Sub-hourly forecasting of fire potential using machine learning on time series of surface weather variables

The role of advanced and adaptive fire forecasting systems

Fires play a significant role in terrestrial ecosystems, influencing the carbon cycle and ecological progression through disturbance and regeneration (Bond and Keeley 2005; Scott and Glasspool 2006; Bowman et al. 2009). However, fires can have adverse effects on humans, particularly where wildland and urban environments intersect, leading to significant impacts on infrastructure, ecosystem services, property and human lives (Pettinari and Chuvieco 2020; Burke et al. 2021). The increasing frequency and intensity of fires highlight the need for improved early warning systems. With projected climate shifts (Klausmeyer and Shaw 2009), these systems become crucial for effective fire adaptation and mitigation.

Forecasting the outbreak and spread of fires is inherently complex owing to the interactions of various drivers across diverse spatial and temporal scales, including climate, weather, topography, vegetation and human activity (Archibald et al. 2009). Traditional indices like the Fire Weather Index (FWI; Van Wagner 1974), Fire Behaviour Index (FBI; Hollis et al. 2024) and the Fire Forest Danger Index (FFDI; Griffiths 1999) predict the likelihood and potential danger of fires over different time scales, from daily forecasts to seasonal outlooks. Accurate short-term predictions are essential for immediate resource allocation and emergency response (Abt et al. 2009).

However, many existing systems are not designed for rapidly changing environmental conditions (Sharples et al. 2009; Fox-Hughes 2015). Their primary limitations include unreliability and lack of validation across different contexts. These indices often struggle with diverse geographic conditions, are sensitive to meteorological inputs and time scales (Brody-Heine et al. 2023), and fail to handle rapidly changing scenarios, leading to inaccurate forecasts (Stocks et al. 1989; Gong et al. 2023). Additionally, the static nature of these indices does not accommodate the dynamic changes in fire drivers, which can vary significantly over short periods.

These limitations underscore the need for time-scale-agnostic models that provide continuous monitoring rather than one-off forecasts. Such systems could lead to more adaptive and reliable fire danger forecasting by continuously integrating new data and adjusting to changing conditions, such as shifts in weather patterns, vegetation growth and human activity (Canadell et al. 2021).

Integrating comprehensive fire drivers into modern fire danger assessments

Fire danger assessments now incorporate vegetation, weather, topography and ignition sources (Beverly et al. 2010; Müller et al. 2020; Kaur 2023; Wang et al. 2023), leading to diverse research on fire susceptibility using expertise-based assessments, statistical methods and machine learning (ML) models (Hong et al. 2018; Leuenberger et al. 2018; Zheng et al. 2020). ML models handle large datasets and uncover complex patterns but require extensive data and can lack transparency. Empirical models are easier to interpret but may not capture physical processes accurately. Physics-based models provide detailed simulations but are computationally intensive and require precise input data. Combining these approaches could enhance model accuracy and informativeness.

Physics-based models provide valuable insights into physical phenomena, whereas data-driven ML models are particularly strong in predictive tasks (Jung et al. 2013; Leuenberger et al. 2018; Zhang et al. 2019, 2021; Huot et al. 2020; Bergado et al. 2021; Bjånes et al. 2021; Le et al. 2021; Kondylatos et al. 2022). With their capacity to process complex datasets and model intricate variable relationships, ML models are well suited for producing fine-scale predictions and adapting to new data in real time. Although full-physics models offer detailed representations, they can be computationally intensive, which can limit real-time processing capabilities in certain contexts. In contrast, simpler surrogate or ML models can produce faster inferences, capturing nuanced, dynamic patterns and enhancing forecast reliability through continuous learning and adaptation.

ML models for short-term fire danger provide monthly, weekly (Bjånes et al. 2021; Le et al. 2021; Zhang et al. 2021) and daily (Huot et al. 2020; Bergado et al. 2021; Kondylatos et al. 2022) forecasts. Fires are influenced by complex, rapidly changing weather patterns. Forecasting models that assimilate continuous data and provide frequent updates have an advantage. Existing ML techniques often rely on static data, so there is a need for models that adapt swiftly to changing conditions to reduce prediction uncertainty, especially in regions with variable weather patterns (e.g. Australian bushfire-prone areas; Bell 2019).

