Accounting for approximately 31% of the global land surface, forests play a critical and indispensable role in sustaining ecological balance. Among the various pressures affecting forests, forest fires constitute a major environmental problem, posing significant threats to natural ecosystems while disrupting forestry activities and leading to substantial economic losses. In Türkiye, forest fires are predominantly observed in the Aegean, Mediterranean, and Marmara regions. This study presents a proactive, explainable GeoAI framework for wildfire risk analysis by integrating multi-source geospatial data. The methodology calculates a physical hazard index by evaluating advanced ensemble machine learning algorithm, specifically extreme gradient boosting (XGBoost) with the latter demonstrating superior predictive performance. The model is driven by topographic, meteorological, biotic and anthropogenic parameters. In parallel, a comprehensive vulnerability matrix encompassing physical, social, economic and environmental dimensions is constructed utilizing the Best-Worst Method (BWM). These two main components are spatially synthesized in a Geographic Information Systems (GIS) environment. The model's interpretability is enhanced focusing on feature importance to identify the primary drivers of fire ignition and spread. Results demonstrate that anthropogenic pressures and Wildland-Urban Interface (WUI) dynamics are the most significant risk determinants. The resulting high-resolution maps provide a robust Spatial Decision Support System (SDSS) for local authorities, enabling targeted mitigation strategies and contributing to the development of sustainable, smart and resilient cities.
Since 2022, the Kinburn Peninsula, Ukraine, has been located within an active war zone, where frequent and recurrent wildfires have affected heterogeneous coastal sandy ecosystems. Burned-area mapping in this landscape is complicated by wetlands, salt flats, exposed sands, reedbeds, artificial pine plantations, cloud cover, smoke and restricted field access. This study applies an integrated multi-sensor approach to map wildfire damage and recurrence during 2022–2025 using Sentinel-1 SAR, Sentinel-2 and Landsat 8/9 optical imagery, NASA FIRMS thermal anomalies and official fire-monitoring materials. Optical burn indicators were used to identify spectral evidence of burning, while Sentinel-1 VV/VH backscatter changes and Radar Forest Degradation Index (RFDI) were used to detect structural vegetation degradation. Optical burn candidate masks and SAR-derived structural damage masks were spatially overlaid to delineate confirmed fire-damage areas. The results show that wildfire impact was spatially recurrent rather than isolated. SAR-derived candidate areas increased from 637.06 ha in 2022 to 4297.97 ha in 2025, while mean ΔVH remained consistently negative, indicating reduced vegetation volume scattering. The largest confirmed optical–SAR fire-damage area was detected in 2025, reaching 598.26 ha. Fire recurrence mapping revealed zones of accumulated pyrogenic stress, particularly within artificial forest plantations, psammophytic steppe complexes and protected coastal ecosystems. The proposed optical–SAR framework improves wildfire damage mapping under wartime and data-limited conditions.
Accurate and spatially explicit mapping of forest fuel types is essential for developing integrated wildfire risk management strategies, estimating fire danger, and modeling fire propagation in fire-prone Mediterranean ecosystems. While time series of Copernicus Sentinel-2 multispectral imagery at 10 m spatial resolution have proven effective for regional fuel mapping, severe spectral and spatial ambiguities persist in distinguishing complex canopy structures, heterogeneous shrublands, and pastures. This study explores the synergistic integration of Sentinel-2 time series and high-resolution PlanetScope satellite acquisitions (3 m spatial resolution) to produce a 10 m fuel type map (9 classes) over a pilot area in Sardinia, Italy. A total of 114 features were synthesized, combining dynamic Sentinel-2 metrics (kNDVI, FAPAR, NDMI, ReNDVI harmonics), PlanetScope seasonal kNDVI zonal statistics and Gray-Level Co-occurrence Matrix (GLCM) textural features, and static terrain/canopy parameters. Using an XGBoost machine learning classifier trained on 245 photo-interpreted Regions of Interest (80/20 spatial split), three sets of features were compared. The combined model achieved a test overall accuracy of 74.46%, compared to 74.05% for Sentinel-2 alone and 74.66% for PlanetScope alone. Notably, incorporating PlanetScope textural features significantly improved classification metrics for dense canopy classes, increasing the F1-score of broadleaf evergreen trees and needleleaf evergreen trees. The PlanetScope GLCM kNDVI autumn entropy emerged as the most important feature overall. Conversely, PlanetScope performed poorly on the grass class. These findings highlight the critical trade-offs and high complementarity between fine spatial texture and broad spectral coverage for operational wildfire management.
When a satellite observes the Earth at an angle, the apparent position of elevated features, such as mountains, clouds or fires in high terrain, can be shifted from their true ground location. This effect, which may be particularly pronounced in the case of geostationary platforms providing observations under systematically high viewing angles in many places, is known as parallax displacement and becomes larger as terrain elevation and satellite viewing angle increase.Meteosat satellites operated by EUMETSAT provide frequent near-real-time observations that are essential for wildfire detection and monitoring. The new Meteosat Third Generation (MTG) satellites, equipped with the Flexible Combined Imager (FCI), offer improved spatial, temporal and spectral capabilities compared with previous systems. These improvements support better detection and tracking of active fires.The LSA SAF Fire Radiative Power product (FRP-PIXEL) provides information on the location, timing and radiative energy of active fires, which can be used as an indicator of fire intensity and biomass consumption. However, standard geostationary processing assumes that observed features are located at sea level. As a result, fires occurring in elevated terrain may be incorrectly geolocated due to parallax displacement.This work presents an updated MTG FRP-PIXEL product that includes both the original fire and terrain-corrected coordinates. The correction aims to improve the spatial accuracy of active-fire locations, especially in mountainous regions. Preliminary comparisons with active-fire information derived from high-resolution Sentinel-2 imagery show that the correction can improve the location of FRP detections over high-elevation terrain.
Fire detection and monitoring are essential for mitigating the environmental and societal impacts of wildfire events. Earth Observation (EO) satellites have long played a central role in this domain, with polar-orbiting sensors such as MODIS and VIIRS providing global fire detections twice daily. However, their limited revisit frequency constrains the timeliness of fire alerts. Geostationary satellites, by contrast, offer high temporal resolution through multiple scans per hour, enabling near-real-time fire detection and rapid alerting. Despite this advantage, fire detection algorithms applied to geostationary imagery have predominantly relied on single-timestamp processing, limiting their ability to exploit the full temporal frequency of these observations. In this paper, we propose a multi-temporal approach for active fire detection using data from the Meteosat Third Generation (MTG) Flexible Combined Imager (FCI). The method first applies multi-temporal cloud masking based on the 10.5 μm thermal infrared band, before implementing a multi-temporal fire detection algorithm inspired by the Robust Satellite Techniques (RST) methodology of (Filizzola et al., 2017). To improve detection reliability, a pixel is confirmed as an active fire only after repeated detection across successive timestamps. Compared to the operational single-timestamp MTG FRP-pixel product, the multi-temporal algorithm achieves higher recall with a slight reduction in precision. Additionally, the method demonstrates sensitivity to regional land-use characteristics, excluding persistent industrial heat sources and continuous gas flaring from fire detections. Finally, the algorithm retrieves Fire Radiative Power (FRP) estimates with associated uncertainty bounds, yielding an average uncertainty window of approximately 4.5 MW.
