Sentinel-2 RGB composite of rice paddy fields in Pavia Province, Lombardy, Italy — study area for SAR-based conservation agriculture classification

Satellite-Based Conservation Agriculture Monitoring Using Sentinel-1 SAR and Sentinel-2 in Pavia Rice Fields

SAR-based remote sensing and machine learning to classify post-harvest management practices across 75 rice fields in Lombardy, Italy.

Sentinel-1 SAR Sentinel-2 Optical Random Forest Google Earth Engine SHAP Explainability

Objective

This project investigates whether satellite remote sensing can distinguish conservation agriculture practices, cover cropping, residue retention, and winter flooding, from conventional tillage in rice paddies, using only freely available Sentinel-1 (SAR) and Sentinel-2 (optical) imagery.

The study covers 75 rice fields (382 ha) in Pavia Province, Lombardy, Italy, over one agricultural season from May 2020 to March 2021. No ground-truth labels were available; pseudo-labels were derived through unsupervised clustering and physical signature interpretation. This is a proof-of-concept, not an operational system.

Data & Field Selection

Fields were selected through a multi-step filtering pipeline: DUSAF 2021 rice polygons filtered to 0.5–11 ha, cross-validated with SIARL 2020 records, stratified by area quartiles, and spatially spread on a 500 m grid to ensure geographic diversity, yielding 75 fields.

Sentinel-2 Surface Reflectance (10–20 m, ~5-day revisit) provided NDVI, NDTI, BSI, and NDRE indices. Sentinel-1 GRD (C-band, 10 m, 12-day revisit) provided VV and VH backscatter. December 2020 had zero usable Sentinel-2 images due to persistent Po Valley fog, SAR provided continuous coverage through this critical winter window.

Methodology

1
Field Selection DUSAF + SIARL filtering, area stratification, 500 m grid spacing → 75 fields
2
Feature Extraction 151 features from 5 temporal windows (harvest, early/mid/late post-harvest, recovery) via GEE
3
Pseudo-Labeling K-Means + GMM + DBSCAN clustering, physical signature interpretation to assign 4 management classes
4
Classification RF, XGBoost, SVM across SAR-only, Optical-only, and Fusion feature sets
5
Spatial Block CV 4 geographic quadrants to prevent spatial autocorrelation leakage
6
Explainability & Impact SHAP feature importance, IPCC Tier 1 carbon impact estimation

Temporal Profiles

Multi-temporal signatures across the five analysis windows reveal distinct physical behaviors for each management class. Cover crops show strong NDVI recovery in winter, flooded fields exhibit characteristic low backscatter, and residue retention is marked by elevated NDTI and VH/VV ratios.

Multi-temporal NDVI, NDTI, VH, and VV profiles for rice paddy conservation agriculture classes including crop residue retention, winter flooding, cover crop, and conventional tillage across five analysis windows
Temporal profiles of spectral and SAR indices across management classes and analysis windows.

Key Result: SAR Outperforms Optical

The SAR-only Random Forest achieved a weighted F1 of 0.85, substantially outperforming the optical-only model (F1 = 0.70). Fusion of both sensor types did not improve over SAR alone. This result reflects SAR's ability to penetrate cloud cover and capture structural surface properties critical for distinguishing post-harvest management.

0.85
SAR-only F1 (weighted)
0.70
Optical-only F1 (weighted)
Random Forest
Best classifier
Classification performance comparison of Sentinel-1 SAR-only, Sentinel-2 optical-only, and SAR-optical fusion using Random Forest, XGBoost, and SVM — SAR achieves F1 0.85 vs optical F1 0.70
Classification performance across sensor configurations and classifiers. SAR-only consistently outperforms optical and fusion.

Classification Results

The spatial distribution of predicted management classes across the 75 study fields. Class assignments were derived from the best-performing SAR-only Random Forest model.

Spatial classification map of conservation agriculture practices across 75 rice fields in Pavia Province showing residue retention, winter flooding, cover crop, and conventional tillage
Spatial distribution of predicted conservation agriculture practices across Pavia Province.
Management Class Fields Area (ha) Key Signature
Residue Retention 35 169 High NDTI, elevated VH/VV ratio
Winter Flooding 23 132 Low VH backscatter, specular reflection
Conventional Tillage 15 71 Bare soil BSI signal, no winter recovery
Cover Crop 2 10 Strong winter NDVI recovery

SHAP Explainability

SHAP analysis confirms that the model relies on physically meaningful features. The VH/VV polarization ratio is the top predictor overall, while different management classes are detected through different physical mechanisms, flooding through low backscatter, residue through surface roughness, and cover crops through vegetation indices.

SHAP explainability analysis showing feature importance per conservation agriculture class — VH/VV polarization ratio as top SAR feature for crop residue detection and winter flooding classification
SHAP feature importance per management class. Each class is detected through distinct physical mechanisms.

Carbon Impact Estimation

Using IPCC Tier 1 literature values (not field measurements), conservation practices across the 382 ha study area would correspond to the following estimated impacts. These are indicative estimates based on published emission factors, not site-specific measurements.

333
tCO₂eq/yr reduction
42
tC/yr sequestered
0.87
tCO₂eq/ha/yr average
IPCC Tier 1 carbon impact estimation showing CO2-equivalent reduction and soil organic carbon gain for rice paddy conservation agriculture practices including residue retention, winter flooding, and cover crop
Estimated CO₂-equivalent reduction and soil organic carbon gain by management practice (IPCC Tier 1 literature values).

Limitations & Next Steps

Limitation Impact Next Step
Pseudo-labels from clustering Circular validation risk, model may learn cluster artifacts Collect field-verified ground truth for independent validation
Small sample (75 fields) Limited generalizability, especially for cover crop (n=2) Expand to additional provinces and seasons
Single season (2020–21) Temporal signatures may vary across years Multi-year analysis to assess robustness
Literature-based carbon factors Not site-specific, inherently uncertain Partner with agronomists for in-situ measurements
No independent test set Performance may be overestimated Reserve held-out fields or use external validation data

Tools & Skills

Google Earth Engine Python scikit-learn XGBoost SHAP Sentinel-1 SAR Sentinel-2 MSI Pandas Matplotlib Seaborn Remote Sensing Spatial Block CV Unsupervised Clustering Feature Engineering IPCC Carbon Accounting