Mapping Soil Carbon from Space: Machine Learning Meets Conservation Agriculture

Accurate soil organic carbon (SOC) monitoring is essential for sustainable agriculture and climate change mitigation, yet traditional laboratory methods remain expensive, time-consuming, and spatially limited. Our recent study published in Smart Agricultural Technology demonstrates how freely available satellite data combined with advanced machine learning can transform SOC assessment—achieving 91% prediction accuracy across two ecologically contrasting sites.

The Challenge: Monitoring Carbon Across Diverse Landscapes

Traditional SOC measurement involves labor-intensive soil sampling and chemical testing that captures only point-based snapshots. For farmers adopting conservation agriculture (CA) practices like no-tillage and mulching, proving carbon sequestration benefits requires scalable, high-resolution monitoring—particularly for accessing emerging carbon markets.

We developed a dual-sensor framework integrating:

• Sentinel-1 synthetic aperture radar (SAR) for soil moisture and structure
• Sentinel-2 multispectral imagery (MSI) for vegetation indices
• XGBoost machine learning with active learning optimization

The framework was tested across contrasting environments:

Niigata, Japan (temperate, sandy soils)

• 37°51'N, 138°56'E
• 1,800 mm annual rainfall
• Sandy loam (80% sand, 15% silt, 5% clay)
• Low organic matter (2%)

Zio, Togo (tropical, clayey soils)

• 6°42'N, 1°15'E
• 1,200–1,500 mm annual rainfall
• Clay-rich soil with high organic carbon (5%)
• Tropical agricultural calendar (March–September)

Key Findings: XGBoost Outperforms Traditional Models

Our machine learning comparison revealed clear performance hierarchies:

XGBoost (eXtreme Gradient Boosting)

• Cross-validation R² = 0.88
• Test R² = 0.91
• RMSE = 0.17 t C ha⁻¹
• Superior handling of nonlinear relationships and heterogeneous data

Random Forest

• Test R² = 0.87
• RMSE = 0.27 t C ha⁻¹
• Competitive performance with higher interpretability

Support Vector Machine

• Test R² = 0.80
• RMSE = 0.34 t C ha⁻¹
• Challenges with large, multivariate datasets

The XGBoost model achieved exceptional accuracy while maintaining generalizability across both temperate and tropical conditions—a critical requirement for operational carbon monitoring systems.

Vegetation Indices as SOC Predictors

Spectral vegetation indices showed strong correlations with SOC content, confirming the dominant role of plant cover in shaping carbon storage:

• SAVI (Soil-Adjusted Vegetation Index): r = 0.75
• EVI (Enhanced Vegetation Index): r = 0.72
• NDVI (Normalized Difference Vegetation Index): r = 0.71
• NDWI (Normalized Difference Water Index): r = 0.65

These correlations enabled accurate SOC prediction based on vegetation health and biomass inputs—key indicators farmers can influence through management decisions.

Spatial Variability Reveals Management Opportunities

High-resolution SOC maps revealed significant spatial patterns:

Niigata Site

• SOC range: 1.2 to 3.8 t C ha⁻¹
• Higher values in densely vegetated zones
• Sandy soil challenges requiring continuous organic inputs

Zio Site

• SOC range: 0.9 to 3.2 t C ha⁻¹
• Clay-rich soils support better carbon stabilization
• Year-round biomass input under tropical conditions

These maps enable farmers to identify carbon-deficient zones and prioritize where organic amendments or conservation practices will yield the greatest benefits.

Conservation Agriculture Impact: Climate and Soil Texture Matter

Our results confirm that both climatic regime and soil texture jointly control SOC stocks under conservation agriculture:

Sequestration Rates

• Tropical clayey soils (Togo): 1.83 t C ha⁻¹ yr⁻¹
• Temperate sandy soils (Niigata): 1.04 t C ha⁻¹ yr⁻¹

Fine-textured clay soils retain organic matter through micro-aggregate protection and reduced decomposition rates. Sandy soils depleted quickly without residue inputs but responded positively to no-tillage combined with mulching.

The key insight: vegetation cover and residue management, rather than climate alone, are the primary levers for building SOC under conservation agriculture.

Methodological Innovation: Dual-Sensor Fusion

While XGBoost is established, its integration with Sentinel-1 SAR + Sentinel-2 MSI feature stacks across ecologically divergent sites is rare. Our workflow delivers several advantages:

Data Economy

• Free Sentinel-1/2 imagery eliminates expensive UAV flights
• Requires only 10–20 ground samples per new area
• Standard CPU processing (no GPU costs)

Active Learning Optimization

• Binary land-use classification (93.45% accuracy) guides sampling
• Targeted bare-soil point selection improves SOC estimation
• Five-fold cross-validation ensures robustness

Geographic Transferability

• Combined-site model maintains high accuracy across both regions
• Sensor fusion captures complementary information (SAR structure + MSI spectral data)
• Addresses limitations of single-sensor workflows

Practical Applications for Climate-Smart Agriculture

1. Carbon Market Verification

Fine-resolution SOC maps enable:

• Documentation of carbon-credit baselines and additionality
• Quantification of sequestration gains from CA adoption
• Monitoring, reporting, and verification (MRV) for carbon markets

