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 the US Corn Belt, the Central Valley of California, and the Canadian Prairies spanning three growing seasons. Notice the seasonal patterns and anomalies that signal irrigation failures or pest attacks.

The Open Earth Observation Backbone: Landsat & Copernicus

Western remote sensing rests on two pillars of free, open imagery, complemented by a thriving commercial sector:

Satellite / Constellation	Resolution	Revisit	Primary Use
Landsat 8/9 (NASA/USGS)	15m PAN, 30m MX, 100m TIR	8 days (combined)	Long-term land change, agriculture, thermal
Sentinel-2 (ESA Copernicus)	10m / 20m / 60m, 13 bands	5 days	Agriculture, land cover, vegetation
Sentinel-1 (ESA Copernicus)	5-40m C-band SAR	6-12 days	All-weather imaging, flood mapping, subsidence
Planet (Dove/SkySat)	3m / 0.5m	Daily (Dove)	High-frequency monitoring, change detection
Maxar (WorldView)	0.3m	1-3 days	Defense, mapping, disaster response
Capella / ICEYE	0.5-1m X-band SAR	On-demand	Persistent all-weather commercial SAR

The resolution-coverage trade-off is fundamental: Maxar WorldView gives you building-level detail but covers a narrow swath. Sentinel-2 covers a 290 km swath but cannot resolve individual trees. Geostationary weather satellites (GOES, Meteosat) cover entire hemispheres every few 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 Sentinel-2 resolution by 2-4x, enabling agricultural monitoring at near-commercial quality across much wider areas at zero imagery cost.

Cloud Masking: The Persistent Challenge

The peak growing season in many regions coincides with maximum cloud cover — precisely when satellite monitoring is most critical. In a humid summer month, 50-70% of optical satellite passes over the US Southeast or Northern Europe can be 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, the Landsat CFMask training data)

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

SAR fusion — Synthetic Aperture Radar (Sentinel-1, Capella, ICEYE) 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, infrastructure growth

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 satellite imagery pairs — before/after suburban expansion in Phoenix, wildfire burn scars in California, and agricultural intensification in the Netherlands.

Applications in Depth

US Corn Belt Yield Monitoring

US food and feed security depends heavily on the Midwest corn and soybean belt. AI + satellite monitoring enables:

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

Crop type classification — distinguishing corn from soybeans, wheat, and fallow land using multi-temporal spectral signatures, validated against USDA Cropland Data Layer

Yield prediction — correlating mid-season NDVI with historical USDA NASS yield data, achieving 85-90% accuracy at county level before harvest

Stress and tile-drainage detection — identifying drought stress and field anomalies that feed into crop insurance and commodity-market models

Everglades & Coastal Wetland Tracking

The Florida Everglades and coastal Gulf wetlands are shrinking due to hurricanes, saltwater intrusion, and development. AI monitors:

Vegetation density from NDVI/EVI gradients — distinguishing healthy, moderate, and degraded wetland zones

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

Hurricane damage assessment — pre/post event comparison after storms, quantifying canopy and marsh loss within 48 hours using Maxar tasking and Sentinel-1 SAR

Phoenix Urban Sprawl

One of the fastest-growing US metro areas provides a laboratory for urban remote sensing:

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

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

Land-use change detection — AI classifies new subdivisions and commercial development from spectral and texture features, informing planning and water-management decisions

Wildland-urban interface monitoring — tracking development encroachment into fire-prone terrain

Planet & Maxar Commercial Constellations

US commercial providers have redefined cadence and resolution. Planet flies a fleet of Dove cubesats that image the entire landmass daily at 3m, plus SkySats at 0.5m, while Maxar delivers 30 cm imagery from WorldView. Combined with hyperspectral newcomers, 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 (e.g., the Carbon Mapper / EMIT programs)

Open data/land-use-classifications.json for a labelled dataset of land use categories across different North American and European 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 the biggest operational challenge — peak-season monitoring requires SAR fusion, temporal compositing, or cloud-removal networks. Optical-only pipelines fail during cloudy months.

Landsat and Copernicus provide a free, open data backbone — between Landsat 8/9 and Sentinel-1/2, organizations have sovereign-grade Earth observation data at zero cost. AI multiplies its value.

Hyperspectral and daily commercial cadence are the frontier — Planet's daily imaging, Maxar's 30 cm detail, and emerging hyperspectral missions enable AI applications (mineral mapping, early disease detection, pollution monitoring) that legacy multispectral data physically cannot support.

This is chapter 2 of AI for Aerospace & Drones (Global).

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