Machine learning framework for short-term fire potential forecasting

In this study, we introduce a ML framework for the forecasting of sub-hourly changes in fire potential that incorporates the extraction of precursor drivers and patterns as detectable changes or anomalies in environmental conditions that precede a fire – from a single weather station’s time series, employing time series feature engineering (Christ et al. 2018). This method, originally developed for volcano eruption forecasting (Dempsey et al. 2020), has been adapted for the fire potential forecasting problem. It involves the development of a classification model to estimate the likelihood of a fire occurring, assuming an ignition event takes place. The model looks for weather conditions that have historically allowed fires to ignite and spread. Whereas typical fire danger metrics indicate potential levels of fire intensity, rate of spread, or overall fire behaviour, our models specifically classify the conditions under which fire ignition occurs more frequently. They also calculate the probability of fire in a future window, usually several days, updated at the frequency of the data resolution, which for this study is every 30 min. While still in the prototype phase, our model identified pre-fire conditions and provided advanced warnings for previously unanticipated fire events. During pseudo prospective testing on one station in the Sunshine Coast (Queensland, Australia), it achieved this for 11 out of 15 recent significant fire events, outperforming standard metrics.

In this study, we distinguish between two important concepts: ‘fire danger’ and ‘fire potential’. Fire danger typically refers to an assessment of the fire environment, combining elements such as weather, fuels and topography to estimate the ease of ignition, potential spread rate, difficulty of control and likely impact of fires. Commonly expressed through indices like the Fire Danger Index (FDI), fire danger ratings support fire management strategies and public warnings. Conversely, ‘fire potential’ in this study specifically denotes the likelihood or probability of fire occurrence under certain environmental conditions, assuming an ignition source is present. Rather than evaluating potential fire behaviour, our fire potential focuses on identifying patterns in weather (and, ideally, fuel conditions, though the latter are not considered in this study) historically associated with fire likelihood, enabling real-time updates to anticipate conditions conducive to fire ignition and spread. This differentiation between fire danger and fire potential is important to developing adaptive forecasting systems that respond dynamically to changing conditions.

Data processing

Model training and testing

Early warning model

Forecast performance and benchmark

In this research, a good forecasting model is one that outputs a strong positive response (values approaching 1) in the days before a fire and a relatively depressed response (values closer to 0) during the extended periods between fires. The discriminability of these models is quantified by their ROC curve (see Methods) and associated AUC. Discriminability assesses the ability of forecasting models to differentiate between different classes (Ardid et al. 2022), specifically, pre-fire and non-fire data in this context. When discriminability is low, it is probable that forecasts will have unacceptably high rates of false positives.

Fig. 2 provides an overview of the forecast performance of the three models throughout the entire study period, juxtaposed with the reported FFDI values (Fig. 2a). Notably, the values preceding fires from the three models show a distinct separation, or discriminability, from the non-fire data (Fig. 2d, e, f). In contrast, the FFDI values do not display such a clear differentiation (Fig. 2g).

Fig. 2.

(a–c) The solid lines represent the consensus values of two models, showcasing the output of the trained models over the entire dataset. High values of the models (green and red curves) indicate high fire potential. We test whether the model correctly anticipated fires and examine its values during non-fire periods. The model’s performance over the entire period is evaluated quantitatively later in the histograms (d–g). Additionally, the FFDI values for the entire period are provided. A reference threshold of 0.7 is denoted by a horizontal dashed gray line, while recorded fires in the region are marked by vertical dashed black lines. (d–g) Histograms display the consensus values of three ML models, distinguishing between pre-fire and non-fire values. In panel (f), the same distributions are shown for FFDI values. For these histograms, the greater the separation between the distributions, the more distinct the pre-fire values are from the non-fire values as output by the model and FFDI. This separation is quantified by the Kolmogorov-Smirnov (KS) statistic. A larger split indicates better model performance. Please note that for this histograms, pre-fire values encompass a period of 10 days before the reported fire times, whereas non-fire values cover all days except those within 15 days around the reported fire times.

