Technical note — Colombia run (wildfire Bayesian posterior + expected exposure)

 

 

This note documents a single Colombia execution of the calibrated Bayesian wildfire-risk prototype in Google Earth Engine. It is a prospective model run (not a fire detection report): the main output is a posterior probability surface (“final_threat”, 0–1) derived from fusing (i) a learned/empirical prior of where fires tend to occur, with (ii) a near-term fire-weather likelihood proxy. The summary table reports posterior-weighted expected exposure (expected value under the posterior), not observed damage.

Run configuration (as executed)

• AOI: Colombia (country boundary)

• runMode: debugging

• Posterior computation scale: fineScale = 2000 m

• Map display scale: displayScale = 5000 m

• Forecast window: ≤ 24 h

• Smoothing: smoothMeters = 0 (disabled in debugging)

• Exposure aggregation: posterior-weighted expected exposure without classes

• Key exposure scales: population 1000 m; roads raster 9000 m; line buffer 500 m (for line-length approximation)

Model outputs produced in this run

1. Posterior risk probability map (“Wildfire posterior risk probability”): a spatial surface of posterior probability after log-odds fusion.

2. Expected exposure panel (“Posterior-weighted expected exposure (NO classes)”): percent of each asset total that lies under posterior probability mass, plus absolute expected exposure.

Colombia results (posterior-weighted expected exposure)

From the run panel:

• Population (expected exposed): 1.43% → 469,797 people (of 32,927,972)

• Cropland area (expected exposed): 0.54% → 2.51 ha (of 460.84 ha)

• Urban area (expected exposed): 2.03% → 5.11 ha (of 252.43 ha)

• Oil palm area (expected exposed): 8.09% → 7.61 ha (of 93.99 ha)

• Protected areas (expected exposed): 1.79% → 215.78 ha (of 12,030.74 ha)

• Houses proxy (built-up footprint, expected): 4.06% → 2,808.55 ha (of 69,215.12 ha)

• Electric grid length (expected exposed): 2.80% → 64.08 km (of 2,287.86 km)

• Roads length / Oil & gas sites / Wind farms / Solar farms: N/A in this run (no value returned in the panel).

Interpretation of this run (what these numbers mean)

• At national scale, the posterior implies modest expected exposure for population (≈ 470k expected exposed out of ≈ 32.9M). This is consistent with a posterior surface that concentrates probability over a subset of the country rather than spreading high probability broadly.

• The built-up footprint expectation is non-trivial (2,808.55 ha, 4.06% of built-up total in the mask used). In this run, the posterior overlaps with built-up areas more than with cropland or protected areas (in percentage terms), suggesting the probability mass is not confined to remote landscapes only.

• The oil palm percentage is the highest among the land-use masks (8.09%). Given the relatively small total oil-palm mask area reported (93.99 ha in this configuration), this percentage is sensitive to spatial coincidence: a concentrated posterior overlapping a small mask can produce a high percentage even when the absolute exposed area is small (7.61 ha).

• Protected areas show 215.78 ha expected exposed (1.79%). This indicates some posterior probability mass intersects protected polygons, but at a relatively low share of the protected-area total.

• Electric grid exposure is measurable (64.08 km, 2.80%), indicating the line-buffer expected-length method executed successfully for that layer in this run.

• The N/A outputs (roads and point infrastructure) should be treated as “not available for this execution” rather than zeros; they are not interpretable as evidence of no exposure without confirming why the values did not resolve.

Notes on comparability

These results are specific to debugging mode (coarser fineScale, no smoothing, short forecast window). They are best interpreted as a documented baseline run of the pipeline for Colombia; direct comparison to other runs should hold constant the runMode, forecast horizon, and exposure masks/scales.