Evaluation of data pre-processing and regression models 
for estimation of soil organic carbon using Vis–NIR 

spectroscopy in acid soils of Colombian tropical savanna

María Fernanda Rondón Fernández
Tropical Forages

• Soil organic carbon (SOC) plays a critical role 
in terrestrial ecosystem functioning and 
global climate regulation.

• The SOC pool in agricultural soils maintains 
fertility, supports sustainable development, 
and is essential for national food security 
(Bangelesa et al. 2020).

• Rapid and accurate SOC estimation using 
near-infrared (NIR) spectroscopy can 
support sustainable agricultural 
development and climate strategies—
especially in tropical savanna soils.

Introduction

Objective

• To identify effects of spectral data pretreatment on the correlation between SOC and 
hyperspectral data.

• To evaluate the impact of different spectral pretreatments on SOC estimation models.

• To develop and compare multivariate models (PLS, SVM, SPA-MLR) optimized for SOC prediction 
in Colombian tropical savanna soils.

Methodology and Technologies applied

Soil sample collection and source 
of heterogeneity

Analyzed 629 soil samples previously 
collected at four depths (0–10, 10–30, 30–
50, 50–100 cm) from native savanna and 
cultivated pasture in HSJ, Vichada.

Preliminary Results

Conclusions

• Vis-NIR spectroscopy with proper pre-processing 
and multivariate modeling enables rapid, non-
destructive, and cost-effective soil organic 
carbon (SOC) assessment, outperforming 
traditional lab methods.

• Vis-NIR spectra are strongly correlated with lab-
measured SOC; models using both pre-
processed (T2-PLS) and raw spectra (T0-PLS) 
achieved high predictive accuracy (RPD > 2.7), 
indicating reliability for SOC prediction.

• SVM with specific pre-processing (T4-SVM) 
performed well in calibration but was less reliable 
in validation (RPD = 1.88), highlighting the need 
for cross-validation and periodic re-testing.

• Vis-NIR-based models, with proper calibration 
and validation, provide a solid basis for SOC 
inventory and can support carbon credit 
applications if quality control is maintained.

Lessons learned and next steps

Lessons Learned
• Pre-processing has a major impact: Careful spectral pre-treatment is critical for maximizing SOC 

prediction accuracy and model robustness.
• Raw spectra can still yield valuable predictions: Even without intensive processing, good models 

(like T0-PLS) are possible for many cases, but pre-processing further improves results.
• Cross-validation is essential: Strong calibration results do not always guarantee reliability, 

independent validation and data split strategies are necessary for credible application.

Next Steps
• Expand testing across regions and soil types to verify model portability and robustness for 

Colombian soils and beyond.
• Integrate Vis-NIR-based SOC monitoring into carbon inventory frameworks and explore 

applications for carbon credit markets.
• Develop practical guidelines and standard protocols for in-field Vis-NIR measurement, 

calibration, and regular re-validation.

For more information about the references of the project please contact:
Jacobo Arango |  j.arango@cgiar.org

Spectral data acquisition, using NIR DS3 FOSS (400–
2500 nm range)

Spectral data preprocessing
• Applied multiple spectral pretreatment 

techniques including smoothing, 
derivatives, SNV, detrending, and 
scattering corrections.

• Selected feature wavelengths via 
Successive Projections Algorithm (SPA).

• Calibrated and validated models using 
a 3:1 split, evaluated with R², RMSE, 
and RPD.

Multivariate statistical analyses
Four algorithms were applied to develop 
predictive models from Vis-NIR spectral 
data:

1. Successive Projection Algorithm (SPA)
2. Multiple Linear Regression (MLR)
3. Partial Least Squares (PLS)
4. Support Vector Machine (SVM)

• SOC showed high variability (CV > 64%) ranging from 0 
to 2.66%.

• Positive correlation between spectral absorbance and 
SOC, with key absorption bands at ~1410, 1918, and 
2207 nm.

• SPA consistently extracted informative bands in the 
visible and NIR regions across pretreatments.

• Pretreatments significantly improved SOC-spectra 
correlations and model accuracy.

• T2-PLS and T0-PLS models showed high confidence 
(>80%) with validation 𝑅^2≥0.86 and RPD > 2.7.

• Narrow error margins (±0.10–0.12% SOC) support field 
applicability.

• T7-SPA-MLR model demonstrated robust performance 
with ~78% confidence and ±0.15% SOC error.

• T4-SVM had lower confidence (~58%) and wider error 
(±0.20% SOC), indicating potential overfitting risk.

• These findings support Vis-NIR spectroscopy combined 
with advanced models for rapid, nondestructive SOC 
monitoring in tropical savanna soils.


	Slide 1