To assess discriminability across various thresholds, we computed ROC curves, which are a graphical measure of the performance of a classification model across all possible thresholds (Fig. 3). As discriminability improves, a ROC curve will depart from the diagonal towards the top left corner, denoting a reduced false positive rate. ROC curves for all forecast models trained on different window size have comparable discriminability, indicating all timescales recover relevant pre-fire patterns information. All models have similar or better discriminability than the existing FFDI (Fig. 3).

Fig. 3.

ROC curves (receiver operating characteristic) and AUC (area under the curve) illustrating and comparing the performance of three ML forecasting models and a simplified forecaster based on FFDI using thresholds (see Methods). The diagonal line represents a random model (a model that makes predictions with no discrimination between classes), and the AUC in the legends indicates the area under the curve. Each point on the ROC curve corresponds to a threshold (100 evenly spaced thresholds between 0 and 1; see Methods). As the performance of the model improves, the ROC curve departs from the diagonal towards the top left corner, indicating a reduced false positive rate. Using AUC to evaluate performance, we find that ML models improve average performance by 47% compared with the FFDI approach. Among the ML models, the 10-day model outperforms the others regarding discriminability and accuracy.

Although all time windows capture relevant pre-fire patterns, the 10-day window generally performs better than the 2-day window, likely owing to its ability to account for longer-term, cumulative weather conditions that may gradually elevate fire risk. The 2-day window, which focuses on short-term fluctuations, may be more susceptible to rapid weather changes that do not consistently indicate pre-fire conditions, leading to a less stable signal. In some instances, like the fire in 2007 (Fig. 2b), the models trained on different windows may appear anti-correlated. This could be due to short-term weather variations affecting the 2-day model’s predictions while the 10-day model, which averages out short-term noise, maintains a more consistent signal. Thus, the 10-day window may better capture slowly building conditions conducive to fire, providing improved predictive performance in many cases.

Models can be directly compared using the area under the ROC curve (AUC), which is a single numerical measure of average model performance. Using AUC to evaluate performance, we find that ML models improve average performance by 47% over the FFDI approach (Fig. 3).

Interpretability: data streams significance and features that strengthen prior to fires

The combination of time series features engineering with training Random Forest classification models allows us to investigate how the data inform the models. The choice of Random Forest can help to quantify variables’ attribution, interpret models’ predictions and ultimately improve our understanding of the drivers of fires.

We can identify which data streams and features hold more significance for the accurate forecasting of fire activity. Our ML models exploit this relationship at the time series feature level. For instance, quantile features derived from wind speed and relative humidity data are shown in Fig. 4a before a fire, displaying clear increases and peaks before the event. This and other fires are later used for training a model, where these features were identified as relevant when training the Random Forests among dozens of others. Fig. 4b shows the model output when run over an out-of-sample fire, model that increase before fire, indicating the efficacy of these identified precursors when forecasting an out-of-sample event. This is further emphasised in Fig. 4c, wherein the same features identified as significant when training (Fig. 4a) showed elevated values prior to the 2009 fire.

Fig. 4.

Illustration of precursor identification and transfer to out-of-sample target fire. The figure demonstrates how significant features identified during model training (in-sample) can be transferred to forecast fire potential in out-of-sample events. The aim here is to show that two features identified during training are triggering the model’s response on out-of-sample data. (a) Wind speed quantile 0.1 (magenta) and relative humidity quantile 0.3 (blue) show elevated levels before the 2004 fire (in-sample). (b) The machine learning model’s fire potential forecast for the 2009 fire (out-of-sample) is shown, with higher consensus values indicating increased fire potential. (c) The same features identified in (a) exhibit elevated levels prior to the 2009 fire (out-of-sample), supporting the transferability of these precursors. Note: the machine learning models developed here exploit learning from weather data at the time series feature level. For example, wind speed quantile 0.1 and relative humidity quantile 0.3 are discriminated in pre-fire sequences from the training ensemble. Later, these same features are seen to strengthen prior to an out-of-sample fire in the target area. This demonstrates the functional basis of precursor transfer among different fire events.

