Fireline: forecasting tomorrow's wildfire from satellites

Fireline predicts which 375 m pixels of an active wildfire will be burning tomorrow, from the last five days of satellite, weather and terrain data. It also tests whether self-supervised pretraining (a JEPA) helps when labelled fires are scarce.

Dixie Fire, Northern California · 2021, test year

Forecast Scored

10 km

Aug 18 → Aug 19: 2,499 fire pixels today, 1,909 tomorrow.

• Fire today
• Predicted fire tomorrow (darker = more likely)
• Actual fire tomorrow

• Predicted and burned
• Missed
• False alarm
• Fire today

←

Aug 18 → Aug 19 351 km² burning today → 268 km² tomorrow · F1 0.65

→

Start Jul 13 Largest day Aug 4 Typical day Aug 18 Best day Aug 30

Three findings carry the report. The model is good where the fire already is and poor where it jumps: most of its skill is within 750 m of today’s fire, and the largest day of the Dixie Fire in the timeline is mostly missed (Analysis). JEPA pretraining on the same fires did not beat a randomly initialised frozen encoder by more than run-to-run noise (Does JEPA pretraining help?). And the benchmark’s vegetation channels contain information from after the prediction day (A leak in the benchmark).

Task

Given five consecutive days of data for one fire, predict a probability of active fire for every pixel on the next day. The label is the next day’s VIIRS active-fire detection. Fire pixels are rare (0.3% of valid pixels), so results are reported as average precision (AP) over all pixels, and as F1 at a probability of 0.5.

Data

WildfireSpreadTS Gerard et al., WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction (NeurIPS 2023 Datasets and Benchmarks) proceedings.neurips.cc has one image per fire and day for 607 US fires from 2018 to 2021, with 23 channels: active fire, satellite reflectance and vegetation, observed and forecast weather, terrain and land cover.

The split is by year, as a forecaster would face it: train on 2018–2019, choose every model and setting on 2020, and score 2021 once at the end. The forecast wind direction channel is dropped (it is wrongly encoded in the dataset), and observed wind is converted to east and north components.

Model

The model has to combine three kinds of information: where the fire is now and how it moved over the last days, the landscape it is moving into, and the weather. It is a U-Net, the standard architecture for per-pixel prediction, with one change: an encoder that reads one day at a time, followed by a step that fuses the five days. The figure follows one real sample through the network, the Dixie Fire on 18 August 2021, predicting 19 August.

AInputsfive days per sample · Aug 14 – Aug 18, predicting Aug 19

1. day −4 Aug 14
2. day −3 Aug 15
3. day −2 Aug 16
4. day −1 Aug 17
5. today Aug 18

Active fire

no composite for this day

Vegetation (NDVI)

Energy release component

Max temperature

Wind speed

Elevation (shaded)

Six of 18 daily channels on Aug 18, plus 21 static channels (terrain, downslope direction, 17 land-cover classes). Weather comes at about 4.6 km, so it looks blocky at 375 m; vegetation composites have gaps, like this one.

↓ each day separately

BEncoderone CNN, the same weights for every day · 621,824 parameters

1. Residual conv block 32 channels · 1 scale (375 m pixels) · full resolution
2. Residual conv block, stride 2 64 channels · ½ scale (750 m pixels)
3. Residual conv block, stride 2 128 channels · ¼ scale (1.5 km pixels)
4. Residual conv block, stride 2 128 channels · ⅛ scale (3 km pixels) · latent grid

Each block: two 3 × 3 convolutions with group normalisation and GELU, plus a skip connection. Output: four feature maps per day, from fine detail at full resolution to a coarse 1/8 grid.

Full-resolution features, 32 channels

Aug 14

Aug 15

Aug 16

Aug 17

Aug 18

1/8 features, 128 channels

Aug 14

Aug 15

Aug 16

Aug 17

Aug 18

Colours are the first three principal components of each scale's features. The full-resolution maps follow terrain, vegetation and the fire; on the coarse grid the active fire areas stand out, and they move from day to day.

