Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction

arXiv:2608.07472 · cs.LG · Submitted 2026-04-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction".

Jane: The paper was written by Nicolas Caron, Hassan Noura, Christophe Guyeux and Benjamin Aynes from Université Marie et Louis Pasteur and FEMTO-ST and SAD Marketing.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone. Today we're looking at a paper that's got a title that just rolls off the tongue — "Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction." And Jane, I gotta say, this one feels different.

Jane: Oh, absolutely, Tom. And I think the key word in that title is "segmentation." Most wildfire prediction work you see out there just throws a grid over a map and says, okay, we'll predict fire risk in each of these squares. This paper says, what if we let the data decide where the boundaries should be?

Tom: Right, and that's a wild idea when you think about it. We've been using grids for everything from weather to population studies, but fires don't care about grids. They follow terrain, wind, fuel, human activity. So why are we forcing them into these artificial boxes?

Jane: Exactly. And the authors — Nicolas Caron, Hassan Noura, Christophe Guyeux, and Benjamin Aynes — they're from Université Marie et Louis Pasteur in France, and they've got this really clever approach. They take historical fire locations and use image processing techniques to carve out what they call "fire zones." Think of it like drawing a map of where fires actually happen, rather than imposing a checkerboard on top.

Tom: And that's the part that got me excited. They're not just tweaking a model. They're questioning the fundamental way we set up the prediction problem. It's like if you were trying to predict where puddles form after rain, and someone said, hey, maybe we should look at the actual dips in the ground instead of just measuring in a grid.

Jane: That's a great analogy, Tom. And the implications are huge. If we can predict fires better, we can allocate resources better. Fire departments can position crews where they're actually needed. Evacuation warnings can be more targeted. This isn't just an academic exercise.

Tom: And they're showing real results, too. We'll get into the numbers later, but the short version is that their fire-zone approach beats the grid approach across multiple departments in France and multiple forecasting models. That's not a fluke.

Jane: Right, and the fact that it's unsupervised is a big deal. They don't need someone to hand-label fire zones. The algorithm finds them from the data itself. That means it can scale to new regions without expensive manual work.

Tom: So stick around, because we're going to break down exactly how they do this, what the results look like, and whether this could change how we think about wildfire prediction everywhere.

Jane: And maybe even how we think about spatial prediction in general. This could have implications beyond fires — floods, disease outbreaks, air quality. If the data can define the zones, why are we stuck with grids?

Tom: Good question. Let's dig into the paper and find out.

Summary: Jane: So, Tom, we're back with "Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction," and I want to get into what this paper actually does. Because the title is one thing, but the method is where it gets interesting.

Tom: Yeah, and the first thing that struck me is how they frame the problem. They're saying the way you slice up your study area matters more than which model you use. That's a bold claim, and they back it up with experiments across six French departments and six different forecasting models.

Jane: Right, and the models they tested are pretty standard stuff — logistic regression, XGBoost, CatBoost, GRU, LSTM, and a multilayer perceptron. So it's not like they're using some fancy new architecture that only works in a lab. These are workhorse models that people actually use.

Tom: And the results hold across all of them. That's what makes the claim so strong. It's not that their segmentation helps one particular model. It helps everything. The mean IoU improvement is somewhere between three and six percent, depending on the spatial scale they used.

Jane: IoU, for our listeners, is intersection over union — it's a way of measuring how well the predicted fire risk overlaps with what actually happened. Higher is better. And they're seeing consistent gains.

Tom: And the method itself is clever. They take historical fire locations, smooth them into a continuous risk surface, then use a watershed algorithm to find natural basins of fire activity. Then they merge those basins until they hit a target size that matches the prediction scale they want.

Jane: It's like finding the valleys in a mountain range. The watershed algorithm literally looks at the risk surface like a landscape, finds the high ridges, and the areas between them become zones. It's a really intuitive way to think about it.

Tom: And here's the kicker — the whole thing runs in under ten seconds per configuration. That's not a heavy computation. You could run this on a laptop.

Jane: And they're not asking you to change your model. You keep your XGBoost or your LSTM, you just feed it different data. That's a low-barrier improvement. Any team doing wildfire prediction could adopt this without retraining their whole pipeline.

Tom: Plus, they handle the messy real-world stuff. Fire data is sparse — most days, most places have no fires. They deal with that by undersampling the zero days, and they tune that proportion for each configuration.

