Mapping soil particle fractions by training digital soil mapping models with surrogate measurements obtained from laboratory and satellite Vis-NIR spectral data
Abstract
Digital Soil Mapping (DSM) is widely used to generate spatial soil information, but its accuracy is often limited by sparse field observation. To overcome this limitation, less costly soil data on soil properties (surrogate data) can be added for densifying the spatial sampling of DSM model inputs. In this study we evaluated the added value of complementing a spatial sampling of soil property measurements (sand and clay) by soil property estimations based on either laboratory Visible and Near-Infrared (Vis-NIR, 400-2500 nm) or satellite (Sentinel-2) data. This work was conducted on a cultivated watershed located in the southwestern part of India.Our results highlight first that, regardless of the soil property, i) adding few data (166), accurately estimated by laboratory Vis-NIR spectra and partial least-squares regression (PLSR) model, slightly improve the DSM model performances, while ii) adding a large set of data (3000), modestly predicted by Sentinel-2 data and Random Forest (RF) model over bare soil pixels, highly improved the DSM model performances. In addition, we also showed that Sentinel-2 data alone are more suitable for predicting the soil property associated to spectral characteristics (clay), whereas DSM approaches that integrate environmental covariates and ensure full spatial coverage are more effective for soil property without spectral characteristics (sand). In conclusion, using soil properties predicted from multispectral remote sensing to train Digital Soil Mapping models can significantly improve the quality of soil maps. With the worldwide availability of high-temporal-frequency Sentinel-2 data, this approach has considerable potential for broad implementation.
Related
- Insights into magmatic processes at Shishaldin Volcano based on satellite and airborne volcanic gas measurements and magma degassing models
- Critical role of single-particle partial coating fractions in reproducing black carbon absorption and refining climate simulations
- From laboratory measurements to field predictions of Ionizable organic micropollutants: a critical review and a proposed prospective validation framework
- Particle behavior and water quality response of wildfire ash compared to laboratory heated materials
- Ion and particle size distribution measurements in the Indo-Gangetic Plain: seasonal insights and challenges in new particle formation event characterization
BibTeX
@article{dharumarajan2026mapping,
title = {{Mapping soil particle fractions by training digital soil mapping models with surrogate measurements obtained from laboratory and satellite Vis-NIR spectral data}},
author = {S. Dharumarajan and Mlmaw Weerasekara and Philippe Lagacherie and Cécile Gomez},
journal = {Geoderma},
year = {2026},
month = oct,
doi = {10.1016/j.geoderma.2026.117981},
url = {https://hal.inrae.fr/hal-05784124v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Mapping soil particle fractions by training digital soil mapping models with surrogate measurements obtained from laboratory and satellite Vis-NIR spectral data
AU - S. Dharumarajan
AU - Mlmaw Weerasekara
AU - Philippe Lagacherie
AU - Cécile Gomez
PY - 2026
DA - 2026/10/07
JO - Geoderma
DO - 10.1016/j.geoderma.2026.117981
UR - https://hal.inrae.fr/hal-05784124v1
AB - Digital Soil Mapping (DSM) is widely used to generate spatial soil information, but its accuracy is often limited by sparse field observation. To overcome this limitation, less costly soil data on soil properties (surrogate data) can be added for densifying the spatial sampling of DSM model inputs. In this study we evaluated the added value of complementing a spatial sampling of soil property measurements (sand and clay) by soil property estimations based on either laboratory Visible and Near-Infrared (Vis-NIR, 400-2500 nm) or satellite (Sentinel-2) data. This work was conducted on a cultivated watershed located in the southwestern part of India.Our results highlight first that, regardless of the soil property, i) adding few data (166), accurately estimated by laboratory Vis-NIR spectra and partial least-squares regression (PLSR) model, slightly improve the DSM model performances, while ii) adding a large set of data (3000), modestly predicted by Sentinel-2 data and Random Forest (RF) model over bare soil pixels, highly improved the DSM model performances. In addition, we also showed that Sentinel-2 data alone are more suitable for predicting the soil property associated to spectral characteristics (clay), whereas DSM approaches that integrate environmental covariates and ensure full spatial coverage are more effective for soil property without spectral characteristics (sand). In conclusion, using soil properties predicted from multispectral remote sensing to train Digital Soil Mapping models can significantly improve the quality of soil maps. With the worldwide availability of high-temporal-frequency Sentinel-2 data, this approach has considerable potential for broad implementation.
SN - 0016-7061
SN - 1872-6259
ER -