2. Site-Specific Management

• Locate sub-field carbon deficits
• Prioritize organic amendments and cover crops
• Optimize residue allocation across spatial gradients

3. Extension Services

• Evidence-based recommendations for conservation practice adoption
• Visual communication of carbon benefits to farmers
• Tracking multi-year SOC trajectories at farm scale

4. Regional Carbon Inventories

• Scalable framework for national soil carbon accounting
• Low-cost alternative to traditional laboratory-based surveys
• Integration with climate mitigation policies

Seasonal Dynamics: Temperate vs. Tropical

Our temporal analysis of five spectral indices revealed distinct patterns:

Niigata (Temperate)

• Stronger seasonal amplitude (NDVI range ≈ 0.45)
• Sharp spring rise and post-summer decline
• Well-defined growing cycle
• Surface moisture dips during dry periods

Togo (Tropical)

• Stable high values (NDVI range ≈ 0.35)
• Continuous greenness year-round
• Prolonged rainy season maintains vegetation
• Consistently high water availability

These patterns reflect how climate and management interact to control carbon inputs—critical for adapting SOC strategies to local conditions.

Limitations and Future Directions

Multi-Year Validation

Our framework requires 5–10 year confirmation under multi-year climate variability to verify long-term SOC accumulation rates.

Sampling Depth

The 0–15 cm topsoil focus captures management-driven changes but misses deeper carbon pools. Future surveys should extend to 30–60 cm for whole-profile dynamics.

Temporal Resolution

Cloud interference limits Sentinel-2 revisit frequency. Fusing Landsat, commercial CubeSats, or UAV hyperspectral data could close temporal gaps.

Model Evolution

Hybrid architectures embedding CNN-derived spatial features into gradient-boosted ensembles may lift accuracy 5–10% while retaining interpretability.

Comparative Performance: Literature Context

Our dual-sensor XGBoost pipeline delivers three clear advantages over existing approaches:

Accuracy Across Land-Use Mosaics Most optical-only studies report solid but site-limited performance (R² ≈ 0.82–0.86). Our framework sustained Test R² = 0.91 simultaneously in temperate Japan and tropical Togo, demonstrating stronger ecological robustness.

Data Economy and Cost Studies using UAV multispectral + LiDAR plateau at R² ≈ 0.74 with high logistical costs. Our framework relies solely on free Sentinel imagery and minimal ground samples, cutting financial barriers while exceeding UAV-based accuracies.

Model Simplicity Many studies feed dozens of ancillary layers (soil texture, lithology, socioeconomic data) into models and reach R² ≈ 0.73–0.75. Our active-learning feature selection showed that compact SAR backscatter + key vegetation indices + topographic metrics suffice to exceed 0.9 R², balancing accuracy with transferability.

The Bottom Line for Practitioners

✅ Adopt conservation agriculture for measurable carbon benefits—no-tillage + mulching increased SOC by 1.04–1.83 t C ha⁻¹ yr⁻¹

✅ Use free satellite data for cost-effective monitoring—Sentinel-1/2 imagery provides operational SOC tracking without expensive field campaigns

✅ Leverage machine learning for spatial insights—XGBoost models identify carbon-deficient zones for targeted interventions

✅ Document carbon sequestration for markets—high-accuracy SOC maps support carbon credit verification and additionality claims

✅ Plan for the long term—conservation agriculture outcomes appear after 5–10 years of continuous practice

⚠️ Consider soil texture—clayey soils sequester carbon faster than sandy soils, requiring adapted management strategies

Advancing Global Soil Carbon Monitoring

The Sentinel-1/2 + XGBoost framework offers an economically accessible, technically robust tool for routine SOC surveillance. By proving effectiveness across temperate and tropical conditions, our work provides a replicable model for:

• National soil carbon inventories
• Regional carbon accounting frameworks
• Farm-scale carbon credit verification
• Climate-smart agriculture extension services

As agriculture faces mounting pressure to mitigate climate change while enhancing productivity, scalable SOC monitoring becomes essential. Remote sensing and machine learning transform carbon measurement from laboratory-bound snapshots into continuous, high-resolution spatial assessments—empowering farmers, policymakers, and researchers to make evidence-based decisions for sustainable land management.

Research Publication: Beisekenov, N., Banakinaou, W., Ajayi, A.D., Hasegawa, H., & Tadao, A. (2025). Remote sensing-based soil organic carbon monitoring using advanced machine learning techniques under conservation agriculture systems. Smart Agricultural Technology, 11, 101036. doi:10.1016/j.atech.2025.101036

Study Sites: Niigata University experimental field, Japan (37°52'N, 138°56'E) and Agbelouve research station, Togo (6°42'N, 1°15'E)

Key Methods: Sentinel-1 SAR + Sentinel-2 MSI fusion; XGBoost, Random Forest, and SVM comparison; five-fold cross-validation; active learning for sampling optimization; variance partitioning analysis

Data Availability: Satellite data freely available through Copernicus Open Access Hub; field measurements provided in supplementary materials

Open Access: This research is published under CC BY license, supporting open science and reproducible research in soil carbon monitoring.