The significance of various data streams within the context of our trained Random Forests provides an extra layer of interpretability, offering insights into the frequency of a data stream’s use in informing decision trees (Fig. 5). Noteworthy trends emerge, with key variables such as temperature (T), RH, wind speed (V), VPD and FFMC being recurrently employed by decision trees. This observation underscores their relevance and role in forecasting fire potential.

Fig. 5.

Significance of data streams in the trained ML models. This visual depicts the frequency of a data stream’s utilisation in informing all decision trees in the random forest, serving as an indicator of its importance for forecasting fire potential. Higher log occurrences indicate greater importance in the forecasting model. The x-axis represents the log occurrence of each data stream being used in the model’s decision-making process. The data streams include T (temperature), DP (dew point), RH (relative humidity), V (wind speed), Theta (potential temperature), KBDI (Keetch–Byram Drought Index), DF (Drought Factor), VPD (Vapour Pressure Deficit), FFMC (Fine Fuel Moisture Code) and Grass FFMC.

Supporting decisions in warning systems

The objective of natural hazard forecasts is to provide decision-makers with objective, quantitative information regarding the likelihood of an extreme event occurrence within a specified warning period (Sorensen 2000). Here, we consider that warnings could be triggered when the ML model output exceeds a threshold and end once model output has dropped below the threshold for more than 48 h (see Methods).

Fig. 6 shows a modified version of the ROC curve designed specifically for the 10-day optimised warning system that takes into account various metrics such as true positives, false positives, true negatives and false negatives based on a warning window (see Methods). This metric is useful for computing conditional probabilities of having a fire for a certain threshold using historical data and hazard ratios (see Methods). The model applied over the 15.4-year dataset with a threshold of 0.7 accurately forecast 75% of fire potential conditions (11 out of 15; Supplementary Table S1) occurring within the designated warning period, constituting approximately 5% of the total time span, equivalent to 282 days (Fig. 6). This equates to a ~4% probability of fire occurrence in a 48-h period when a warning is in effect, as opposed to a 0.08% chance of a fire during no-warning.

Fig. 6.

Alternative ROC curve design for an optimised warning system. This figure illustrates the performance of the warning system based on the 10-day ML model tested on the Sunshine Coast record (15.4 years). The ROC curve plots the proportion of fires during alerts (y-axis) against the proportion of days under alert (x-axis) for various threshold values. For a threshold of 0.7 (indicated by the star), the system spends 5.5% of the time under warning, covering 282.6 days, and captures 75.0% of the fires, equating to 11 fires. The calculated rates of fire occurrence are λ_in = 0.0389 when a warning is in effect and λ_out = 0.00075 when no warning is active. These translate to probabilities of 3.89 and 0.0749% for fire occurrence in warning and non-warning periods, respectively, yielding a hazard ratio of 51.9. The warning system becomes active when the output of the forecast model reaches or exceeds a predetermined trigger threshold. This system helps to effectively balance the trade-off between missed fires (false negatives) and false alarms (false positives).

Specifically, the variables indicated in the Methods sections for computing the conditional probabilities and hazard ratio are: M_in = 11, M_out = 4, T_in = 282.6 days, T_out = 5369.4 days. The calculated fire rates (per 48 h) are λ_in = 0.0389 and λ_out = 0.00075, and the probabilities are P_in = 3.89 and P_out = 0.0749. Finally, the hazard ratio is 51.9 (Fig. 6). Given the inherently low probability of fire events, it is more meaningful to express conditional probabilities as hazard ratios. These ratios quantify forecast accuracy in terms of the relative likelihood of a fire event under different warning conditions. In the example above, fires are predicted to be 50 times more likely during an active warning compared with non-warning conditions.

Practical challenges on operational forecasting

Demonstrating the forecasting skill of transferred fire precursors is a necessary but not sufficient step for their operational use. Fire monitoring teams routinely integrate a wide range of signals across disparate time scales (Ghali et al. 2020; Mohapatra and Trinh 2022; Ardid et al. 2024a) and use these to inform mental models (expert judgement) of the evolving fire hazard. If they are to provide complementary input to this process, forecasting models need to be interpretable on a real-time time-series basis. Although this problem remains outstanding, a presentation of individual forecast characteristics is useful here to define its challenging aspects.

