8 min

Satellite & Remote Sensing AI

Extracting Intelligence from Earth Observation Data

Beyond RGB: Spectral Intelligence

A satellite camera is not a phone camera. Earth observation sensors capture data across electromagnetic bands that reveal information invisible to human eyes. Each band tells a different story about the surface below.

Band	Wavelength (nm)	What It Reveals	Key Applications
Blue	450-520	Water depth, atmospheric scattering	Coastal mapping, bathymetry
Green	520-600	Vegetation vigour, turbidity	Crop health, water quality
Red	630-690	Chlorophyll absorption	Vegetation stress, urban mapping
Red Edge	700-750	Early stress detection in plants	Precision agriculture, forest health
NIR	750-1000	Cell structure reflectance	Biomass estimation, flood mapping
SWIR	1500-2500	Moisture content, mineral identification	Drought monitoring, geological survey
Thermal IR	10000-12000	Surface temperature	Urban heat islands, fire detection

AI processes these bands not individually but in combination — computing spectral indices, learning band-ratio signatures, and identifying patterns across millions of pixels that no human analyst could review manually.

NDVI and Beyond: Vegetation Intelligence

The Normalized Difference Vegetation Index (NDVI) is the most widely used spectral index, calculated as:

NDVI = (NIR - Red) / (NIR + Red)

Healthy vegetation strongly reflects NIR and absorbs Red light, producing high NDVI values. The interpretation scale:

NDVI Range	Surface Condition	Action Trigger
-1.0 to 0.0	Water, snow, clouds	Mask from analysis
0.0 to 0.2	Barren soil, rock, urban	Land use classification
0.2 to 0.4	Sparse/stressed vegetation	Alert: drought/disease investigation
0.4 to 0.6	Moderate vegetation	Normal monitoring
0.6 to 0.9	Dense, healthy vegetation	Healthy crop / forest canopy

But NDVI saturates in dense canopies and is sensitive to soil background. Modern AI-driven analysis uses enhanced indices:

EVI (Enhanced Vegetation Index): Corrects for atmospheric and soil effects, better in dense canopies

SAVI (Soil-Adjusted Vegetation Index): Adds a soil brightness correction factor L

NDWI (Normalized Difference Water Index): Uses Green and NIR to map water bodies and moisture content

NBR (Normalized Burn Ratio): NIR and SWIR combination for fire scar mapping

Open data/ndvi-readings.csv — it contains time-series NDVI data for agricultural plots across Punjab, Maharashtra, and Karnataka spanning three crop seasons. Notice the seasonal patterns and anomalies that signal irrigation failures or pest attacks.

ISRO's Earth Observation Constellation

India operates one of the world's most capable civilian Earth observation programmes through ISRO:

Satellite Series	Resolution	Revisit	Primary Use
Cartosat-2/3	0.65m PAN, 2m MX	5 days	Urban planning, infrastructure mapping
ResourceSat-2/2A	5.8m (LISS-IV), 23m (LISS-III)	24 days	Agriculture, water resources
EOS-04 (RISAT-1A)	1-50m SAR	12 days	All-weather imaging, flood mapping
EOS-06 (OceanSat-3)	360m-1km	2 days	Ocean colour, sea surface temp
INSAT-3D/3DR	1-8 km	30 min	Weather, cyclone tracking

The resolution-coverage trade-off is fundamental: Cartosat gives you building-level detail but covers a narrow swath (few km). ResourceSat covers 140 km swaths but cannot resolve individual trees. INSAT covers the entire subcontinent every 30 minutes but at kilometre-scale resolution.

AI helps bridge this gap through super-resolution — training models on paired low-res/high-res imagery to predict high-resolution outputs from low-resolution inputs. A well-trained model can effectively increase ResourceSat resolution by 2-4x, enabling agricultural monitoring at near-Cartosat quality across much wider areas.

Cloud Masking: The Persistent Challenge

India's monsoon season (June-September) coincides with the kharif crop season — precisely when satellite monitoring is most critical and cloud cover makes it nearly impossible. In a typical monsoon month, 60-80% of optical satellite passes over central India are cloud-contaminated.

