Wildfire AI models need more transparency before they can earn trust

The role of machine learning and deep learning in wildfire prediction remains limited by geographic concentration, uneven data practices and a major shortage of open code, reveals a systematic review which claims that wildfire prediction research is advancing quickly, yet its ability to support global fire management is limited by weak transparency and uneven representation of high-burn regions.

Published in Fire, the study "Machine Learning and Deep Learning for Wildfire Prediction: A Systematic and Bibliometric Review of Methods, Data Practices, and Reproducibility (2020–2025)" reviewed 341 peer-reviewed studies from 2020 to 2025 and conducted a detailed methodological analysis of 110 articles from 2024, showing that deep learning and ensemble machine learning dominate the field while only 7.7% of studies provide publicly accessible code.

Wildfire AI research grows as climate risks intensify

Wildfires are becoming a more urgent global threat, driven by hotter temperatures, lower humidity, longer dry seasons, changing vegetation conditions and expanding human activity in fire-prone landscapes. Fires affect a large share of global forests and impose rising ecological, social and economic costs. In that setting, predictive tools are no longer only academic exercises, but are increasingly seen as part of early warning systems, prevention planning, risk mapping and operational decision support.

Traditional wildfire models have long relied on physical combustion processes, meteorological indices and empirical risk maps. These methods remain useful, but they often struggle to capture nonlinear interactions among climate, vegetation, terrain and human activity. Machine learning and deep learning offer a way to process large, complex and multi-source datasets, including weather data, vegetation indices, satellite imagery, topography and historical fire records. These methods can identify patterns that are difficult to capture through conventional modelling.

The review shows that research output in this area grew sharply between 2020 and 2025. The number of eligible studies increased steadily during the review period, reflecting the consolidation of wildfire prediction as a major computational research field within environmental science. The rise has been supported by better satellite data, larger environmental datasets, cloud computing, open geospatial tools and growing demand for fire-risk decision systems.

However, the growth has been uneven. China and the United States dominate publication output, followed by countries such as Australia, Brazil, Germany, France, the United Kingdom, India and South Korea. The authors warn that this concentration has consequences for global wildfire science. Many models are being developed in data-rich regions with temperate, Mediterranean or boreal fire regimes, while regions that experience large burned areas, including parts of Africa, South America and Siberia, remain underrepresented.

Models trained on one set of vegetation, climate and ignition patterns may not work well in another. A system built mainly on data from the western United States or China may have limited transferability to savannas, tropical forests or regions where fire is closely tied to land use, poverty, agricultural burning and weak fire-management infrastructure. The review frames this as both a scientific and equity problem. If predictive tools are built mainly in wealthy or data-rich regions, the countries most exposed to fire impacts may remain least represented in the models designed to help them. That could reinforce gaps in decision-support capacity just as climate change increases fire risk in vulnerable regions.

The authors also identify linguistic and database biases. Because the review relied on major English-language research databases, studies published in local languages or regional outlets may be underrepresented. Even with this limitation, the pattern is clear: wildfire prediction research is expanding, but its geography remains heavily skewed.

Deep learning advances, but data choices remain narrow

The review finds that machine learning and deep learning methods are being used for a wide range of wildfire prediction tasks, but most studies focus on occurrence or risk prediction. About three-quarters of the reviewed studies focused on predicting wildfire occurrence or risk. Smaller shares focused on burned area, severity, detection, monitoring, spread or propagation. This shows that current AI wildfire research is more focused on identifying where fires are likely than on modelling how fires move once they ignite.

The study classifies methods into major algorithm families. Tree-based and ensemble methods, including Random Forest and XGBoost, accounted for 26.7% of the detailed 2024 subset. Deep learning architectures together accounted for 59.4%. Within deep learning, convolutional neural networks were prominent, especially for spatial and image-based tasks. Hybrid and specialized architectures, feedforward networks, recurrent and temporal models, generative models, transformers and vision transformers also appeared across the literature.

The coexistence of traditional machine learning and deep learning is important. The field is not simply replacing older methods with more complex neural networks. Instead, different models are being matched to different data types and prediction tasks. Tree-based ensemble models remain attractive for tabular environmental data because they are efficient, robust and relatively interpretable. Deep learning is more common where researchers use satellite imagery, spatiotemporal sequences or complex remote sensing workflows.