A reliable forecast should consistently demonstrate accuracy, free from significant errors or biases, thereby establishing confidence in the forecasting process and its outcomes. In this study, we considered that a fire is anticipated by our forecast fire potential model, i.e. a true positive, if the model output (consensus) surpasses a predetermined threshold the days defined by training window (2, 5, and 10 days) before the recorded fire. Once the model output meets this criterion, it is deemed to have been anticipated. For instance, Fig. 7 shows the out-of-sample performance of the model trained using a 10-day window before the largest fire in our catalogue, which occurred in January 2017. The model’s predictions steadily increase over 2 weeks, reaching a peak a day before the event. The reference threshold is crossed ~4 days before the fire, triggering a warning that continues until the fire occurs. This extended warning period is designed to ensure that high-risk conditions are captured and maintained if fire-prone conditions persist. The warning remains active as long as the model’s output consistently exceeds the threshold, signalling sustained fire potential. If the consensus drops below the threshold for 48 continuous hours, the warning would be cancelled; however, in this instance, conditions remained above the threshold. The FFDI also reaches high values a couple of days before the fire, but its highest values for this period are recorded 2 weeks earlier, over 4 days (7 to 9 January), while the ML model remains low during this period (Fig. 7a). The difference between the consensus model output and FFDI values, despite both using the same weather variables, arises from the time series feature engineering and selection process within our ML approach. This approach allows our model to capture nuanced, longer-term and hidden patterns in the data that may indicate increased fire potential. By identifying these underlying risk factors, our model can provide early indications of fire potential, complementing the FFDI’s real-time responsiveness to immediate conditions.

Fig. 7.

Performance of the forecasting model trained with a 10-day window prior to the largest fire event in the catalogue, which occurred in January 2017 and burned 843 ha (out-of-sample, i.e. this fire was not considered during training). (a) The model’s consensus output (green line) and FFDI (black line) are plotted against time. The consensus steadily increases over 2 weeks, peaking a day before the fire. The warning threshold (grey dashed line) is crossed approximately 4 days before the fire, triggering a warning period (grey shaded area) that persists until the fire occurs. The fire event is highlighted by a red shaded area. The FFDI also shows increasing values a couple of days before the fire, with the highest values recorded 2 weeks earlier (7 to 9 January), a period during which the ML model remains low. It is important to note that an FFDI of 7 falls into the low/mid category. (b) Time series of wind speed (magenta line), temperature (orange line) and relative humidity (sky blue line) during the same period. The fire event is again highlighted by a red shaded area. This figure illustrates the model’s capability to anticipate fire events by showing a clear increase in the consensus output leading up to the fire, compared with traditional indices like the FFDI.

It is notable that the FFDI peaks precisely during the fire event, whereas the event occurs as the consensus from our model begins to decline. This discrepancy may arise from the differences in how FFDI and our model respond to short-term versus accumulated conditions. FFDI is highly sensitive to immediate weather conditions and reflects rapid fluctuations, whereas our model’s consensus score incorporates longer-term fire-prone patterns detected over the prior 10-day period. This approach captures fire risk conditions that build up and signal fire potential in advance rather than aligning with immediate peaks in weather-driven indices. This complementary behaviour between FFDI and our model could provide a large-scale warning system, where early indications from our model can be reinforced by high FFDI values closer to the event, improving operational response.

