validation strategies should be aligned with the intended prediction objective and that a single performance metric may inadequately describe model applicability.
SN - 2666-0172
ER -