🤖 machine learning

Uncertainty-Aware Deep Learning for Wildfire Danger Forecasting

This paper presents an uncertainty-aware deep learning framework that jointly models epistemic and aleatoric uncertainties to significantly improve the accuracy, calibration, and reliability of short-term wildfire danger forecasting while providing actionable insights for decision support.

Original authors: Spyros Kondylatos, Nikolas Papadopoulos, Gustau Camps-Valls, Ioannis Papoutsis

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Spyros Kondylatos, Nikolas Papadopoulos, Gustau Camps-Valls, Ioannis Papoutsis

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 trying to predict when a dry forest might catch fire. It's a bit like trying to guess if a pile of kindling will suddenly ignite. Sometimes, the weather is perfect for a fire; other times, even if it looks risky, nothing happens. And sometimes, a tiny spark in the wrong place causes a massive blaze.

This paper is about building a "smart weather forecaster" for wildfires using Artificial Intelligence (AI). But the authors realized that standard AI has a big flaw: it's often overconfident. It might say, "I'm 99% sure this will burn," when it's actually just guessing. If a fire chief acts on that wrong guess, they might waste resources or, worse, miss a real danger.

To fix this, the researchers built an AI that doesn't just give an answer; it also tells you how sure it is about that answer. They call this "Uncertainty-Aware Deep Learning."

Here is how they did it, using simple analogies:

1. The Two Types of "Not Knowing"

The paper explains that there are two different reasons an AI might be unsure. They call these Aleatoric and Epistemic uncertainty.

Aleatoric Uncertainty (The "Chaos" Factor):
Think of this as the inherent randomness of nature. Imagine you have a perfect map of a forest, the exact temperature, and the humidity. Yet, a fire might still start or not start just because of a random gust of wind or a stray spark. This is noise in the data that cannot be fixed, no matter how much you study. It's like trying to predict the exact path of a single raindrop; the system is just too chaotic.
- The Paper's Finding: As you try to predict further into the future (like 10 days ahead), this "chaos" gets louder. The weather becomes harder to predict, so the AI's "noise" level goes up.
Epistemic Uncertainty (The "Knowledge Gap"):
This is uncertainty because the AI hasn't learned enough yet. Imagine a student taking a test. If they get a question wrong because they never studied that topic, that's a knowledge gap. If they get it wrong because the question was tricky or the answer key was blurry, that's the "chaos" factor.
- The Paper's Finding: This type of uncertainty comes from the model itself. The researchers found that even when looking 10 days ahead, the AI's "knowledge gap" stayed about the same. The model didn't get "stupider" over time; it just faced more chaotic data.

2. The "Super-Student" Approach

To handle these two types of uncertainty, the researchers didn't just build one AI model. They built a system that acts like a panel of experts (specifically using a method called "Bayes by Backpropagation").

The Analogy: Imagine you ask one student to predict a fire. They might be confident but wrong. Instead, the researchers asked 100 slightly different versions of the student to look at the same data.
- If all 100 students agree, the AI is certain.
- If the students are arguing with each other, the AI knows it is uncertain (Epistemic).
- If the students agree but the data itself is messy (like a blurry photo of a fire), the AI knows the data is noisy (Aleatoric).

By combining these two, the AI gives a final prediction that is not just a number, but a number with a "confidence rating."

3. What They Discovered

The researchers tested this system to predict wildfires in the Mediterranean (places like Greece, Spain, and Italy) for the next day, and even up to 10 days in advance.

Better Accuracy: The "uncertainty-aware" AI was actually better at predicting fires than the standard, overconfident AI. It improved its accuracy score by about 2.3%.
Better Honesty: The standard AI was often overconfident (saying "100% sure" when it was wrong). The new AI was much more honest. It calibrated its confidence so that when it said "80% sure," it was actually right 80% of the time.
The "Reject" Button: Because the AI knows when it's unsure, you can set a rule: "If the AI is less than 90% sure, throw that prediction away." The paper showed that if you do this, the remaining predictions are incredibly accurate (jumping to 93.6% accuracy). This helps decision-makers ignore the "guesses" and focus only on the "facts."
Complementary Insights: In easy cases (like a clear, dry day with no fire), both types of uncertainty are low, so they look the same. But in tricky, dangerous situations, they tell different stories. One might say, "The data is messy" (Aleatoric), while the other says, "I haven't seen a situation like this before" (Epistemic). Knowing the difference helps humans understand why the prediction is hard.