AI-driven cloud masking goes beyond simple threshold-based methods:

Semantic segmentation models trained on labelled cloud/shadow datasets (e.g., Sentinel-2 Cloud Probability)

Temporal compositing — combining multiple partially cloudy images from a 10-15 day window to produce a cloud-free composite

SAR fusion — Synthetic Aperture Radar (like ISRO's EOS-04) penetrates clouds. AI models fuse SAR and optical data to provide continuous monitoring regardless of weather

Thin cloud removal — Deep learning models that subtract haze and thin cirrus contamination, recovering usable surface reflectance

Change Detection: What Moved, What Grew, What Disappeared

AI-powered change detection compares multi-temporal imagery to identify:

Urban expansion — new construction, road development, informal settlements

Deforestation — forest cover loss, encroachment into protected areas

Water body changes — reservoir levels, river course migration, wetland shrinkage

Agricultural patterns — crop type rotation, fallow land identification, irrigation infrastructure

The technical pipeline:

Prompt: "Analyze these two satellite image patches from the same location, dated March 2025
and March 2026. Both are 10m Sentinel-2 multispectral composites. Identify and classify
all changes into categories: new construction, vegetation loss, vegetation gain, water
body change, agricultural change. Report change area in hectares and confidence score."

Open data/satellite-image-metadata.json for metadata on a curated dataset of Indian satellite imagery pairs — before/after urban expansion in Gurugram, deforestation patches in Jharkhand, and agricultural intensification in Tamil Nadu.

Indian Applications in Depth

Punjab Wheat Monitoring

India's food security depends heavily on Punjab's wheat belt. AI + satellite monitoring enables:

Sowing date estimation from NDVI time-series onset (critical for yield forecasting)

Crop area estimation — distinguishing wheat from mustard, potato, and fallow land using multi-temporal spectral signatures

Yield prediction — correlating mid-season NDVI with historical yield data, achieving 85-90% accuracy at district level by February (harvest is in April)

Residue burning detection — identifying active fires and burn scars from thermal and NBR data, feeding into the Commission for Air Quality Management's enforcement system

Sundarbans Mangrove Tracking

The world's largest mangrove forest straddles the India-Bangladesh border and is shrinking due to cyclones, salinity intrusion, and human encroachment. AI monitors:

Mangrove density from NDVI/EVI gradients — distinguishing dense, moderate, and degraded mangrove zones

Shoreline erosion — tracking the land-water boundary migration at sub-pixel accuracy using spectral unmixing

Cyclone damage assessment — pre/post event comparison after events like Cyclone Amphan (2020), quantifying canopy loss within 48 hours

Mumbai Urban Sprawl

India's most densely populated city provides a laboratory for urban remote sensing:

Impervious surface mapping — tracking concrete/asphalt expansion vs green space loss

Heat island analysis — thermal band data showing 4-8°C temperature differentials between built-up and vegetated areas

Informal settlement detection — AI classifies slum areas from spectral and texture features, informing urban planning and disaster preparedness

Coastal zone monitoring — tracking reclamation and mangrove destruction along the western shoreline

Pixxel's Hyperspectral Constellation

Bengaluru-based Pixxel is deploying the world's highest-resolution commercial hyperspectral constellation — capturing 150+ narrow spectral bands at 5m resolution compared to Sentinel-2's 13 bands at 10m. This enables:

Mineral identification from spectral fingerprints (each mineral has a unique absorption signature)

Crop disease detection before visible symptoms appear (subtle chlorophyll fluorescence changes)

Water quality assessment (algal bloom species identification from spectral signature)

Methane plume detection from SWIR absorption features

Open data/land-use-classifications.json for a labelled dataset of land use categories across different Indian landscapes, suitable for training a multi-class classifier on satellite imagery.

Key Takeaways

Multispectral analysis is the foundation — AI on satellite imagery is only as good as the spectral features you feed it. Understanding which bands reveal which phenomena is prerequisite knowledge.

Cloud cover is India's biggest operational challenge — monsoon-season monitoring requires SAR fusion, temporal compositing, or cloud-removal networks. Optical-only pipelines fail for 4 months of the year.

ISRO provides a strong data backbone — between Cartosat, ResourceSat, and EOS-04, India has sovereign access to high-quality Earth observation data. AI multiplies its value.

Hyperspectral is the next frontier — Pixxel and similar constellations will enable AI applications (mineral mapping, early disease detection, pollution monitoring) that multispectral data physically cannot support.

This is chapter 2 of AI for Aerospace & Drones.

Get the full hands-on course — free during early access. Build the complete system. Your projects become your portfolio.

View course details

Ch. 1: AI for Flight Path Optimization

Ch. 3: Predictive Maintenance for Aircraft