Core Concepts

Rasters are grids of pixels (satellite imagery, DEMs). Vectors are points, lines, and polygons (boundaries, roads, alert zones). Your pipeline ingests rasters and outputs vector alerts.

The math that maps Earth's curved surface to flat coordinates. Cuenca uses EPSG:32717 (UTM 17S) for accurate area calculations. Using the wrong CRS silently corrupts every area measurement.

All flat maps distort reality. UTM preserves area locally (good for 'how many hectares were cleared?'). Web Mercator (EPSG:3857) distorts area badly at the equator — never calculate forest area in it.

The standard raster format — a TIFF with embedded CRS and georeferencing metadata. Cloud-Optimized GeoTIFFs (COGs) allow reading just the pixels you need without downloading the whole file.

A database index structure that makes 'find all alerts within this polygon' queries fast. Without it, PostGIS scans every row. With it, queries drop from seconds to milliseconds.

Satellites measure reflected energy at specific wavelengths. Visible bands show color; near-infrared reveals vegetation health; shortwave-infrared detects moisture and bare soil. Sentinel-2 has 13 bands across these regions.

L1C measures what the satellite sensor receives (includes atmospheric haze). L2A corrects for the atmosphere to estimate what the ground actually reflects. Always use L2A for analysis.

Clouds block the ground in optical imagery. In Cuenca, 50-70% of scenes are cloud-contaminated. Masking methods (s2cloudless, Cloud Score+) identify and remove cloud pixels before analysis.

Active microwave sensor that transmits its own energy and measures the return. Clouds are transparent to microwaves, so SAR provides continuous monitoring regardless of weather — critical for the Andes.

The ground area each pixel represents. Sentinel-2 gives 10m pixels (a pixel covers 10×10 meters). Higher resolution detects smaller clearings but generates more data.

How often a satellite revisits the same area. Sentinel-2 revisits every 5 days, Sentinel-1 every 12 days. More frequent revisits detect changes faster.

(NIR - Red) / (NIR + Red). Healthy vegetation reflects strongly in NIR. NDVI ranges from -1 to 1; dense forest is typically 0.6-0.9. Saturates in very dense canopy — less useful for detecting degradation under intact canopy.

(NIR - SWIR) / (NIR + SWIR). Sensitive to moisture and vegetation structure changes. dNBR (differenced NBR between two dates) is the standard metric for detecting clearing and burn severity.

Improved NDVI that corrects for atmospheric effects and reduces saturation in dense vegetation. Better than NDVI for tropical forests.

Detects forest degradation (selective logging, thinning) that NDVI misses. Uses spectral unmixing to estimate sub-pixel fractions of green vegetation, soil, and non-photosynthetic vegetation.

Nearby pixels are more similar than distant ones. This violates the independence assumption of standard ML validation. If your training and test data are spatially close, accuracy is inflated because the model memorizes local patterns instead of learning generalizable features.

Split data into spatial blocks (e.g., 5km grid cells) and ensure no block appears in both training and test sets. This gives honest accuracy estimates that predict real-world performance. Typically shows 5-15% lower accuracy than random splits.

Measures how much accuracy drops when a feature's values are randomly shuffled. Use this instead of sklearn's default impurity-based importance, which is biased toward correlated and high-cardinality features.

Maps where your model is making predictions within its training domain vs. where it's extrapolating into unknown feature space. Shows clients exactly where your predictions are trustworthy.

The standard methodology for reporting land-change accuracy. Requires stratified random sampling, area-weighted accuracy metrics, and confidence intervals on area estimates. Non-negotiable for credibility with government and scientific clients.

Don't just count correctly classified pixels — weight by the area each class represents. A model that gets 99% of 'no change' right but only 50% of 'deforestation' right is useless for monitoring, even though overall accuracy looks great.

University of Maryland's near-real-time deforestation alerts using Landsat (30m, GLAD-L) and Sentinel-2 (10m, GLAD-S2). Global coverage. Tuned for humid tropical lowland forest — underperforms in montane and fragmented forest.

Wageningen University's radar-based alerts using Sentinel-1 SAR. Works through clouds. Lower spatial detail than optical but continuous temporal coverage. Critical for cloudy regions like the Andes.

World Resources Institute platform that integrates GLAD, RADD, and DIST-ALERT into a single dashboard. The benchmark your product competes with — and the system whose gaps define your market opportunity.

PostgreSQL extension for spatial data. Stores geometries, supports spatial queries (ST_Intersects, ST_Within, ST_Area), and uses GiST indexes for performance. The standard spatial database for production systems.

A standard API for searching and discovering satellite imagery. Instead of downloading scenes manually, you query a STAC catalog for imagery matching your AOI, date range, and cloud cover — then stream just the pixels you need.

A GeoTIFF organized so you can read any spatial subset via HTTP range requests without downloading the whole file. The foundation of cloud-native geospatial architecture.

A dynamic raster tile server that reads COGs and serves map tiles on-the-fly. Lets web maps display your detection results without pre-rendering millions of tiles.