The season for large fires in Southern California is projected to lengthen in a changing climate

The study addresses how future anthropogenic climate change will alter the timing and frequency of large wildfires in coastal southern California, a region with Mediterranean climate, high biodiversity, and large population. Despite statewide increases in wildfire activity linked to warming-driven aridity, the coastal southern California area has shown no significant trend in burned area over recent decades, likely due to high interannual climate variability, ignitions/suppression changes, and land cover shifts. Conflicting prior studies alternately downplay or emphasize climate’s role in lower-elevation southern California fire regimes, and highlight the potential interaction between climate-driven fuel aridification and ubiquitous human ignitions. Given projected increases in temperature, vapor pressure deficit (VPD), and extreme fire danger, the authors ask: Which climatic conditions produce large fire days in this region, and how will intra- and interannual variability in large fire days respond under moderate (RCP4.5) and high (RCP8.5) emissions? The purpose is to provide high-resolution, station-based, daily projections to overcome spatial/temporal limitations of prior work and to quantify seasonal changes in large fire season onset, intensity, and duration.

Prior work shows warming and drying have lengthened fire seasons and increased wildfire activity across much of the western United States, including California, with notable extreme seasons in 2017, 2018, and 2020. In coastal southern California, Santa Ana wind (SAW) events drive autumn-winter fire peaks, while summer fires are linked to hot, dry conditions; yet annual/seasonal burned area trends have not significantly changed over recent decades. Studies disagree on climate’s future influence at lower elevations and latitudes: some predict increased fire probability and fire-danger days, others suggest decreases or strong roles of human development and management. Evidence indicates large fires are more closely tied to weather/climate than small fires, and that climate change plus persistent human ignitions may elevate large fire risk. Projections of continued warming and higher VPD motivate fine-scale analyses. Prior studies often use coarse spatial/temporal resolutions; the present work applies station-based downscaling and machine learning to better resolve small-region dynamics and seasonal drivers (e.g., VPD, fuel moisture, SAWs).

- Study area and data: Coastal southern California (~41,000 km²). Fire perimeters from California Department of Forestry and Fire Protection (FRAP) for 1950–2019. Observations from 49 Remote Automatic Weather Stations (RAWS) for 1996–2010. Daily historical (1950–2005) and future (2006–2099) climate from 14 CMIP5 Earth System Models (ESMs) under RCP4.5 and RCP8.5, statistically downscaled to the 49 stations using the multivariate adaptive constructed analogues method (MACA). - Predictors: Daily meteorological variables and National Fire Danger Rating System (NFDRS) indices: vapor pressure deficit (VPD), wind speed (WS), precipitation (as percentiles), energy release component (ERC), burning index (BI), spread component (SC), ignition component (IC), 100-h (F100) and 1000-h (F1000) dead fuel moisture. Elevation and canopy density were tested but excluded due to minimal contribution. VPD computed from daily Tmax, Tmin, and RH. - Response variable: Large fire presence/absence by station-day for fires >40 ha. Fire perimeters converted to a binary indicator per station-day using a station-centered buffer to associate fires with stations. - Buffering and sampling: Tested buffer radii of 5, 10, 25, 50, 100 km to link station conditions to nearby fires; selected 10 km based on cross-validated AUC. Generated separate datasets for dry (April–September) and wet (October–March) seasons reflecting distinct regimes (non-SAW vs SAW-driven). - Modeling approach: Random forest classification models trained separately for dry and wet seasons using standardized anomalies of predictors (except precipitation, used as percentiles). Five-fold cross-validation with out-of-bag samples defined as all data from three consecutive years within 1996–2010 to assess generalization. Parameter grid: maxnode 10–1000, mtry 2–8, ntree 10–2000; best dry-season parameters: buffer 10 km, maxnode 500, mtry 2, ntree 500; best wet-season parameters: buffer 10 km, maxnode 100, mtry 2, ntree 500. - Class imbalance handling: Applied k-means clustering based undersampling of the majority class (nonfire days) to balance classes and improve minority (fire) detection. - Bias correction: Balanced training led to overestimation of large fire probability (LFP). A post hoc linear regression calibration was applied between predicted and observed decadal mean daily LFP (1950–2019), explaining 67.4% of variance, to reduce bias. - Projections and diagnostics: Applied trained and bias-corrected models to downscaled ESM historical and future daily data to compute daily LFP and identify large fire days (LFDs) as days when simulated LFP exceeds an observed baseline annual LFP threshold. Computed monthly climatologies and interannual time series for baseline (1970–1999), mid-century (2040–2069), and late-century (2070–2099). - Evaluation and interpretation: Performance assessed with AUC, accuracy, precision, recall. Variable importance via permutation-based decrease-in-accuracy. Accumulated Local Effects (ALE) plots used to interpret predictor-LFP relationships. Low-pass filtering and LOESS smoothing applied to visualize trends; uncertainties discussed with respect to data limitations and scenario spread.