Jane: And they're honest about limitations too. The optimal parameters differ by department. What works for Bouches-du-Rhône doesn't perfectly transfer to Hérault. They flag that as a real challenge for deployment.

Tom: But even with that caveat, the message is clear. Grids are a bottleneck. Letting the data define the zones is better. And that's a message that could change how a lot of operational systems are built.

Jane: So next, let's talk about the actual improvements they're proposing and how they got those numbers. Because the details matter here.

Improvements: Tom: Alright, Jane, so we've covered the big picture. Now let's get into the weeds of what they actually improved and how. And I think the best place to start is the core insight — that grid-based segmentation is actively hurting prediction.

Jane: Yeah, and they make a really specific argument about why. When you use a grid, you get cells that might be half water, half forest. You get cells where a fire happened once in ten years, and cells where fires happen every summer, all lumped together. That noise confuses the model.

Tom: So their fix is to build zones that follow the actual fire patterns. They do this in three stages. First, they take the raw fire occurrence data and turn it into a continuous risk signal using a Laplacian-based filter that smooths over a twenty-kilometer radius.

Jane: That smoothing matters because raw fire points are scattered and sparse. You need to fill in the gaps to see the underlying structure. Once they have that risk surface, they use K-means clustering to reduce noise and identify the main fire-prone areas.

Tom: And then comes the clever part — the merging step. Because the watershed algorithm gives you zones, but they might be too small or too big for your prediction scale. So they merge undersized zones with neighbors, dilate them to reach further if needed, and erode oversized ones until everything fits within a target size range.

Jane: And that target size is controlled by a scale parameter, which they test at zero point two, zero point three, and zero point four degrees. That's roughly twenty to forty kilometers. And they also have a tolerance parameter so zones don't have to be pixel-perfect.

Tom: Right, and the tolerance is set to zero point three, which gives them a range of acceptable sizes. The whole thing is a balancing act — you want zones big enough to have enough fire events for the model to learn from, but small enough to be locally meaningful.

Jane: And the results show that the medium scale — zero point three degrees — gives the most consistent gains across departments. That's a practical recommendation for anyone deploying this.

Tom: Now, one thing I really appreciate is that they don't just report the wins. They dig into the parameter coupling problem. The optimal number of dilations and intensity levels changes from department to department. Bouches-du-Rhône needs two dilations, Hérault needs three.

Jane: And they measured what happens if you ignore that — if you use one department's optimal settings on another, you lose about two to four percent IoU. That's not catastrophic, but it's not nothing either.

Tom: So they're essentially saying, you can't just set it and forget it. You need to tune per region. But even with that tuning cost, the fire-zone approach wins in every scenario they tested.

Jane: And that's the improvement in a nutshell. Better spatial units, better predictions, without changing the underlying models. It's an elegant result.

Tom: So now I want to zoom out a bit and look at the first page of the paper, because there's some context there that really sets the stage for why this matters.

First Page: Jane: So, Tom, let's step back and look at how the paper opens, because the introduction does a lot of work in setting up why this problem is worth solving.

Tom: Yeah, and the first thing they hit you with is the scale of the problem. Nearly half a million wildfires are reported globally every year, and over ninety percent are attributed to human activity. That's a staggering number.

Jane: And the economic burden is real too. Suppression costs, infrastructure damage, public health impacts — they routinely exceed billions of dollars per fire season. So this isn't just an environmental issue. It's an economic and humanitarian one.

Tom: And they position their work against the state of the art. A lot of recent work uses deep learning — U-Nets, Vision Transformers — to detect active fires from satellite imagery. But that's a different problem. Those models tell you where a fire is burning right now. This paper is about predicting where fires will start before they happen.

Jane: Right, and they're careful to say their approach is complementary to those detection models. You could use their segmentation as a preprocessing step, then feed those zones into a detection or classification pipeline. They're not competing — they're building the foundation.

Tom: And there's a really interesting point about why grids are so entrenched. It's not that anyone thinks grids are perfect. It's that they're easy. You just divide the map into equal squares and you're done. No thinking required.

Jane: But the paper argues that this convenience comes at a cost. Grids dilute the ignition signal. They introduce spatial noise from water bodies, from unreliable historical encoding, from areas where fires simply never happen. And that noise degrades model performance.

