Reading the report.

Level-2A is the ESA processing level that converts top-of-atmosphere radiance into surface reflectance, removing the atmospheric column. We exclusively use L2A as our raster basis because the spectral indices below are only meaningful on surface reflectance, not raw radiance.

Every project is bounded by an AOI. We buffer 500 m beyond the polygon to capture pixels whose convolution windows would otherwise include nodata, and clip all final deliverables back to the original polygon.

We use the L2A scene-classification layer plus a custom shadow refinement. Scenes with >10% AOI cloud coverage are dropped entirely; partial cloud is masked per pixel and the median composite fills the gaps.

A simple red/blue ratio that exploits the broad Fe³⁺ absorption in the visible. High values flag oxidised iron — a primary signal for weathered sulphide caps and goethite-rich altered ground.

The 2.20 µm Al-OH bond absorption straddles Sentinel-2 bands 11 and 12. Elevated B11/B12 suggests phyllic / argillic alteration zones common above porphyry and epithermal systems.

High NDVI means more vegetation, less visible rock. We invert it to weight bare-ground pixels and to suppress false anomalies driven by chlorophyll absorption rather than mineralogy.

Extracted from Sobel-filtered hillshade plus a Hough transform on the multi-azimuth derivatives. Filtered for length, persistence across azimuths and proximity to known structural trends.

Spectral indices, raw bands, terrain derivatives (slope, aspect, hillshade), distance-to-lineament, distance-to-drainage and a handful of texture metrics — all resampled to a common 10 m grid and stacked.

We use XGBoost in pairwise-ranking mode. It handles missing data, scales to millions of pixels and — crucially for clients who want to question the model — exposes feature-importance and per-prediction SHAP values.

Standard k-fold leaks geographic information. We hold back whole licences from the regional analogue set, train on the rest, and evaluate. The reported AUC and ROC are always from licences the model has never seen.

A scalar in [0, 1]. 0.5 is random, 1.0 is perfect. Our published rolling mean is 0.89 across twelve regional analogues; expect ±0.05 variation per project depending on data quality and analogue overlap.

TP / (TP + FP). High precision means few false alarms, but high precision usually trades off against recall. We tune the operating point per project after a kickoff conversation about tolerance for false negatives.

TP / (TP + FN). For early-stage targeting, recall typically matters more than precision — a missed deposit is more expensive than a wasted hole.

A = ≥80% model confidence, multiple converging signals. B = 70–80%, a single dominant signal. C = 60–70%, marginal but worth ground-truthing if cheap. Below 60% is not reported.

Eight layers: prospectivity, each spectral index, lineament density, drainage proximity. Cloud-Optimised GeoTIFF, EPSG:32xx UTM matched to the AOI, internal overviews built.

Printed on the report cover. Send it back with any question about a specific figure and we can spin up the exact build container that produced it.