Month: September 2026

Most commercial agricultural dashboards rely heavily on optical remote sensing metrics—chief among them the Normalized Difference Vegetation Index (NDVI) and 2-band Enhanced Vegetation Index (EVI2) derived from Sentinel-2 or Landsat constellations. While optical indices perform reliably in semi-arid zones or during clear-sky winter (Rabi) cycles, they break down in tropical and subtropical monsoonal environments.

For agricultural risk underwriting and yield forecasting, relying solely on optical indices introduces critical vulnerabilities:

• Persistent Cloud Contamination: During the peak vegetative and reproductive stages of monsoon crops (July to September in South and Southeast Asia), cloud cover often exceeds 80–90% across consecutive satellite overpasses. An insurer or lender cannot afford a 45-day blind spot during tillering and panicle initiation.
• Signal Saturation: Optical NDVI asymptotically saturates once canopy closure occurs (Leaf Area Index LAI > 3), failing to distinguish between vegetative growth, biomass accumulation, and early-stage moisture or nutrient stress.
• Standing Water vs. Moisture Blindness: Optical reflectance cannot reliably quantify standing water beneath an established crop canopy, blinding risk models to early-season submergence or flash flood impacts.

The solution lies in active microwave remote sensing: Synthetic Aperture Radar (SAR) via the Sentinel-1 constellation.

---

1. The Physics of C-Band SAR in Crop Monitoring

Unlike optical sensors that measure reflected solar radiation, Sentinel-1 is an active microwave system emitting pulses at C-band frequency (~5.405 GHz, wavelength λ ≈ 5.6 cm). Because C-band wavelengths are larger than cloud droplets and rain cells, the signal penetrates weather phenomena unimpeded, providing guaranteed observations every 6 to 12 days regardless of daylight or cloud conditions.

     Cloud Cover (Transparent to ~5.6 cm C-band microwaves)
     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
                          │            ▲
                 Emitted  │            │ Backscattered
                  Signal  │            │ Signal (sigma-0)
                          ▼            │
         ┌───────────────────────────────────────────┐
         │ Canopy (Volume Scattering - VH dominant)  │
         ├───────────────────────────────────────────┤
         │ Soil / Water Surface (Double-bounce /     │
         │ Specular reflection - VV dominant)        │
         └───────────────────────────────────────────┘

Polarization Dynamics: σ⁰_VV vs. σ⁰_VH

Sentinel-1 operates in dual-polarization (VV and VH in Interferometric Wide swath mode):

• σ⁰_VV (Vertical transmit, Vertical receive): Highly sensitive to rough surface scattering and vertical stem structures. In flooded fields (such as transplanted paddy), specular reflection redirects incoming energy away from the sensor, causing σ⁰_VV to plunge significantly (often below −18 dB).
• σ⁰_VH (Vertical transmit, Horizontal receive): Controlled by depolarizing cross-reflections within the crop canopy (volume scattering). As crops branch, tiller, and accumulate above-ground biomass, σ⁰_VH increases systematically.
• The Cross-Ratio (σ⁰_VH / σ⁰_VV or VH − VV in dB): A powerful metric that normalizes variations in environmental moisture and soil roughness, serving as a radar-based structural vegetation index.

---

2. Temporal Phenology Profile of Paddy Rice

Tracking the backscatter trajectory across biological growth stages enables automated phenological anomaly detection:

If a region fails to show the signature “specular drop” during June–July, sowing was delayed or hindered by drought. If σ⁰_VH collapses abruptly during heading, it indicates lodging, pest defoliation, or inundation damage.

---

3. End-to-End Implementation in Python (Google Earth Engine)

Below is a modular script using the Earth Engine Python API (ee) to extract and compare the temporal profile of Sentinel-1 SAR and Sentinel-2 optical data for an agricultural region in Assam.

Prerequisites & Setup

import ee
import pandas as pd
import matplotlib.pyplot as plt

# Authenticate and initialize Earth Engine
ee.Authenticate()
ee.Initialize()

# Define Area of Interest (Agricultural area in Dhemaji, Assam)
aoi = ee.Geometry.Polygon([
    [
        [94.45, 27.35],
        [94.65, 27.35],
        [94.65, 27.55],
        [94.45, 27.55],
        [94.45, 27.35]
    ]
])

start_date = '2024-05-01'
end_date = '2024-11-30'

Ingesting and Filtering Sentinel-1 SAR Backscatter

def get_sentinel1_timeseries(geometry, start, end):
    """
    Extracts calibrated, speckle-filtered Sentinel-1 backscatter (VV and VH in dB).
    """
    s1 = (ee.ImageCollection('COPERNICUS/S1_GRD')
          .filterBounds(geometry)
          .filterDate(start, end)
          .filter(ee.Filter.eq('instrumentMode', 'IW'))
          .filter(ee.Filter.listContains('transmitterReceiverPolarisation', 'VV'))
          .filter(ee.Filter.listContains('transmitterReceiverPolarisation', 'VH'))
          .filter(ee.Filter.eq('orbitProperties_pass', 'DESCENDING')))

    def compute_ratios(image):
        vv = image.select('VV')
        vh = image.select('VH')
        ratio_db = vh.subtract(vv).rename('VH_VV_ratio')
        