Tom: And they also point out a practical issue with fine-resolution grids. When you predict at kilometer scale, most cells have zero fires, which forces you to undersample and biases your validation metrics. Coarser grids help, but then you're back to the question of how to draw the boundaries.

Jane: And that's exactly the gap they're filling. They're saying, let's not guess. Let the historical data tell us where the boundaries should be. That's the core contribution.

Tom: And the authors are upfront about what they're not claiming. They're not saying this segmentation is optimal. They're not comparing it to every possible alternative. They're saying, here's a method that works, it's reproducible, and it beats the standard approach.

Jane: And that's refreshing. A lot of papers oversell. This one is measured and honest about its scope.

Tom: So we've covered the problem, the method, and the results. Let's wrap this up with our final thoughts on what this means for the future.

Conclusion: Tom: Alright, Jane, let's bring it home. We've spent this whole episode on "Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction," and I think the takeaway is pretty clear.

Jane: It really is, Tom. The paper shows that how you divide up your prediction area matters just as much as which model you use. And their fire-zone segmentation — built from historical fire data using watershed and clustering — beats the standard grid approach across every department and model they tested.

Tom: And the gains aren't tiny. Three to six percent mean IoU improvement is meaningful when you're talking about lives and property. And it costs almost nothing to implement — ten seconds per configuration, no labeled data needed.

Jane: The medium scale, zero point three degrees, is their recommended default. It gives the most consistent results across regions. And they're honest that you need to tune parameters per department, but even with that overhead, the method wins.

Tom: And I think the bigger implication is that this way of thinking could spread. If data-driven zones work for wildfires, why not for flood prediction? For disease spread? For air quality? Any spatial prediction problem where the underlying phenomenon doesn't respect grid lines.

Jane: That's a great point. The paper is specifically about fires, but the principle is general. Let the data define the units of analysis. That's a philosophy that could reshape a lot of fields.

Tom: And they've laid out a clear path forward. They mention transfer learning — training a model on data-rich regions and using it to segment data-scarce ones. That could make this work in places without long fire histories.

Jane: And they're open about the limitations. Cross-boundary zones, departmental coupling, scalability at finer resolutions. They're not pretending this is the final answer. They're saying it's a solid step in the right direction.

Tom: So we're saying goodbye to this paper, but I have a feeling we'll be seeing follow-up work. The idea is too good to ignore.

Jane: Agreed. And to our listeners, if you're working on wildfire prediction, or any spatial prediction problem, give this paper a read. It might change how you think about your data.

Tom: Thanks for joining us. We'll be back soon with the next paper. Until then, stay curious.

Jane: And stay safe out there. Bye everyone.

Nicolas Caron, Hassan Noura, Christophe Guyeux, Benjamin Aynes

Université Marie et Louis Pasteur · FEMTO-ST · SAD Marketing

cs.LG

Submitted: 2026-04-23

Updated: 2026-08-12

Comments: Accepted at 22nd AIAI 2026

Journal ref: Artificial Intelligence Applications and Innovations (AIAI 2026), IFIP AICT, vol. 795, Springer, Cham, 2027

DOI: 10.1007/978-3-032-30809-2_5

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 69/100

Key concepts

Fire-Zone Segmentation
This method uses historical fire locations and image processing techniques to carve out natural areas of fire activity. Instead of using arbitrary grids, the data itself defines the boundaries, allowing prediction models to learn from within these specific zones.
Mean IoU Improvement
Intersection over Union (IoU) measures how well predicted fire risk overlaps with actual fire locations. The paper demonstrates consistent gains in this metric—between 3 and 6 percent—when using the new segmentation method compared to standard grid-based approaches.

Terminology

Summary

Summary

This paper challenges the prevailing paradigm in wildfire prediction that discretizes study areas into uniform grids, arguing instead that how data is discretized matters more than which model is used. The authors propose an unsupervised fire-zone segmentation algorithm that combines watershed detection with K-means clustering to define prediction units directly from historical fire patterns, rather than relying on arbitrary geometric grids.

