🤖 machine learning

Set Prediction for Next-Day Active Fire Forecasting

This paper introduces WISP, a query-based set prediction model that reformulates next-day active fire forecasting as a point-set prediction task to generate high-resolution, ranked fire cluster centers globally, achieving superior localization and coverage performance compared to traditional grid-based approaches.

Original authors: Yuchen Bai, Georgios Athanasiou, Xin Yu, Diogenis Antonopoulos, Ioannis Papoutsis, Stijn Hantson, Nuno Carvalhais

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuchen Bai, Georgios Athanasiou, Xin Yu, Diogenis Antonopoulos, Ioannis Papoutsis, Stijn Hantson, Nuno Carvalhais

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict exactly where a wildfire will start tomorrow.

Most current computer models try to do this by drawing a giant grid over a map and coloring every single square. They say, "There is a 10% chance of fire here, a 5% chance there, and a 2% chance over there." While this is good for knowing if a whole region is "at risk," it's like trying to find a specific needle in a haystack by just saying, "The needle is probably somewhere in this pile of hay." It doesn't tell you exactly where the needle is.

This paper introduces a new model called WISP (Wildfire Ignition Set Predictor) that changes the game. Instead of coloring the whole map, WISP acts like a spotlight or a hunting dog. It doesn't try to guess the probability for every square inch. Instead, it looks at the data and says, "I am 90% sure a fire will start right here, and maybe another one over there." It outputs a specific, short list of predicted locations.

Here is how the paper explains this, broken down into simple concepts:

1. The Problem: Too Much Noise

Wildfires are rare events. If you look at a map of the world, 99.9% of the land is not on fire.

Old Way: The old models tried to predict fire for every single pixel on the map. Because fire is so rare, the computer got confused by all the "no fire" data. It was like trying to learn to recognize a cat by looking at a million pictures of empty rooms.
The Paper's Solution: WISP treats fire prediction like finding a few specific dots on a page. It ignores the empty space and focuses only on finding the "dots" (the fire centers).

2. The Ingredients: A 48-Hour History

To make its predictions, WISP looks at a "time capsule" of data from the last two days (48 hours) and the weather forecast for the next day. It mixes together:

Weather: Temperature, wind, humidity.
Vegetation: How dry the plants are (like checking if the grass is dead or green).
The Land: Is it a mountain? A flat plain?
Fire History: Has there been a fire here recently? (Fire often starts near where it burned before).

3. The Magic Trick: "Hungarian Matching"

This is the technical heart of the paper, but think of it as a perfect seating arrangement.

The model has a fixed number of "guesses" (let's say 10 or 50 guesses) it can make.
The "Truth" (what actually happens) is a set of real fire locations.
The model tries to match its guesses to the real fires. It uses a math trick called Hungarian Matching to pair them up one-to-one.
The Analogy: Imagine you have 10 students (the model's guesses) and 3 real fires (the truth). The model tries to assign the 3 best students to the 3 fires. The other 7 students are told, "You didn't find a fire, so you are 'No Fire'."
The paper notes a tricky part: The model has to decide which guesses are the "winners" and which are "losers" while it is learning. They had to invent a special rule to make sure the model didn't get confused about whether a guess was for "finding a fire" or "saying there is no fire."

4. The Results: How Good is It?

The authors tested WISP on a global dataset covering 11 years. They compared their "spotlight" model against the old "grid" models.

Precision: WISP successfully identified the center of fire clusters about 38% of the time (which is a big deal for such a difficult, rare event).
Coverage: It managed to "cover" about 53% of the total fire energy (how much fire is actually burning) with its predictions.
Accuracy: About 54% of the time, the model's predicted spot was within 5 kilometers of the actual fire center.

5. The Trade-Off: More Guesses vs. Fewer Mistakes

The paper found that you can tune the model like a radio dial:

Fewer Guesses (Conservative): If you tell the model to only make 10 guesses, it is very careful. It rarely cries "Wolf!" when there is no wolf, but it might miss some fires.
More Guesses (Aggressive): If you tell the model to make 50 guesses, it catches way more fires (better coverage), but it also makes more false alarms (predicting fire where there is none).

Summary

The paper claims that by changing how we ask the computer to solve the problem—from "guess the probability of every square" to "give me a list of specific fire locations"—we can predict wildfires much more accurately. WISP is the first model to successfully use this "list of dots" approach for high-resolution fire forecasting, proving that sometimes, looking for specific needles is better than staring at the whole haystack.