4. Real-World Use: The Fire Danger Map

The paper shows how this looks on a map.

Scenario A (The Clear Danger): The map shows a red zone (high fire danger) with a green "low uncertainty" tag. This means the AI is very sure a fire will happen here. Decision-makers can send firefighters immediately.
Scenario B (The Foggy Danger): The map shows a yellow zone (medium danger) but with a red "high uncertainty" tag. This means the AI sees some risk, but it's not sure. It's like a foggy day where you can't see clearly. The advice here is to "be extra cautious" or wait for more data, rather than panicking or ignoring it.

Summary

The paper argues that for high-stakes tasks like predicting wildfires, an AI that says "I don't know" is more valuable than an AI that confidently guesses wrong. By teaching the AI to distinguish between "the world is chaotic" and "I don't know enough," they created a tool that is not only more accurate but also more trustworthy for the people who need to save lives and protect forests.

Technical Summary: Uncertainty-Aware Deep Learning for Wildfire Danger Forecasting

1. Problem Statement

Wildfires pose severe threats to human life, infrastructure, and ecosystems, with their frequency and intensity amplified by climate change. While Deep Learning (DL) has demonstrated promise in predicting wildfire danger, its adoption in disaster management is hindered by concerns regarding the reliability, transparency, and calibration of model outputs. Standard DL models often produce overconfident predictions and lack mechanisms to quantify uncertainty.

Wildfire forecasting is inherently uncertain due to:

Stochasticity: Fire ignition is a stochastic process where similar environmental conditions may or may not lead to ignition.
Data Scarcity: Fire events are rare (e.g., a fire-to-non-fire ratio of $5 \times 10^{-7}$ in the study dataset), leading to class imbalance and distributional shifts between training and real-world scenarios.
Complex Drivers: The task involves modeling complex, non-stationary interactions between meteorological, vegetation, topographical, and anthropogenic variables.

The core problem addressed is the lack of uncertainty quantification in existing DL frameworks for wildfire danger, which limits their trustworthiness for operational decision-making.

2. Methodology

The authors propose an uncertainty-aware DL framework that jointly models two distinct types of uncertainty:

Epistemic (Model) Uncertainty: Arising from the model's lack of knowledge, reducible with more data or better models.
Aleatoric (Data) Uncertainty: Arising from inherent noise in the data (e.g., label noise, stochastic ignition), irreducible even with more data.

Data and Task Formulation

Dataset: The study utilizes Mesogeos, a 1km $\times$ 1km data cube covering the Mediterranean basin (2006–2022).
Inputs: A 55-day time series of dynamic variables (vegetation indices, land surface temperature, meteorological data, etc.) and static features (elevation, land cover).
Task: Binary classification of wildfire danger for a target day $t$ . Positive samples are ignition points; negative samples are drawn from locations >62km away from ignitions.
Forecast Horizons: The study evaluates next-day forecasting ( $n=1$ ) and extends analysis up to 10 days ahead ( $n=10$ ).

Model Architecture and Uncertainty Estimation

The framework employs Long Short-Term Memory (LSTM) networks as the base architecture. Uncertainty is estimated using three primary approaches integrated into a unified framework:

Epistemic Uncertainty Estimation:
- Bayesian Neural Networks (BNNs): Specifically Bayes by Backpropagation (BBB), which places probability distributions over weights using Variational Inference (VI).
- Deep Ensembles (DEs): Training multiple deterministic networks with different random initializations.
- MC Dropout: Interpreting dropout as approximate Bayesian inference during inference.
Aleatoric Uncertainty Estimation:
- A heteroscedastic uncertainty model (Collier et al., 2020) is applied. This places a Gaussian distribution over the logits of the softmax classifier, allowing the model to predict input-dependent noise ( $\sigma_c(x)$ ) alongside the mean logits ( $f_c(x)$ ).
Total Uncertainty (Joint Modeling):
- The framework integrates the aleatoric module into the Bayesian formulation.
- Double Monte Carlo (MC) Sampling: To approximate the predictive distribution, the model samples $N$ weight sets (for epistemic) and, for each weight set, samples $S$ noise vectors (for aleatoric).
- Decomposition: Total uncertainty is decomposed into Epistemic Uncertainty (variance of mean predictions across weight samples) and Aleatoric Uncertainty (average variance across noise samples within each weight set).