The problem formulation states that the work "addresses the detection and spatial segmentation of fire-prone zones as a preprocessing step for wildfire prediction. Specifically, we investigate whether partitioning the study area into semantically meaningful fire zones—rather than relying on arbitrary uniform grids—can improve the accuracy of short-term wildfire forecasts generated by AI models." The authors note that contemporary work typically relies on uniform grids, citing Michail et al. who coupled GraphCast with a temporal encoder on a 0.25° grid, Chen et al. who partitioned Portugal into ten cells, and Koh et al. who analyzed U.S. risk on a 0.5° grid. They also distinguish their work from supervised deep-learning segmentation approaches (U-Net, Vision Transformers) which address a different problem: detecting or segmenting fires after ignition, rather than defining prediction zones before forecasting. Their approach is unsupervised, deriving prediction units directly from historical ignition distributions without requiring segmentation labels.

The proposed method comprises three main stages: (A) generating a continuous 3D signal of fire risk, (B) detecting fire-prone areas, and (C) merging fire areas to match a specified target size. The algorithm takes four input parameters: Scale (s) determining zone size in degrees, intensity levels (r) controlling intensity thresholds for fire regions, max dilations (a) specifying maximum dilation iterations to merge undersized zones, and tol (t) defining tolerance on region size. The continuous risk signal is generated by transforming the sparse discrete fire occurrence raster into a continuous spatial risk surface using a Laplacian distribution-based filter applied to a 3D raster of fire occurrences, which smooths the signal based on the average duration of fire sequences within a 20 km radius. Fire-prone areas are then detected by clustering cumulative risk into groups using K-means. The merging step ensures each detected fire-prone region matches a target size corresponding to the spatial scale at which predictions will be made, with zones smaller than minimum size merged with neighboring zones, undersized zones removed after max dilations iterations, and oversized zones eroded.

The authors justify their algorithmic choices, stating that watershed naturally detects 'basins' in the density surface generated from historical ignitions, producing zones that follow the topography of fire risk rather than imposing arbitrary geometric shapes, while K-means provides explicit control over the number of merged regions via the intensity levels parameter. They rejected SLIC superpixels, Mean-Shift, and DBSCAN for various reasons including poor adaptation to sparse irregular point data, sensitivity to bandwidth selection, and difficulty with varying density.

The method was validated on two datasets organized by French departments. The Firefighter dataset compiles wildfires requiring intervention across four departments: Ain (01), Doubs (25), Rhône (69), and Yvelines (78), spanning 2017 to 2024. The BDIFF dataset (Forest Fire database) covers Bouches-du-Rhône and Hérault departments, spanning 2017 to 2023. Wildfire forecasting is treated as an ordinal multi-class task with five-level occurrence labels: class 0 for no fires, and four ordered categories (Normal, Medium, High, Extreme) created by K-means grouping of positive instances. Features span six categories: Meteorological, Topographic, Socio-Economic, Air Quality, Hydrological, and Historical, transformed into 3D rasters at 2 km resolution. Datasets were split into training (2017–2021), validation (2022), and testing (2023) subsets.

Six forecasting models were evaluated: Logistic Regression, XGBoost, CatBoost, GRU, LSTM, and MLP. The authors selected fast-to-train models, noting that gradient-boosted decision trees (XGBoost, CatBoost) remain highly competitive with—and often outperform—deep learning models on tabular data. A grid-search over max dilations ∈ 1, 2, 3, 4, 5 and intensity levels ∈ 2, 3, 4, 5, 6 was used, with tol fixed at 0.3. Three spatial scales were analyzed: 0.2°, 0.3°, and 0.4°. Class imbalance was addressed by testing different proportions of zero-class samples, with 0.3–0.5 proving optimal. Performance was evaluated using the IoU metric, which accounts for class uncertainty and preserves class ordinality—predicting class 1 instead of 4 is penalized less than predicting 0.

Results demonstrate consistent improvements with fire-area segmentation. For the Firefighter dataset, Doubs shows the strongest uplift (∆= + 0.05 across all scales); Ain benefits only at the mid- and high-thresholds, while Yvelines gains mainly at mid-scale. For the BDIFF dataset, Bouches-du-Rhône gains most at the lower scale (∆= + 0.06 at s = 0.2) while Herault peaks at the mid-scale (∆= + 0.06 at s = 0.3). Across all configurations, In no scenario does grid-based segmentation surpass fire-area segmentation. The authors recommend the medium scale (s = 0.3◦) as yielding the most robust cross-department performance, noting that at finer scales gains remain positive but metrics are more volatile, while at coarser scales the performance gap narrows. A reproducibility test using logistic regression revealed result variability of approximately 0.01 for both approaches, confirming that observed differences reflect genuine methodological improvements.