↓ five days per scale

CFusion over daysat each of the four scales

Stack the five days, mix with a 1 × 1 convolution

5 × C channels → C channels, then group normalisation and GELU. The convolution learns how much each day and feature matters at every pixel: what burned today, what burned before, and how the weather changed.

• 1: 160 → 32
• ½: 320 → 64
• ¼: 640 → 128
• ⅛: 640 → 128

Fused 1/8 features

↓ coarse to fine

DDecoderU-Net upsampling with skip connections

1. Upsample ⅛ → ¼, join fused ¼ featuresresidual conv block, 128 channels
2. Upsample ¼ → ½, join fused ½ featuresresidual conv block, 64 channels
3. Upsample ½ → 1, join fused full-resolution featuresresidual conv block, 32 channels
4. Join today's fire maskraw, 1 channel, at full resolution
5. Residual conv block, 1 × 1 convolution, sigmoidprobability of fire tomorrow

Output for Aug 19: predicted (blue), actual (red outline), today's fire (orange)

Fire fronts are a few pixels wide. Feeding today's fire mask straight to the last layer keeps them sharp; the deep path alone blurs them.

Training. 1,508,481 parameters. Loss: binary cross-entropy + soft Dice, over valid pixels (Dice counters the 0.3% fire share). AdamW, learning rate 0.001, weight decay 0.0001; 1 warm-up epoch, then cosine decay; gradient norm clipped at 1; batch 32; 30 epochs on 128 × 128 crops biased towards fire, with random flips and rotations. The epoch with the best validation AP is kept.

Inference. Whole images, padded by reflection to a multiple of 32; the padding is excluded from all scores.

Two design choices matter for the rest of the report. Because the encoder sees a single day, a self-supervised encoder of a single day can later be swapped in and frozen without changing anything else (the JEPA experiment). And because today’s fire mask reaches the output layer directly, the model can always fall back to “the fire stays where it is”, which is a strong prior: about half of tomorrow’s fire is burning today.

Results

How the Forecasts Are Scored

The model outputs a probability for every pixel. Turning that into a yes/no map needs a threshold, and every threshold trades two kinds of error: flag too much and most flagged pixels do not burn (low precision); flag too little and most of the fire is missed (low recall). Move the threshold below on one real day and watch the trade-off.

Threshold

0.50

Flag a pixel as "fire tomorrow" when the model's probability is at least this.

Hits

–

flagged, and burned

False alarms

–

flagged, no fire

Misses

–

burned, not flagged

Precision

–

hits ÷ flagged: when it says fire, how often is it right?

Recall

–

hits ÷ burned: how much of tomorrow's fire does it find?

F1

–

balance of the two at this threshold

This day's AP

–

AP (average precision) is the shaded area: precision averaged over every recall level, i.e. over all thresholds at once. 1.0 means every burning pixel is ranked above every other; random scores get the fire share, here –.

Average precision summarises the whole precision–recall curve in one number, so it does not depend on a chosen threshold. It is the right headline for a rare class: a model that says “no fire” everywhere is 99.7% accurate here, but its AP is the fire share, 0.003. All AP values in this report pool every valid pixel of every evaluated fire-day, so large fire days weigh more than small ones. F1 at a threshold of 0.5 is reported alongside, as the score of a fixed yes/no map.

Model Against Baselines

Two baselines use only today’s fire. Persistence predicts that tomorrow’s fire is today’s. The neighbourhood score is the share of fire among each pixel’s 5 × 5 neighbours (1.9 km), so pixels next to the fire rank above distant ones.

On the test year the model reaches AP 0.567 against 0.407 and 0.287. Its gain over persistence in F1 is smaller (0.59 against 0.53), because persistence is already a decent yes/no map: most of tomorrow’s fire is today’s. What the model adds is ranking: it puts the pixels most likely to burn first, including pixels next to the fire that persistence cannot flag.

Scores are higher on 2021 than on the 2020 validation year for every method, the baselines included (persistence 0.287 against 0.196). 2021 has less fire far from the previous day’s, so it is an easier year. Methods are compared within one year throughout.