Technical Summary: Wildfire Ignition Set Predictor (WISP)

Problem Formulation
Current global wildfire forecasting models typically operate on coarse, kilometre-scale grids (1–9 km) to predict fire probability or danger scores. While useful for regional risk assessment, these dense probability fields fail to represent the discrete, localized nature of actual fire ignitions. Furthermore, treating next-day active fire forecasting as a dense semantic segmentation problem at high resolution (e.g., 375 m) creates severe foreground-background imbalance, as the prediction domain is dominated by non-fire pixels while the learning signal is concentrated in sparse fire events.

The authors propose reformulating next-day active fire forecasting as a sparse set prediction problem. The goal is to predict a variable-length set of future fire-cluster centres over a 24-hour horizon, given historical Earth-system states and future-known covariates, rather than estimating a probability for every grid cell.

Methodology: The WISP Architecture
The proposed model, Wildfire Ignition Set Predictor (WISP), adapts the DETR (Detection Transformer) paradigm for wildfire forecasting.

Input and Target Representation:
- Inputs: The model ingests 48 hours of multi-source covariates ( $X_h$ ) and 24 hours of future-known drivers ( $X_f$ ). These include meteorology (ERA5), satellite vegetation products (PROBA-V, VIIRS), static land descriptors (elevation, hydrology, landforms), and historical fire observations (VIIRS active fire and Fire Radiative Power - FRP). All data are harmonized to a 375 m grid.
- Targets: Future active fire detections are aggregated into connected components (clusters) using a 7×7 footprint (~1 km). The target is a set of FRP-weighted cluster centres ( $G_K$ ).
- Output: WISP predicts a fixed-size set of $Q$ query hypotheses (default $Q=10$ ), where each query consists of a fire probability score and a normalized 2D location. Unmatched queries are trained as "no-fire" samples.
Architecture:
- Encoders: A spatio-temporal backbone encodes the historical context ( $X_h$ ) into a multi-scale feature pyramid. Parallel encoders process static land data, historical fire sequences, and fire-summary maps.
- Condition Injection: Future drivers ( $X_f$ ) are encoded and injected into the historical context via a gated multi-scale condition injection mechanism. This allows future weather and vegetation states to modulate historical features in a scale-aware manner, with stronger conditioning applied to lower-resolution semantic features.
- Query Decoder: A DETR-style decoder uses learnable queries to attend to the future-conditioned memory. It iteratively refines reference points (locations) and classification logits through self-attention and cross-attention mechanisms.
Training Objective:
- The model is trained end-to-end using Hungarian matching to establish a one-to-one assignment between predicted queries and ground-truth clusters.
- Loss Function: The total loss combines classification loss (for all queries) and localization loss (for matched queries).
- Key Innovation: The authors identify a conflict where a single classification score is used for assignment, ranking, and query activation. To mitigate this, they employ asymmetric classification–localization weighting. Specifically, they down-weight the "no-fire" class in the matching cost and loss to prevent unmatched queries from dominating the training signal, and adjust the balance between classification and localization weights in the matching cost to control query compactness.

Key Contributions

Formulation Shift: The paper reformulates next-day active fire forecasting from dense grid estimation to point-set prediction, introducing WISP as the first query-based model for this specific task.
Score-Role Conflict Resolution: The authors identify and address a specific challenge in sparse set prediction where the classification score serves multiple conflicting roles (assignment, ranking, activation). They propose asymmetric weighting in matching and loss to stabilize training.
Benchmark Construction: The authors constructed a globally distributed, hourly, multi-source benchmark using VIIRS-derived targets, covering 11 years (2014–2024) across 14 GFED regions.

Results
Evaluated on a held-out test set spanning global fire regions, the best WISP variant ( $Q=50$ ) achieved:

38.2% Average Precision (AP) for ranked fire-centre detections at a 14-pixel (~5 km) tolerance.
53.4% coverage of fire cluster mass, weighted by Fire Radiative Power (FRP).
54.1% of observed clusters localized within 5 km.

Regime-Specific Performance:

Continued Fire: The model performs significantly better when historical fire evidence is present (AP@14 of 43.8% for $Q=50$ ), confirming historical fire as a strong predictive cue.
New Ignition: Performance is lower for new ignitions without historical cues (AP@14 of 12.5% for $Q=50$ ), highlighting the difficulty of predicting spontaneous ignitions.
Query Budget Trade-off: Increasing the query budget ( $Q$ ) improves coverage and recall but increases the number of false positives and duplicate detections. A smaller budget ( $Q=10$ ) provides more compact predictions with fewer false alarms.

Significance and Claims
The paper claims that sparse set prediction is a viable formulation for high-resolution wildfire forecasting, offering a direct representation of discrete fire events rather than smoothed probability fields. The work establishes a new benchmark for this regime and demonstrates that query-based models can effectively localize future fire clusters. The authors note that while the model shows promise, future work is needed to improve assignment stability, score calibration, and the integration of ignition–spread coupling. The results are presented as establishing a foundation for future research in this specific forecasting regime.