Southern California, a biodiversity hotspot and home to over 23 million people, faces a significant challenge from wildfires. While the annual wildfire area hasn't shown a substantial change in recent decades, the impact of future climate change on fire regimes remains a critical concern. This study focuses on the coastal Southern California area (CSCA), a region experiencing some of the highest property losses from wildfires in the United States due to its unique fire characteristics, large population, and wildland-urban interface development. California's Mediterranean climate, with its wet winters and dry summers, is naturally conducive to wildfires. Anthropogenic warming has increased aridity and drought risk, contributing to increased fuel aridity, longer fire seasons, and more wildfire activity across much of the state. However, the CSCA presents a unique case, as there has been no significant trend in annual or seasonal total burned area over the past five decades. This lack of trend is likely due to a combination of factors including high interannual climate variability, reduced ignitions, improved fire suppression, and land cover change. Existing research offers conflicting views on climate's role in determining future fire activity in the CSCA, with some suggesting climate will not be a major determinant, while others emphasize its crucial role. This disparity underscores the need for high-resolution spatial and temporal analysis to accurately predict future fire risks in this geographically small yet climatically diverse region. Therefore, this study employs station-based climate projection data and a machine learning approach to model the relationship between daily climate and large wildfire probability, projecting changes in fire occurrence in response to future climate scenarios.

The literature surrounding the impact of climate change on wildfire risk in Southern California is mixed. Some studies predict an increase in fire probability, burned area, or fire-danger days, while others suggest a decrease. These differing projections may stem from the coarse spatial and temporal resolution employed in previous studies, which limits their ability to provide detailed information for smaller regions like the CSCA. Studies have highlighted the influence of anthropogenic warming on increased aridity and drought risk in California, leading to more frequent and intense wildfires across much of the state. However, in the CSCA, the lack of a clear trend in annual or seasonal burned area over several decades has led to different interpretations of climate’s role. Some research emphasizes other factors, such as fuel management and human ignitions, as more significant drivers of fire activity in lower elevation areas. Conversely, other studies show that climate change alone has driven an increase in large fires, suggesting that the interaction between climate change and the presence of human ignition sources requires further investigation. This study attempts to reconcile these conflicting conclusions by improving both the spatial and temporal resolution of the analysis.

This study uses a sophisticated approach combining high-resolution climate data with machine learning techniques to model daily wildfire probability in the CSCA. The researchers statistically modeled the relationship between daily climate and the probability of large (>40 hectares) wildfires at a local scale. The data utilized included meteorological variables (vapor pressure deficit (VPD), wind speed (WS), precipitation) and fire-danger indices from the National Fire Danger Rating System (NFDRS), such as energy release component (ERC), burning index (BI), spread component (SC), ignition component (IC), and 100-h (F100) and 1000-h (F1000) dead fuel moisture. Data from 49 weather stations across the CSCA were used, along with climate simulations from 14 CMIP5 Earth System Models (ESMs) for historical (1950–2005) and future (2006–2099) periods under both RCP4.5 (moderate emissions) and RCP8.5 (high emissions) scenarios. These ESM data were downscaled to each weather station. Random forest algorithms were employed to model daily large fire probability (LFP), allowing for the capture of non-linear relationships between predictor variables. The model was trained and validated using a five-fold cross-validation procedure, employing out-of-bag samples to assess model performance. Separate models were developed for the dry season (April–September) and wet season (October–March) to account for seasonal differences in fire behavior. The relative importance of predictors was determined using a permutation-based approach, and accumulated local effect (ALE) analysis was used to visualize the relationships between LFP and key predictors. To address the issue of imbalanced data (more non-fire days than fire days), a cluster-based resampling technique was applied. A bias-correction linear regression model was used to adjust for overestimation of LFP observed after resampling. The study projected changes in LFP under different climate scenarios by applying the validated random forest models to the future climate simulations.

The random forest model accurately predicted large fire occurrence (82-84% accuracy). Key findings include: * **Key Drivers of Large Fires:** The most important predictors of large fire probability (LFP) varied seasonally. In the dry season, VPD, IC, F1000, and ERC were the most influential, while in the wet season, F1000, VPD, WS, and IC were dominant. Higher VPD generally increased LFP linearly in the dry season, while in the wet season, high VPD only increased LFP at very high VPD levels. Abnormally dry fuels (low F1000) exponentially increased LFP in the wet season. * **Projected Increases in LFP:** LFP is projected to increase significantly across all months under both RCP4.5 and RCP8.5 scenarios. The annual mean LFP increase is approximately 39% for RCP4.5 and 62% for RCP8.5 by the late 21st century. The most substantial increases in LFP are projected in the spring and autumn seasons. * **Seasonal Shifts:** The large fire season is predicted to lengthen, with an earlier onset and later end. This is due to increased LFP in spring and autumn. The most substantial relative increases occur in the late spring and early summer transition period. * **Interannual Variability:** The high-emission scenario (RCP8.5) will lead to significantly more large fire days annually than the moderate emission scenario (RCP4.5). By 2070-2099, the average number of large fire days is projected to increase from 36 days (1970-1999) to 58 days (RCP4.5) and 71 days (RCP8.5). This divergence between RCP4.5 and RCP8.5 begins to appear in the mid-21st century. * **Influence of Santa Ana Winds:** While Santa Ana winds are expected to decline slightly, this may not offset the increased LFP driven by other factors such as higher VPD and lower fuel moisture.

This study's findings strongly suggest that Southern California will experience substantial increases in large wildfire events, significantly expanding the length and intensity of the fire season. The projected near-doubling of large fire days by 2100 under the RCP8.5 scenario highlights the severity of the potential risk. The methodology used, employing high-resolution, station-based data and advanced machine learning techniques, provides a more refined and accurate prediction than previous studies. The seasonal shift towards an extended fire season, particularly in the spring and autumn, emphasizes the need for proactive adaptation strategies. The key drivers, elevated VPD and reduced fuel moisture, clearly indicate the dominant role of climate change. While the projected slight decrease in Santa Ana winds might offer some mitigation, it is unlikely to offset the larger effects of warming and drying. The significant increase in fire risk is linked to broader circulation changes associated with global warming, including the expansion and intensification of the Hadley Circulation, which leads to regional drying. The improved methodology employed allowed the researchers to connect general circulation model outputs to local fire risks more effectively than prior work. Future research needs to incorporate factors like fuel management and human ignition patterns to fully understand the complex dynamics of wildfire occurrence in this region.

This study provides compelling evidence for a substantial increase in large wildfire days in Southern California under various climate change scenarios. The high-resolution modeling approach, incorporating station-based data and advanced machine learning, offers a crucial advancement in predicting fire risks. The results emphasize the need for proactive and comprehensive wildfire management strategies, considering both climate change impacts and the specific seasonal dynamics of large fire events. Future research should explore the combined effects of climate change, fuel management policies, and human behavior to further refine wildfire risk assessments and inform effective mitigation efforts.

The study acknowledges limitations inherent in using limited historical fire data, which may not fully represent the region's true fire regime. Also, the exclusion of small fires, primarily driven by human ignition, may underestimate the total fire activity. The study assumes no radical changes in fuel management practices, which may not be entirely realistic. The daily scale modelling, while more precise, is constrained by the limited availability of large fire records for this specific region. The study also excludes the impact of fire management practices and human behavior on fire occurrence.