Analysis: Where Tomorrow’s Fire Comes From

A single AP hides where the model succeeds and where it fails. This section splits tomorrow’s fire by where it appears relative to today’s.

How Far Fire Moves Overnight

80% of tomorrow’s fire is either already burning or within 750 m of today’s fire, and that is where the model’s skill is: AP 0.77 on pixels burning today and 0.41 on the first 750 m. Each ring further out roughly halves the AP. Beyond 1.9 km it is below 0.1, and for fire in images with no fire today it is 0.05. The baselines cannot reach beyond 1.9 km at all, so the model’s gain is real at every distance, but small where it matters most for warning.

Good Days and Bad Days

The same split, per day: for each test fire-day, how well the model did against how much of the next day’s fire appeared far from today’s.

Most fire-days are “close” days, and on those the model is consistent (pooled F1 0.64). On the 279 days where most of the next day’s fire is far away, F1 drops to 0.08: those are days the model does not see coming.

A Few Days Hold Most of the Far Fire

Far fire is not spread evenly. In 2020, half of it came from 36 fire-days, 1.1% of all fire-days; in 2021 it took 105 days (3.1%). These are the days that matter for warnings, such as a run or a large spot fire, and they are exactly the days a model trained on all days learns least about.

What the Inputs Can Explain

For each far fire pixel, the inputs were checked for a plausible cause, in order: was there fire nearby on the previous days (a re-flare or a missed detection), was today’s view blocked, and is the pixel downwind of today’s fire?

Most far fire has no visible cause in the inputs. Wind direction is weak evidence: a 120° downwind cone catches about a third of random directions, and 35–41% is only slightly more. The weather inputs are daily means at about 4.6 km, so the afternoon wind shift, gust front or lightning strike that starts a run is not in the data.

A Gap in the Record

The largest day of the Dixie Fire in the timeline shows how data gaps add to the problem.

Aug 3 → Aug 4 today 919 · tomorrow 0 fire pixels 0 hits · 0 missed · 916 false alarms

Aug 4 → Aug 5 today 0 · tomorrow 4,550 fire pixels 446 hits · 4,104 missed · 221 false alarms

Aug 5 → Aug 6 today 4,550 · tomorrow 3,804 fire pixels 2,462 hits · 1,342 missed · 1,263 false alarms

• Fire today
• Predicted and burned
• Missed
• False alarm

There are no fire detections for this fire on 4 August, between 919 fire pixels on the 3rd and 4,550 on the 5th: a gap in the observations, not a pause in the fire. Gaps like this cost twice. The model’s sensible prediction for the 4th is scored as 916 false alarms, and on the 5th it has no “today” fire to extend. It still finds 10% of that day’s fire from the earlier days.

Does JEPA Pretraining Help?

The Idea

Supervised training uses a thin signal. The label is tomorrow’s fire mask, 0.3% of pixels are positive, and the training years provide about 5,200 labelled fire-days. Everything else in the scene, how vegetation dries, how heat builds, how the weather changes, is learned only as far as it helps predict that mask.

A joint-embedding predictive architecture (JEPA) Assran et al., Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA, CVPR 2023) arxiv.org learns by predicting the representation of one part of the data from another, rather than its pixels. For video, the same idea learns motion and dynamics without labels Bardes et al., Revisiting Feature Prediction for Learning Visual Representations from Video (V-JEPA, 2024) arxiv.org . Fire spread is a forecasting problem, so the natural JEPA task is to predict tomorrow’s representation from the last five days.

Setup

The experiment was pre-registered: before any JEPA result, three gates were fixed that decide whether the approach is viable. The representation must not collapse (effective rank above a quarter of its 128 dimensions). The predictor must beat the trivial forecast “tomorrow’s representation equals today’s” by at least 10% on cells near the fire. And a frozen encoder must reach validation AP within 0.03 of the supervised model (0.43 or more).