        # 3x3 focal mean filter to reduce radar speckle noise
        filtered_vv = vv.focal_mean(radius=1.5, units='pixels').rename('VV_filtered')
        filtered_vh = vh.focal_mean(radius=1.5, units='pixels').rename('VH_filtered')
        
        date = image.date().format('YYYY-MM-dd')
        mean_values = ee.Image([filtered_vv, filtered_vh, ratio_db]).reduceRegion(
            reducer=ee.Reducer.mean(),
            geometry=geometry,
            scale=20,
            maxPixels=1e9
        )
        return ee.Feature(None, {
            'date': date,
            'VV': mean_values.get('VV_filtered'),
            'VH': mean_values.get('VH_filtered'),
            'VH_VV': mean_values.get('VH_VV_ratio')
        })

    features = s1.map(compute_ratios).getInfo()['features']
    data = [f['properties'] for f in features if f['properties']['VV'] is not None]
    
    df = pd.DataFrame(data)
    df['date'] = pd.to_datetime(df['date'])
    return df.sort_values('date').reset_index(drop=True)

Ingesting Sentinel-2 for Cloud Gap Comparison

def get_sentinel2_ndvi(geometry, start, end):
    """
    Extracts mean NDVI from Sentinel-2 surface reflectance along with cloud cover.
    """
    s2 = (ee.ImageCollection('COPERNICUS/S2_SR_HARMONIZED')
          .filterBounds(geometry)
          .filterDate(start, end))

    def compute_ndvi(image):
        ndvi = image.normalizedDifference(['B8', 'B4']).rename('NDVI')
        cloud_pct = image.get('CLOUDY_PIXEL_PERCENTAGE')
        date = image.date().format('YYYY-MM-dd')
        
        stats = ndvi.reduceRegion(
            reducer=ee.Reducer.mean(),
            geometry=geometry,
            scale=20,
            maxPixels=1e9
        )
        return ee.Feature(None, {
            'date': date,
            'NDVI': stats.get('NDVI'),
            'cloud_pct': cloud_pct
        })

    features = s2.map(compute_ndvi).getInfo()['features']
    data = [f['properties'] for f in features if f['properties']['NDVI'] is not None]
    
    df = pd.DataFrame(data)
    df['date'] = pd.to_datetime(df['date'])
    return df.sort_values('date').reset_index(drop=True)

Visualizing the Optical vs. Radar Gap

# Execute extractions
df_s1 = get_sentinel1_timeseries(aoi, start_date, end_date)
df_s2 = get_sentinel2_ndvi(aoi, start_date, end_date)

# Plotting the diagnostic comparison
fig, ax1 = plt.subplots(figsize=(12, 6))

# Plot SAR VH backscatter (Continuous biomass tracking)
color = 'tab:blue'
ax1.set_xlabel('Timeline (Kharif Season)', fontsize=12)
ax1.set_ylabel('Sentinel-1 VH Backscatter (dB)', color=color, fontsize=12)
ax1.plot(df_s1['date'], df_s1['VH'], marker='o', color=color, linewidth=2, label='SAR VH (Cloud-Free)')
ax1.tick_params(axis='y', labelcolor=color)
ax1.grid(True, linestyle='--', alpha=0.5)

# Plot Optical NDVI on secondary axis
ax2 = ax1.twinx()
color = 'tab:green'
ax2.set_ylabel('Sentinel-2 NDVI', color=color, fontsize=12)

# Separate clear scenes from cloud-contaminated scenes
clear_s2 = df_s2[df_s2['cloud_pct'] < 40]
cloudy_s2 = df_s2[df_s2['cloud_pct'] >= 40]

ax2.scatter(clear_s2['date'], clear_s2['NDVI'], color='darkgreen', s=45, label='Clear NDVI (<40% Cloud)')
ax2.scatter(cloudy_s2['date'], cloudy_s2['NDVI'], color='red', marker='x', s=40, label='Cloud Contaminated (>=40% Cloud)')
ax2.tick_params(axis='y', labelcolor=color)

plt.title('Monsoon Crop Monitoring: Sentinel-1 SAR vs. Sentinel-2 Optical Blindspot', fontsize=14, fontweight='bold')
fig.tight_layout()
plt.show()

---

4. Translating SAR Metrics into Parametric Risk Underwriting

The operational value of this pipeline lies in quantitative risk metrics for lenders and underwriters:

1. Sowing Failure Index: If regional σ⁰_VV does not drop below −16 dB within the standard planting window (June 15 – July 20), automated risk systems immediately flag a sowing failure or drought delay.
2. Submergence Duration: While optical sensors display only clouds during flood events, SAR penetrates rain clouds. Sustained specular reflection over 7–10 consecutive days post-tillering triggers parametric flood payouts automatically.
3. Standardized Biomass Anomaly: By comparing seasonal VH backscatter against a 10-year rolling SAR baseline, we calculate a standardized anomaly score:

Where μ_historical(t) represents the multi-year mean backscatter for calendar week t, and σ_historical(t) represents the historical standard deviation. This provides lenders with an objective, mid-season yield deviation indicator long before traditional harvest reporting.

---

Conclusion & What’s Next

Optical remote sensing has clear merits, but in monsoonal agricultural landscapes, treating it as a standalone monitoring source introduces unmanageable data gaps. Coupling Sentinel-1 SAR with weather reanalysis provides uninterrupted, physics-grounded agricultural intelligence.

In the next post, we will explore integrating these SAR backscatter profiles with daily ERA5-Land precipitation data inside a predictive Spatiotemporal Transformer model.