- Model performance: Random forest models achieved overall accuracy of ~82% (wet) and ~84% (dry), with AUC >0.7 for most configurations; optimal station-fire buffer was 10 km. Top predictors: dry season—VPD, IC, F1000, ERC; wet season—F1000, VPD, WS, IC. - Predictor sensitivities: ALE analyses show dry-season LFP increases approximately linearly with VPD; wet-season LFP increases nonlinearly with low F1000 (dry fuels) and with very high VPD; WS has elevated importance in wet season, consistent with Santa Ana wind influence. - Recent historical change: Simulated frequency of large fire days increased significantly from ~34 days/year (1950–1979) to ~43 days/year (2000–2019) (p<0.001). - Future annual changes: By ~2050 (2040–2069), both RCP4.5 and RCP8.5 project ~55 large fire days/year. By late century (2070–2099), large fire days increase from ~36 days/year (1970–1999 baseline) to ~58 (RCP4.5) and ~71 (RCP8.5). Annual mean LFP increases by ~39% (RCP4.5) and ~62% (RCP8.5). Scenario divergence becomes pronounced in the latter half of the century. - Seasonal changes and season length: LFP increases throughout the year under both scenarios, with particularly strong relative increases in spring and autumn. Spring–early summer (April–June) LFP increases by ~110% by late century under RCP8.5. July and September show additional large fire days due to already high LFP. The large fire season is projected to start earlier and end later, indicating a lengthened and intensified season. - Drivers of change: Projected warming (+~2.5 °C RCP4.5, +~5.5 °C RCP8.5 by late century, vs 1970–1999) increases VPD; fuel moisture (F1000) declines, especially in spring and autumn; IC and ERC increase, especially in autumn. Autumn/winter WS slightly decreases; studies suggest future suppression and sharpened seasonality of Santa Ana winds, potentially moderating some autumn-winter risks but not offsetting warming-driven aridity effects. - Hydroclimate context: Precipitation increases in winter but decreases in spring and late autumn, with greater interannual variability and fewer wet days punctuated by more intense storms, contributing to fuel aridification between events.

The findings directly address the research questions by identifying the dominant meteorological drivers of large fire probability and quantifying how their projected changes alter the frequency and seasonal timing of large fire days in coastal southern California. Warming-driven increases in VPD (particularly influential in the dry season) and reductions in 1000-h fuel moisture (dominant in the wet season) collectively elevate fire potential, producing both intensification within the traditional summer fire season and a lengthening into spring and fall. Although autumn-winter wind speeds (and potentially Santa Ana wind frequency) may decline slightly, this effect is insufficient to counteract the strong thermodynamic aridification signal. The results reconcile prior conflicting regional projections by using station-based downscaling and daily, local-scale modeling, demonstrating that climate change alone can substantially increase large fire-conducive conditions even in an area without recent increases in burned area. The broader significance is consistent with large-scale circulation changes under global warming (e.g., Hadley Cell expansion), which reduce relative humidity and increase dryness in the subtropics, amplifying wildfire risk in the U.S. Southwest. The approach and results provide actionable information for planning and management in Mediterranean-climate regions where fine spatial resolution is essential.

This study introduces a station-based, daily-resolution machine learning framework that integrates downscaled CMIP5 climate projections with NFDRS indices to model and project large fire probability in coastal southern California. It shows that large fire days are projected to increase from ~36 days/year (1970–1999) to ~58 (RCP4.5) and ~71 (RCP8.5) by 2070–2099, driven primarily by higher VPD and reduced 1000-h fuel moisture, and that the large fire season will both intensify and lengthen (earlier onset, later end). The methodology overcomes limitations of coarser-grained studies and provides detailed seasonal insights. Future research should incorporate dynamic human dimensions (ignitions, suppression, land use, and fuel management), explicitly model Santa Ana wind changes with higher-resolution wind fields, and leverage longer and richer fire datasets to reduce uncertainty and improve daily-scale predictability. The approach may be extended to other Mediterranean-climate regions to inform adaptation and risk mitigation strategies.

- Human factors: The models exclude explicit human activity and fire management changes; assumptions that large fires are primarily climate-driven and that fuel management remains similar may not always hold. - Data limitations: Limited years of station observations and large fire records for a small region reduce statistical power at daily resolution; observed LFP has high variance, and model simulations retain a 1–2 week seasonal phase offset. - Class imbalance and bias correction: Despite k-means undersampling, initial models overpredicted LFP and required linear bias correction (explaining 67.4% variance), introducing additional uncertainty. - Spatial linkage assumptions: Use of a fixed 10 km station-fire buffer may not capture all spatial heterogeneity in complex terrain and varying wind regimes. - Wind projections: Coarse ESM wind fields and modest projected changes in WS, plus uncertainties in future Santa Ana winds, may underrepresent wind-driven extremes. - Generalizability: Results pertain to the coastal southern California domain; extrapolation to other regions requires local calibration and downscaling.