Experiments and Results

None of the variants collapsed (effective rank 125–127 of 128). JEPA-next and JEPA-weighted failed the second gate: their predictor was no better than copying today’s representation (1.01 times the copy error). JEPA-change, which predicts the change, passed it with 0.67 times the copy error, but its frozen encoder reached AP 0.403 on validation, below the 0.43 bar, and 0.514 on the test year. That is 0.015 and 0.017 above the random encoder: the same direction in both years, but within the noise of a single run. The verdict, by the rule fixed in advance, is not viable.

Why It Failed

Three things worked against it, and the ladder above shows the most important one.

The encoder had little room to help. A frozen, untrained encoder with a trained fusion and decoder already reaches 0.497, 88% of the full model. Most of the skill comes from today’s fire, which reaches the decoder directly, and from what the decoder learns. Training the encoder end to end adds 0.07 AP; JEPA-change recovered about a quarter of that.

Copying was almost the right answer. In JEPA-next the representation barely changed from one day to the next: the error of copying today’s representation was 0.013, about 1% of the target’s variance. Terrain, land cover and most of the weather are the same tomorrow, so an encoder can satisfy “predict tomorrow” by encoding what does not change. Weighting the loss towards fire cells (JEPA-weighted) did not change this. Predicting the change (JEPA-change) forced the predictor to model what moves, but what moves from day to day is mostly the fire itself, which the decoder already sees.

The pretraining saw no more data than the supervised model. JEPA results in vision come from pretraining on far more unlabelled data than labelled data. Here both used the same 2018–2019 fires, about 5,200 days, so the test was whether the objective helps on the same data, not whether more data does. And the information that would explain far fire (sub-daily wind, lightning) is not in the inputs, so no objective could learn it.

Where It Could Work

The negative result is specific to this setting, and it points to where JEPA has a better chance:

• Far more unlabelled data. Daily satellite and weather data exist for every day of every year over fire-prone land, with or without fire. That is orders of magnitude more than the labelled fire-days, and it is the regime in which self-supervised pretraining usually pays off.
• Few labels. The probe here had all the labelled fires. Pretraining is most useful when labels are scarce, for example a new region or sensor; training the probe on a tenth of the fires would test that directly.
• Richer dynamics. With sub-daily inputs (geostationary fire detections every few minutes, hourly wind), “predict the next state” contains far more than the static landscape, and copying stops being a good answer.
• A shared representation. A forecasting encoder could serve several tasks at once: spread, burn severity, smoke. That is where a frozen, general encoder would earn its keep over a model trained for one task.

A Leak in the Benchmark

The NDVI and EVI2 channels are 16-day vegetation composites, updated every 8 days. The value on day t comes from the latest composite whose window started within the previous 15 days, so the window always reaches 8 to 15 days past t: part of each composite is observed after the prediction day. If a pixel burns inside that window, its vegetation index already drops at the composite update, before it burns.

The effect is modest but consistent: AUC 0.61–0.63 in every year, and the same direction in 287 of 493 fires. Results here use the channels as published, for comparability with other work on the benchmark; the leak is a property of the data, not of this model. A deployed model would need vegetation composites that end before the prediction day.

What Didn’t Help

If far fire is often downwind, giving the model explicit wind and distance features should help it look in the right direction. This was pre-registered with a target of +20% AP at 0.75–3.75 km from today’s fire.

The model already derives most of this from the fire mask and wind channels; the features only make it learn faster. What is missing is information, not presentation: daily mean wind at 4.6 km cannot show the gust that throws embers a kilometre ahead.

Limitations and Next Steps

Most of the limits on this model are limits of the data, and each has a concrete fix.

Next Steps for Far Fire

The model is already good near the fire; the open problem is fire that appears far from it. The plan runs from cheap to large: first change what the model trains on, with the data it already has; then add the inputs that far fire depends on; then build a larger dataset, which also gives a clean new test season.

Alongside these, every comparison here is a single training run, so several seeds per arm would resolve differences below 0.03 AP, including JEPA-change against the random encoder.