Across all reported fires, the models exhibit good performance when applied to fires that were not part of the training data, i.e. out-of-sample (Supplementary Figs S1 and S2). Models are effective in forecasting fire potential when there is a significant increase in their output before the fires occur (e.g. Supplementary Fig S1a, d, e, f), but less so when no increase is observed beforehand (Supplementary Fig S1b, c). Nevertheless, the models remain valuable when an elevated value is detected several days before the fires (e.g. Supplementary Fig S1b, c, e, f) as we are estimating fire potential conditions that do not always result in actual fires (see Supplementary Fig S2 for all events). Using a reference threshold of 0.7 (e.g. Fig. 7, Supplementary Fig S1), the highest accuracy achieved by the tested forecasters in predicting out-of-sample fires for the 10-day model was 75% (11 out of 15 fires) (Supplementary Fig S2; Supplementary Table S1). In comparison, the 2 and 5-day models achieved 33 and 20% accuracy, respectively, at this threshold. It is important to note that, by lowering the threshold, it is possible to eventually anticipate all fires. However, it is essential to ensure that the associated false positive rate remains acceptable for the chosen threshold. We explored this balance in previous sections, where we tested a range of thresholds and evaluated different metrics to identify the optimal one (Figs 2, 3).

Assimilating diverse data sources for enhanced fire potential forecasting

Expanding the forecasting framework to include a broader array of time series data, such as remote sensing parameters like drought indices, vegetation health and additional weather variables (Huot et al. 2020), can refine fire potential assessments. Integrating remote sensing data provides a comprehensive environmental view, capturing indicators that may not be discernible through on-site weather station measurements alone. Drought indices reveal soil moisture levels, influencing fuel availability and fire ignition. Vegetation health metrics provide valuable information on landscape flammability, identifying areas prone to rapid fire spread. Including terrain ruggedness is also relevant, as fires in rugged terrain typically escalate owing to topography and the difficulty of early suppression efforts. This diverse data integration allows the model to better understand the complex interplay between environmental conditions and fire behaviour, enhancing accuracy and reliability.

Currently, the model focuses primarily on weather metrics, which does not capture all factors influencing fire behaviour. Understanding the site context is key to evaluating whether weather is the primary driver or if other factors play a significant role, such as fuel properties, fuel connectivity, topographical differences, moisture gradients and atmospheric stability. For example, unanticipated fires might have been influenced by these unaccounted factors. Our data include fires mainly in three areas: Noosa, Mount Coolum and Maroochy River National Parks, each characterised by distinct fuel types – primarily eucalypt, heathland and melaleuca, respectively (Supplementary Table S1). Preliminary analysis suggests that different fuel types may influence fire behaviour and the likelihood of anticipation, with heathland and melaleuca being less likely to be associated with anticipated fires. However, there are not enough data to draw definitive conclusions owing to an unbalanced dataset.

Additionally, live fuel moisture, which refers to the water content in living vegetation, directly impacts flammability and is a significant predictor of fire activity and behaviour (Yebra et al. 2013, 2018). Future work should incorporate complex vegetation and topography (Si et al. 2022), as well as hydrological variables like soil moisture or stream flow data, to improve fire potential forecasts (Farahmand et al. 2020). Factoring in wind direction changes, such as sudden shifts from west to southeast, which are crucial in escalating established fires, could also enhance model precision.

Complementing recent advances in fire danger forecasting

The Australian Fire Danger Rating System (AFDRS) is a new fire danger rating system for Australia calculated daily by the Bureau of Meteorology using forecast weather at hourly intervals on a 1.5 × 1.5 km grid across Australia (Matthews et al. 2019; Hollis et al. 2024). The NSW Rural Fire Service trial of the AFDRS during the 2021–22 bushfire season highlighted the advantages of the new system, despite the need for ongoing adjustments (McCoy and Field 2022). Our study introduces an ML-based framework that uses high-frequency weather data and updates for sub-hourly predictions, improving forecast responsiveness. By leveraging ML models, our framework complements the AFDRS by enhancing the accuracy and timeliness of fire danger forecasts, thus supporting operational responsiveness. This aligns with the goals of the AFDRS to support fire management operations and public communications regarding potential fire danger (Hollis et al. 2024). The modular nature of the AFDRS supports continuous and incremental improvements, incorporating the best available science to estimate fire potential accurately (Hollis et al. 2024). Therefore, our approach could be integrated into the AFDRS framework, allowing improvements in predictive accuracy and operational effectiveness to be evaluated.