The discussion identifies several limitations. Parameter coupling across departments is noted: The current parameter optimization assumes that the algorithm's optimal parameters are identical for every department, which is incorrect—for example, Bouches-du-Rhône requires 2 dilations, whereas Hérault requires 3. When using a single configuration across all departments, mean IoU dropped by approximately 0.02–0.04 for Hérault and Ain. Cross-boundary zones are not handled due to departmental organization of French fire services, data volume considerations, and lack of nationwide fire-brigade intervention records. For data-scarce regions, the authors propose two strategies: training a spatial model to learn segmentation from data-rich regions and infer for unseen regions, or constructing an upstream fire-susceptibility map as input. Scalability is addressed, noting the algorithm runs in under 10 seconds per configuration and the full grid-search over 25 configurations completes in under 5 minutes per department, with memory footprint below 2 GB.

The conclusion states: "we demonstrate that conventional grid-based segmentation is poorly suited to wildfire prediction because it introduces spatial noise that weakens the model's ability to learn ignition patterns. To address this limitation, we propose a fire-zone segmentation method that defines prediction units directly from historical fire distributions. Evaluated across six French departments, six forecasting models, three spatial scales, and multiple lead times, this data-driven segmentation consistently outperforms the static grid approach with mean IoU improvements of +3–6%. The method requires no labeled segmentation data, runs in under 10 seconds per configuration, and integrates seamlessly into existing prediction pipelines. Future work may explore exhaustive cross-configuration testing, transfer learning from data-rich to data-scarce regions, or coupling the segmentation with downstream deep-learning models for end-to-end optimization."

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and what the improved system can do:

  1. Replace uniform grid discretization with data-driven fire-zone segmentation

  • Implement the watershed + K-means clustering pipeline as a preprocessing layer before any prediction model

  • Use the three-stage algorithm: continuous risk signal generation → fire-prone area detection → zone merging with size constraints

  1. Add spatial-adaptive input encoding

  • Modify the input layer to accept variable-sized, irregularly shaped prediction zones instead of fixed rectangular cells

  • Encode each zone's historical ignition density as part of the feature vector

  1. Implement scale-adaptive zone merging

  • Add the max dilations, intensity levels, and tol parameters as tunable hyperparameters in the model configuration

  • Optimize these per department/region rather than using global defaults

  1. Integrate unsupervised segmentation with supervised forecasting

  • Add a two-stage training pipeline: first learn the segmentation from historical fire data (unsupervised), then train the forecasting model on those zones (supervised)

  • Use the segmentation as a fixed preprocessing step, not jointly optimized, to maintain stability

  1. Add cross-boundary handling for multi-region deployment

  • Extend the algorithm to merge zones across administrative boundaries when fire patterns cross them

  • Implement tiled processing for large-scale deployment to manage memory constraints

  1. Achieve +3–6% mean IoU improvement in short-term wildfire forecasting compared to grid-based systems, consistently across six departments and six model architectures

  2. Adapt to regional fire patterns automatically — the system learns where fires actually occur rather than assuming uniform spatial distribution, so it works in both dense urban-interface areas (Ain, Rhône) and rural forest regions (Bouches-du-Rhône, Hérault)

  3. Generate prediction-ready zones in under 10 seconds per configuration, fully parallelizable across departments, enabling near-real-time updates when new fire data arrives

  4. Handle data-scarce regions by using transfer learning: train the segmentation on data-rich departments, then infer zones for regions with limited historical fire records

  5. Provide interpretable prediction units — each zone corresponds to a semantically meaningful fire-prone area, so forecasts can be directly mapped to operational response units (fire stations, evacuation zones) rather than arbitrary grid cells

  6. Maintain performance across multiple forecast horizons (0-day, 1-day, 3-day) with consistent gains, making it suitable for both daily operational alerts and short-term planning

  7. Reduce training noise by eliminating water bodies, non-fire pixels, and unreliable historical encoding from the prediction units, improving signal-to-noise ratio for all downstream models (LR, XGBoost, CatBoost, GRU, LSTM, MLP)

Related papers