Conference Agenda

Early Wildfire detection using AI remains challenging in environmental monitoring, particularly when the approach should be flexible enough to handle different sensors and different landscapes. This study presents a multi-stage deep learning framework for real-time smoke and fire detection using imagery from fixed tower cameras and UAVs. The proposed system employs YOLOv11 as the primary detection model for high-speed inference, com-plemented by Faster R-CNN for precision benchmarking and cross-architecture evaluation. Together, these models support an in-depth analysis of detection accuracy, robust-ness, and computational efficiency across diverse envi-ronmental conditions.

An end-to-end pipeline has been developed, integrating real-time image acquisition, asynchronous message han-dling through RabbitMQ, and performance logging via InfluxDB, enabling continuous model evaluation under near-operational conditions. Experimental results indicate that while YOLOv11 achieves high frame-rate perfor-mance and strong detection capability, it remains suscepti-ble to false positives in visually ambiguous scenarios such as haze, fog, or low-contrast backgrounds, where contex-tual patterns closely resemble smoke. Faster R-CNN serves as a complimentary reference to quantify localiza-tion accuracy and analyse error propagation, facilitating threshold tuning and model interpretability.

The presented framework bridges the gap between aca-demic model development and field-deployable fire sur-veillance systems. It establishes a reproducible, scalable foundation for real-time decision support in forest watch-tower networks and autonomous UAV missions aimed at early wildfire detection and response

Wildfires induce significant changes in forest structure and the surface reflectance characteristics. This study evaluated the effectiveness of using airborne LiDAR Intensity data to delineate surface fire burn areas in wildfires. We extracted ground returns from both coniferous and deciduous forests and conducted qualitative assessment of Intensity through Intensity images, as well as statistical evaluation using the non-parametric Mann–Whitney U test to compare burned and unburned areas. We compared the median and standard deviation of Intensity at a 10-m mesh scale, calculating standard deviation at a finer 0.5-m mesh resolution. The results revealed significant differences between the two groups. As a result, a significant difference was observed between the two groups. The effect size r for the median in deciduous forests ranged from 0.55 to 0.84, while the effect size r for the standard deviation in coniferous forests ranged from 0.32 to 0.47. Both indicated a medium to large effect. These findings suggest that LiDAR Intensity can effectively identify surface fire burn areas even under heterogeneous forest floor conditions. The proposed method has the potential to contribute to enhancing post-fire monitoring using airborne LiDAR.

This study investigates two decade (2000-2020) of Aboveground Biomass dynamics in the biodiversity hotspot of Western Ghats, India, focusing on the impacts on forest fire and climate variability. Using machine learning approaches with GEDI LiDAR data and MODIS satellite imagery, we developed a robust annual AGB model. These analysis reveals a consistent decline in AGB across Kodaikanal and Nilgiris. Results shows that rising temperature and vapor pressure deficit are the key driver for increase in burn are and fire intensity. These are pushing carbon rich evergreen forests toward a critical transition from carbon sink to source. An integrated Structural Equation Model confirms that the dominant role of climate in driving fire regimes and subsequent biomass loss. This research provides a critical scientific foundation for fire adaptive forest management and carbon accounting in vulnerable tropical ecosystem.

Wildfires are longstanding natural phenomena with significant impacts on ecosystems and communities. In recent years, Canada has experienced particularly severe wildfire effects, especially in British Columbia (BC), which has endured prolonged and impactful wildfire events. However, there is currently no specialized wildfire time-series dataset for BC that considers long-term temporal sequences and multiple driving factors, which are essential for data-driven approaches. To facilitate future research on data-driven wildfire risk and spread prediction, we have developed a dataset covering the entire BC province, encompassing 683 wildfire events from 2002-2023 at 500m resolution with daily observations. For each wildfire event, the dataset includes 20 driving factors, including vegetation status, meteorological factors, human activities, topographical features, and active fire detection. Based on this benchmark and similar datasets from other regions, we compared multiple deep learning models, including CNN-based, Transformer-based, and Mamba-based architectures, to explore the effectiveness of existing deep learning models in wildfire risk prediction. We found that model F1 scores were below 0.6, indicating that this new dataset presents a challenging non-linear modeling scenario that requires more advanced and tailor-designed deep learning models to improve wildfire risk prediction accuracy.