Probabilistic forecasting techniques, like those in the Global ECMWF (European Centre for Medium-Range Weather Forecasts) Fire Forecast (GEFF) system (Worsnop et al. 2021), enhance accuracy through statistical post-processing. Our framework aligns with this approach by using ML models to update forecasts in real time. Whereas GEFF focuses on 2-week forecasts for longer-term planning, our model provides immediate, short-term predictions, bridging the gap between long-term and real-time needs. Advancements in deep learning have shown effectiveness in capturing spatio-temporal patterns (Kondylatos et al. 2022). Our framework complements these efforts through time series feature engineering and Random Forest models, which offer some interpretability through feature importance scores. However, the reliance on feature importance from decision trees provides limited insight into causal relationships within the data. To address this, future work could incorporate more advanced interpretability techniques, such as SHAP values (SHapley Additive exPlanations; e.g. Qayyum et al. 2024) or causal inference, to enhance model transparency and improve operational decision-making.

Research in mountainous plateau regions has used logistic regression to address topographical and meteorological complexities (Si et al. 2022). Our framework complements this by capturing fine-scale temporal variations and integrating a broader range of weather variables. However, future work should incorporate complex vegetation and topography to improve precision in challenging environments. The integration of hydrological variables, such as soil moisture, is necessary for enhancing fire danger forecasts (Farahmand et al. 2020). Our current model focuses on weather metrics, but incorporating hydrological data and remote sensing could potentially enhance our predictions. Remote sensing would be particularly beneficial in improving fuels data and accounting for topographic complexity, providing a more comprehensive understanding of the factors influencing fire behaviour.

Multi-station data integration and simplifying the complexity of fire regimes

To improve the accuracy and robustness of fire potential assessments, future work should consider integrating data from multiple stations across diverse eco-regions. This multi-station approach would provide a broader and more comprehensive dataset, allowing the model to capture regional differences in fire occurrence and better adapt to the variability across distinct terrains and ecosystems (Fujioka et al. 2008). Such regional differences are critical, as models developed using data from a single region may not generalise effectively to other eco-regions with unique fire environments – including fire weather, fuels and topography. In areas with complex terrain, individual weather stations may not accurately reflect localised fire–weather interactions. Additionally, fuel type and moisture content are also critical factors influencing ignition and fire spread, further underscoring the importance of a diverse geographic data set.

To simplify the complexity of fire regimes and the diversity of fire-prone ecosystems, we can use the dichotomy between fuel-limited and fuel dryness-limited fire regimes. Fuel-limited fire regimes, found in grass-dominated semiarid interiors, are primarily constrained by the availability of combustible material. In contrast, fuel dryness-limited fire regimes, typical of forests and woodlands with litter fuels, depend on the moisture content of the fuel (Bradstock 2010; Boer et al. 2021). Incorporating these distinctions into the model can help address variability in fire behaviour across ecosystems, considering both weather and fuel conditions for a more comprehensive fire forecasting system. Integrating data from multiple stations across various regions will enhance our understanding by providing diverse observations, refining the model to better capture regional fire behaviour nuances.

Addressing data scarcity challenges through transfer machine learning

In light of the escalating impact of climate change on fires, it is increasingly important to address the issue of data scarcity. For instance, countries with scarce historical data that may be facing increasing fire danger due to a combination of climate conditions, vegetation types and human activities, could benefit from ML models like the one presented here. Our research aims to build models capable of training in one region and applying their forecasting capabilities to another, even in the absence of historical data, by leveraging the principle of transfer ML.

Our approach relies on assimilating 30 min increment time series of weather variables as a minimum requirement. This method could revolutionise real-time fire assessment, offering significant advancements in fire management and safeguarding communities across diverse geographical domains. By utilising existing, cost-effective infrastructure, our model’s implementation is straightforward, making it a valuable tool for communities and agencies working towards effective fire management with limited resources. However, understanding how the specific model generalizes to other regions remains an area for further investigation.

With significant climate shifts projected in the coming century, strategic fire adaptation and mitigation measures are critical. Our approach addresses these challenges by providing a scalable and adaptable solution, enhancing the precision and effectiveness of fire warning systems, and ultimately helping to protect vulnerable populations and ecosystems from the increasing threat of fires.