The review also finds geographic variation in method choice. Studies linked to China more often used tree-based and ensemble approaches, especially Random Forest and XGBoost. Studies linked to the United States showed greater use of deep learning architectures, including convolutional and recurrent models. The authors suggest this reflects differences in data ecosystems, research infrastructure and operational priorities rather than a simple contest between superior and inferior methods.

Input data practices show a clear bias toward biophysical variables. Vegetation and fuel-related variables were the most common input domain in the detailed analysis, followed closely by historical fire labels such as fire perimeters, ignition records and occurrence data. Climatic and meteorological variables, remote sensing-derived inputs and topographical factors also featured in the literature. Human and socioeconomic variables were far less common.

The narrow use of socioeconomic data is a major limitation. Wildfires are not driven by climate and vegetation alone. Human activity influences ignition risk through roads, population density, agricultural burning, land management, development patterns and access to prevention or suppression resources. Socioeconomic conditions can also shape exposure, vulnerability and response capacity. If models prioritize hazard variables while ignoring social context, they may perform poorly as real-world risk tools.

The study argues that wildfire AI research remains largely hazard-centric rather than fully risk-integrated. This means many models predict environmental fire likelihood but do not fully capture the human systems that shape ignition, vulnerability and management outcomes. For decision makers, that gap is critical. Fire agencies need tools that can guide prevention, preparedness, resource allocation and community protection, not only produce technical risk scores.

Evaluation practices are also fragmented. Classification studies often used precision, recall and F1-score, while regression studies relied on error metrics such as RMSE and MAE. Other studies reported computational performance, spatial overlap measures, accuracy-based metrics or threshold-independent measures. The authors note that metrics are increasingly differentiated by task type, which is useful. But the diversity of metrics still makes it hard to compare results across studies.

The problem becomes sharper for rare but severe fires. Standard accuracy can be misleading when high-severity fire events are uncommon. A model can appear accurate while missing the most dangerous cases. The review calls for evaluation metrics that fit operational goals, particularly recall-sensitive metrics for early warning and spatial metrics for geospatial prediction. In wildfire management, the cost of a missed event can be far higher than the cost of a false alarm.

Reproducibility crisis limits trust in wildfire decision tools

One of the study's strongest warning concerns reproducibility. Among the 110 studies analyzed in detail, only 7.7% provided publicly accessible and verifiable code repositories. That means more than nine in ten studies did not provide the code needed for outside researchers or agencies to fully test, reproduce or adapt their models.

This is a structural constraint on cumulative knowledge building. In computational science, model descriptions are not enough. Reproducibility depends on access to code, preprocessing steps, feature engineering decisions, training and validation splits, hyperparameter settings, data documentation and assumptions about spatial and temporal sampling. Without those components, even a technically impressive wildfire model may be difficult to verify or reuse.

The review found a statistically significant association between algorithm family and code availability, suggesting that transparency practices vary by method type. This may reflect differences in complexity, institutional norms, intellectual property concerns or barriers to publishing deep learning pipelines. But the result is the same: the fastest-growing parts of the field are not consistently open enough to support validation.

Models may influence where agencies allocate resources, where communities receive warnings, where prevention efforts are prioritized and how governments plan fire response. A closed model that cannot be independently tested may undermine trust, especially if it is deployed in regions different from those where it was developed.

Reproducibility, the authors argue, is not only an academic virtue, but a requirement for operational reliability. If agencies cannot inspect or validate a model, they may hesitate to use it. If researchers cannot compare results across regions, the field risks producing isolated tools that work in one place but fail elsewhere. If models are developed without transparent data and code, errors may remain hidden until they affect real decisions.

The study calls for stronger reporting standards, open-code practices, dataset documentation and clearer links between evaluation metrics and operational goals. Journals and funders could require code deposition, while research communities could build open benchmarks that allow models to be tested across different ecosystems. Such standards would not require every region to use the same model, but they would make it easier to compare, validate and adapt systems.

The review also stresses that more complex models are not automatically better. For wildfire prediction, operational value depends on accuracy, interpretability, uncertainty estimates, transferability and usability by decision makers. A highly accurate deep learning model may be less useful if it cannot explain its predictions, quantify uncertainty or run reliably in an emergency setting. Future research, the authors argue, should focus not only on algorithmic novelty but on operational readiness.

Hence, the next challenge is broader than improving performance scores. Researchers must show whether AI models can be trusted, replicated and deployed in real wildfire management conditions. That includes testing across underrepresented regions, incorporating human and socioeconomic data, aligning metrics with real decisions and opening the computational pipelines behind published claims.