Fire

Forest fires cause severe ecological, economic, and social losses. Machine learning (ML) has been increasingly applied in forest fire prediction. This study examines advances and challenges in ML-based forest fire prediction. The analysis covered 811 publications indexed in the Web of Science from 2010 to March 2026. Bibliometric analysis and VOSviewer were used to examine publication trends, author and institutional co-authorship networks, and keyword co-occurrence. Mainstream ML models and their applications were also compared. The burden–output gap (G_i) quantified differences between affiliation-based publication shares and mean annual forest burned area shares for 2012–2023. Results showed that publication output increased rapidly after 2020, while author and institutional collaboration remained fragmented. Research hotspots mainly involved the integration of ML with geographic information systems and remote sensing. The application analysis covered fire susceptibility assessment, fire spread prediction, and fire danger forecasting and early warning. Key challenges included data imbalance, limited cross-regional generalizability, weak model interpretability, computational demands, and climate-related uncertainty. Future research should improve model interpretability and cross-regional applicability. Multi-source data and climate projections should be integrated into real-time early warning systems. The geographical distribution of affiliation-based publication output differed from that of forest burned area. Several countries with extensive forest burned areas had limited or no authorship representation in the selected publications. The geographical comparison informs country-level investigations of local data needs, model transferability, and research collaboration. These findings inform model selection and international research collaboration.

Current international design standards for timber structures under fire exposure rely on simplified reduced cross-section methods (RCMs) that generally assume constant charring rates and a fixed zero-strength layer (ZSL) thickness, typically 7 mm. These assumptions do not account for the influence of timber density, moisture content, cross-sectional geometry, or fire exposure configuration on charring and thermal degradation. This study presents an enhanced reduced cross-section method (ERCM) for the fire design of Glulam beams by incorporating these parameters. The underlying numerical model was independently validated against experimental thermal and structural fire tests, demonstrating accurate predictions of both charring rates and structural deflection under fire exposure before conducting the 65 coupled thermal–structural simulations detailed in a companion manuscript. Analytical models for average charring rate and ZSL evolution were then formulated using nonlinear regression of this validated numerical dataset. The average charring-rate model incorporates timber density, moisture content, fire exposure duration, and number of fire-exposed faces. The ZSL model is formulated as the product of a Peak ZSL, which depends on density, moisture content, and cross-sectional geometry, and a time-dependent ZSL coefficient that depends on density, moisture content, fire exposure duration, and number of fire-exposed faces. The models yielded coefficients of determination (R²) of 0.982 for average charring rate, 0.973 for Peak ZSL, 0.914 for time-dependent ZSL coefficient, and 0.927 for time at Peak ZSL. The resulting analytical models were incorporated into the reduced cross-section framework and applied to four design examples representing 60 min residential and 90 min office applications, considering Glulam densities of 300 and 450 kg/m³. The ERCM was compared with Eurocode 5 Simplified Method (EC5, SM) and AS/NZS 1720.4 using consistent loading and material assumptions. The comparison demonstrates substantial differences between the design methods, particularly at low timber densities and longer fire exposure durations. The spread between the methods increases as density decreases and fire exposure duration increases, with the constant-parameter EC5 (SM) approach becoming non-conservative for low-density glulam at long fire exposure durations, while AS/NZS 1720.4 becomes overly conservative under the same conditions. The results illustrate the influence of density-dependent charring and ZSL evolution on the predicted residual cross-section and fire utilisation. The proposed ERCM provides a physically representative framework for evaluating fire resistance, though its application beyond the validated parameter ranges requires further experimental confirmation.

Long-term trends, seasonal fluctuations, and short-term disturbances are superimposed in land surface temperature (LST) time series, complicating the identification of persistent coal-fire thermal anomalies. Based on 1188 Landsat scenes of the Midong mining area, Xinjiang, China, acquired during 1990–2025, this study applied Seasonal-trend decomposition using LOESS (STL) to extract long-term temperature trends. The contemporaneous spatial mean of the trend components in non-mining areas was used as the background reference to construct the background-corrected temperature difference ΔLST. Thermal anomalies were identified using year-specific thresholds, and their spatiotemporal evolution and relationship with mining-land changes were analyzed in conjunction with land-use/land-cover (LULC) records. In 2020 and 2024, the matching rates with field-measured high-temperature points were 75.0% and 83.3%, respectively, for Z > 2, compared with 47.2% and 41.7% for Z > 3. To focus on thermal anomalies with larger standardized temperature deviations and to examine their long-term evolution, subsequent analyses used the Z > 3 results. Thermal-anomaly area peaked at approximately 1.7 km² in 2016 and subsequently declined overall, but anomalies were repeatedly detected at some locations within open-pit mining areas during 2016–2025. From 2014 onward, thermal anomalies in these mining areas accounted for more than 80% of the total identified anomaly area. The mining-land area ratio and the thermal-anomaly area ratio were positively correlated (Spearman’s r = 0.376, p = 0.0245), but mining-land area ratios were generally lower in 2016–2025 than in 2005–2015, while thermal-anomaly area ratios remained relatively high. At comparable mining-land area ratios, thermal-anomaly area ratios were generally higher in the later period. Long-term records of thermal anomalies and LULC reveal patterns of localized anomaly recurrence that are difficult to capture using annual area statistics or the current extent of mining land, providing a basis for determining long-term monitoring priorities in mining areas.