← Back to the map

METHOD & INTERPRETATION

What this map shows

This atlas displays existing 2022 predictions from the local Pacific Coast histogram gradient boosting models. The models learn Kelpwatch-style canopy labels from AlphaEarth satellite embedding features. Kelpwatch is a Landsat-derived weak-label reference, not independent field truth. Canopy area is not kelp biomass.

Three model surfaces

Presence probability is calibrated to the saved binary model. Conditional canopy is predicted area given presence. Expected canopy combines the two by multiplying probability and conditional area. Area layers use m² per 30 m source cell, with a fixed 0–900 m² display scale.

Held-out year and evaluation

This map uses the existing local Pacific Coast year-split run. The entire display year, 2022, belongs to its held-out test set. Both error layers cover all available 2022 predictions, rather than selected geographic bands. This evaluates temporal performance within the modeled region; it does not establish spatial generalization.

Signed errors are predicted minus reference area. Presence errors use the saved probability threshold of 0.41; reference presence is canopy area ≥90 m² per 30 m cell (10%). Assumed-background false positives remain distinct from observed-label false positives.

Coverage and scale

Transparent pixels mean missing or excluded coverage, not zero canopy. Label support distinguishes observed labels from assumed background. The domain layer shows retained and excluded cells within the current supported footprint. Retention does not guarantee available embeddings or predictions; missing coverage stays transparent. The PCA colors reuse one fixed coastwide projection and stretch; they describe embedding structure, not canopy classes.

A separate prediction-supported footprint-edge category preserves existing predictions affected by differing geographic and projected polygon-edge conventions. This display reconciliation changes neither the source domain policy nor the predictions. Native footprint edges are resolved using geographic cell centers.

Detail comes from a 30 m source grid. Raster tiles cover zooms 5–13; zooming farther enlarges existing tiles without adding source detail. Overviews use nearest-source-cell sampling, not canopy totals or area averages. Small patches can disappear at low zoom. Display colors clip values beyond the legend endpoints.

Sources and attribution

The browser loads only map tiles and the supported-footprint outline. Raw source data, trained models, and prediction tables are not part of the public map bundle.