Turn any fire lookout
into a calibrated sensor

• Frame screeningBrightness, Laplacian sharpness, contrast and clipping checks drop night, glare and blur before matching.
• SIFT on CLAHE greyscaleContrast-limited histogram equalisation evens out season and time of day; matches pass Lowe's ratio test at 0.75.
• 4-DOF similarity by RANSACShift, rotation and scale only. A full homography overfits to clouds and moving shadows.
• Keyframe trackingFrames match recent keyframes, relocalise against older ones, and start a new stretch after four lost frames.
• Loop closureStretches join through a maximum-inlier spanning tree over their clearest keyframes.
• Relocations and bumpsA camera moved to a new mast becomes a second view; sustained shifts are logged as move dates.

fuelgauge register-archive photos/ out/ridge

• Sky segmentationSegFormer-B2 (ADE20K) through ONNX Runtime on CPU; the skyline is the first non-sky row per column, snapped to the nearest strong edge.
• Rendered terrain panoramaCopernicus 30 m elevation ray-marched along every bearing with earth curvature and refraction (k = 0.13).
• Fisheye lens modelEquidistant projection with one radial term, θ = (r/f)(1 + k₁(r/f)²). Fire cameras are strongly barrel-distorted.
• Robust pose fitBounded Powell search on a trimmed loss that ignores the worst 20% of columns: trees by the mast, low cloud.
• Network lensOne field of view and distortion fitted jointly across all cameras of a model (101.15°, k₁ = 0.080 here).

fuelgauge calibrate ridge frame.jpg --lat 33.4008 --lon -117.1905 --elev 483 --yaw 90

• Back-projectionEach block's viewing ray is intersected with the rendered panorama; the first terrain cell it reaches is its ground point.
• Terrain attributesSlope and aspect from elevation gradients; ESA WorldCover 10 m land cover at the ground point.
• Usable-ground maskSky, the camera housing and blocks past the terrain model's reach are dropped before measuring.

ground = terrain.backproject(cam, pano)   # distance, lat, lon per block

• Haze scoringSobel edge correlation with the anchor frame; a day is clear at 80% of its trailing 90-day norm.
• Automatic regionsSegmentation vegetation labels plus a strong seasonal swing (5th to 95th percentile of weekly medians). No hand-drawn masks.
• Drift-tracking white balanceA von Kries correction against blocks that stay flat all year, smoothed over 61 days so it follows the sensor, not snow.
• IndicesGRVI and GCC per block; camera NDVI from near-infrared twins with exposure correction (Petach et al. 2014).
• Causal smoothingWeekly 90th percentile then a 15-day mean, using only past data, so it is safe as a nowcast.

fuelgauge measure out/ridge   # _daily.csv + _regions.png

• Radial viewshedsRays every 0.1° out to 30 km with curvature and refraction; a sample is visible when it beats every elevation angle before it.
• Smoke line of sightWould a column 100 or 300 m above a point rise into view? That is how a lookout spots a fire before its base is visible.
• Greedy sitingHilltop candidates in every 6 km square; each round takes the one that adds the most unseen ground, counted once.

fuelgauge viewshed cameras.csv coverage.tif
fuelgauge site cameras.csv --k 10