Multi-modal wildfire ignition modeling

Western fire seasons are getting longer. Washington's 2020 season alone produced record PM2.5 exposures and a measurable rise in mortality and hospitalizations.

FireFusion forecasts next-day wildfire ignition probability and ignition cause across Washington on a 2 km grid. On a held-out test set the model reaches an AUPRC of 0.62, against the 0.08 baseline any random scorer would produce on data where ignitions make up roughly 8% of valid pixels.

85-90% of recent Washington wildfires are human-caused, and ignitions now reach west of the Cascades into the densely-populated Puget Sound corridor.

Fires are, in data speak, incredibly rare. To make matter's worse, the drivers are heterogeneous and don't compose linearly - one fire can start from a combination of drought and wind, another from lightning in a generally non-burning area, and so on. Indices like the Fosberg FWI condense wind, temperature, and humidity into a single per-station scalar, but ignores topography, fuel state, and the human pressures (transportation corridors, wildland-urban interfaces, recreational patterns) that explain the the large majority of ignitions. A model that fits this domain needs to learn interactions across not only space and time, but the features themselves, and a LOT of them.

The input cube is 24 channels deep, spanning five categories that no other fire model combines in one end-to-end system (with the exception of the SeasFire DataCube, although this operates globally with a 8d temporal resolution, too high an interval for pinpoint next day prediction). Each was derived from raw products into something the model could use.

• Meteorology. Daily gridded temperature, humidity, wind speed, 100-hr fuel moisture, and 2/5-day rolling precipitation from gridMET.
• Fuels and vegetation. NDVI (a satellite measure of vegetation greenness) anomaly and LAI (leaf area in m² per m² of ground) from MODIS, plus NLCD canopy cover and fractional impervious surface. LAI uses nearest-neighbor in time to preserve sharp post-fire dropoffs.
• Topography. Elevation and slope from LANDFIRE, with aspect handled as components above.
• Human pressure. Housing density (GPW), the WUI index (wildland-urban interface: where housing borders wildland fuels) and distance-to-WUI from USDA, and distance-to-road from TIGER/Line. Long-tailed variables get log1p before standardization.
• Fire history and derived priors. Kernel-density estimators for past lightning and human ignitions, months-since-last-burn from MODIS MCD64A1, a 3x3 rolling mean of recent activity, and the Fosberg FWI. A sinusoidal day-of-year encodes seasonality without learning that day 365 and day 1 are far apart.

The model is a ConvFormer: a CNN encoder followed by three self-attention blocks on three different axes, then a two-head decoder.

Each input batch has a (B, T, C, H, W) shape batch, T lookback days, C feature channels, and an HxW spatial grid. With 24 channels over a 204x200 grid covering Washington.

A residual encoder downsamples spatial extent and lifts the channel dimension into an embedding: (B, T, C, H, W) -> (B, T, D, H', W'), (embedding dimension D, H' and W' denote downsampled spatial dims. This lets the attention blocks operate on a smaller grid while preserving local geography through skip connections.

Three attention blocks follow, each reshaping the tensor to expose a different axis as the sequence dimension before applying multi-head self-attention.

Windowed spatial attention. The (H', W') grid is partitioned into PxP windows (P=4, sized to characteristic topographic scales like a mountain ridge). Within each time window, the P² locations become tokens: (B, T, D, H', W') -> (B, T, P², D) per window. By sampling 4x4 windows, cost stays linear in pixel count rather than quadratic.

Channel-mixing attention. Next, the embedding axis becomes the sequence: (B, T, D, H', W') -> (B, H', W', T, D), with MHA over D. In other words, each channel attends to every other channel at the same location and time. Two channels can be informative individually but redundant together (e.g, drought conditions, but no vegatation cover), or weakly informative individually but strongly together (e.g., low rainfall and wind speed). the attention weights distinguish those cases.

Temporal attention. Finally, the time axis: (B, H', W', T, D) with MHA over T. Each day's embedding attends to every other day in the lookback window. Temporal attention runs last because day identity has to survive through to the decoder, which reads only t=T.

Two heads share the upsampling path: a binary ignition head producing (B, 1, H_out, W_out) and a 4-class cause head producing (B, 4, H_out, W_out). Loss is masked over water, glaciers, currently-burning cells, and out-of-study regions so the model is never penalized for predicting nothing where ignition is physically impossible.

• Efficient information compression This was a learning lesson in the scaling laws, and one of the first experiments I Since attention scales quadratically in cells per window. 50m resolution to catch fuel breaks and narrow canyons wasn't feasible on a single RTX 3070 I shipped at 2km and ran with batch_size=2. Next time I'd design hierarchical attention from the start: coarse globally, fine in high-risk regions, instead of uniform windows over the whole grid.
• Use a stratified hold-out set. I split the test set temporally on the WA grid, which probes generalization across years but not across ecoregions. A more sound approach would have been to randomly sample pixels across time, ecoregion, and space.
• Don't pretend an 8% positive rate is balanced. I oversampled positive windows and weighted the binary loss ~6000:1. Without it the model collapses to no and the AUPRC story disappears.
• Multi-task only when the labels can carry it. The cause head added value during architecture exploration but didn't earn its keep on the held-out set. Next time I'd train the binary head solo, then bolt cause on as a frozen-encoder probe.