3. Key Contributions

Unified Framework: Development of a DL framework that jointly captures epistemic and aleatoric uncertainty for short-term wildfire danger forecasting, demonstrating improvements in predictive skill and calibration over deterministic baselines.
Temporal Analysis: Investigation of uncertainty evolution across 1-to-10 day forecast horizons, revealing that aleatoric uncertainty increases with time (reflecting environmental stochasticity), while epistemic uncertainty remains stable.
Disentanglement Analysis: Demonstration that while epistemic and aleatoric uncertainties are highly correlated in low-uncertainty (easy) cases, they provide complementary insights in high-uncertainty (challenging) cases, particularly for identifying ambiguous scenarios.
Operational Utility: Practical demonstration of how uncertainty estimates can support decision-making through:
- Generating danger maps with accompanying uncertainty layers.
- Establishing uncertainty thresholds to reject low-confidence predictions (Discard Test).
- Improving model calibration to mitigate overconfidence.

4. Results

Predictive Performance and Calibration

Metrics: The best-performing model (BBB with aleatoric uncertainty) improved the F1 Score by 2.3% and reduced the Expected Calibration Error (ECE) by 2.1% compared to a deterministic baseline.
Recall vs. Precision: The BBB model with aleatoric uncertainty increased Recall by 7.9% (better detection of fire events) at the cost of slightly lower Precision, a trade-off deemed acceptable for high-stakes applications where missing a fire is costlier than a false alarm.
Calibration: Uncertainty-aware models, particularly BBB, showed significantly better calibration than deterministic models, with reliability diagrams closely aligning with the diagonal. This was achieved without post-hoc calibration techniques.

Reliability of Uncertainty Estimates

Discard Test: As the most uncertain predictions were discarded, the average loss decreased for models capturing epistemic uncertainty. The BBB model showed a 7.5% F1 score improvement after discarding the top 20% most uncertain samples.
Uncertainty Density: Models capturing epistemic uncertainty (BBB, DEs) clearly separated correctly classified samples (low uncertainty) from misclassified ones (high uncertainty). In contrast, aleatoric-only models and MC Dropout failed to distinguish errors in the positive class effectively.
Correlation: While aleatoric and epistemic uncertainties were highly correlated in the full dataset (Spearman $\rho \approx 0.91$ for BBB), this correlation dropped significantly (becoming negative for some models) when analyzing only the most uncertain samples (top 25%), confirming their complementary nature in difficult cases.

Temporal Evolution (1–10 Days)

Aleatoric Uncertainty: Increased systematically with longer forecast horizons, reflecting the accumulation of stochasticity in environmental inputs (e.g., weather variability).
Epistemic Uncertainty: Remained stable across all horizons, suggesting the models were sufficiently trained to generalize within their respective temporal windows.
Performance: AUPRC declined only modestly (8%) from 1-day to 10-day forecasts, indicating retained predictive skill.

5. Significance and Claims

The paper claims that integrating uncertainty-aware DL into wildfire forecasting significantly enhances the accuracy, reliability, and trustworthiness of predictive systems.

Decision Support: The study argues that uncertainty estimates are not merely theoretical but practical tools. They allow operators to:
- Identify regions where the model is confident (high danger, low uncertainty) for immediate intervention.
- Flag regions where the model is uncertain (high danger, high uncertainty) for further investigation or expert review.
- Reject low-confidence predictions to prevent misguided resource allocation.
Complementarity: The work underscores that modeling both uncertainty types is crucial. While they may be redundant in "easy" cases, their joint modeling provides essential insights in complex, high-risk scenarios where data noise and model limitations coexist.
Broader Applicability: The authors suggest that the challenges addressed (data sparsity, distributional shifts, need for calibration) are common across natural hazard domains, implying the framework's potential utility beyond wildfires.

The study concludes that uncertainty-aware DL, particularly using Bayesian methods like BBB combined with heteroscedastic noise modeling, represents a significant step toward developing trustworthy AI systems for high-impact environmental applications.