Wildfires are known as one of the most important natural hazards, adversely impacting the ecosystems and human lives. Monitoring and management of wildfires is necessary to minimize their negative effects. Global BA products are widely used to study wildfires, but their accuracy is not constant over different environments. In this study, the MCD64A1 BA product was spatially validated using ground truth maps in a fire-prone Zagros Forest over 2021-2023. Our results indicated that its performance varies temporally, as the Kappa coefficient ranged from 0.04 to 0.69. Overall accuracy was higher than 0.96 percent in all years, indicating that MCD64A1 can be considered as a source for studying wildfires; however, its underestimation should be considered. In the next step, the trend of fire and its relationship with precipitation (i.e., obtained from the CHRIPS dataset) were analyzed in three forest ecosystems from 2001 to 2024. In two regions, Marivan and Kermanshah, wildfires experienced an increasing trend, in contrast to the other region, Shiraz, where they decreased over time. Analyzing the correlation between fire and precipitation revealed that spring precipitation is more connected to BA than annual precipitation. Comparing the results of the three regions showed that this matter is also region-related, and the results of one region cannot be referred to another. This study provided information on the performance of MCD64A1 in semiarid forests and the wildfire conditions in the Zagros Mountains to aid wildfire management.

Estimating carbon stock is important for understanding ecosystem dynamics and mitigating climate change. However, biomass mapping in boreal forests faces challenges due to harsh conditions and limited ground truth data for large scale studies. This study presents a parametric model for accurate biomass estimation in the Acadia and Taiga Forest using GEDI Level 4A, Sentinel-2, and SRTM DEM data. We integrated these datasets, and developed the parametric model consisting of spectral bands, vegetation indices, and topographic information with regression techniques, Random Forest and K-nearest neighbour. Results showcase performance of the parametric model with relative weights of variables for accurate Aboveground Biomass Density (AGBD) predictions for the two forest sites. With an average RMSE ranging between 9 Mg/ha to 31 Mg/ha and R^2 values of 0.54 to 0.60, the study reveals the importance of variables like slope, aspect and specific vegetation indices along with raw bands of Sentinel-2 data. Results also demonstrate potential and accuracy limitations of the proposed model with for biomass estimation with high resolution open-source satellite data without ground control. Further research include assessing the model robustness across diverse ecosystems and geographical settings, contributing to sustainable resource management practices.

Accurate estimation of aboveground biomass (AGB) is essential for quantifying forest carbon stocks, monitoring ecosystem change, and supporting greenhouse-gas reporting frameworks. While field measurements remain the benchmark, their limited spatial coverage has driven increasing reliance on remote sensing. Existing global AGB products such as ESA CCI Biomass and GEDI represent major advances but still suffer from signal saturation, sparse sampling, and limited ability to resolve fine-scale structural variation, highlighting an ongoing gap in effectively fusing information from different sensors. Recent studies combining Sentinel-1 (S1) SAR and Sentinel-2 (S2) multispectral imagery have demonstrated improvement in biomass modelling, with machine learning and deep learning achieving R² values from 0.57 to 0.80. However, most work focuses on tropical or broadleaf forests, leaving boreal mixedwood systems underrepresented, and models often rely on short-term composites that overlook multi-year temporal dynamics important for distinguishing long-term growth from seasonal phenology. This study addresses these gaps by developing a Transformer-based deep learning framework that integrates multi-temporal S1 and S2 time series and incorporates Growth & Yield (G&Y) variables as temporal constraints. By leveraging complementary radar–optical interactions and long-range temporal dependencies, the model is designed to reduce signal saturation, enhance structural sensitivity, and improve generalizability across heterogeneous boreal forest conditions.

Fuel Moisture Content (FMC) serves as a vital indicator for monitoring vegetation health and predicting wildfire risk. While existing approaches have largely emphasized temporal variations in FMC, they frequently overlook the inherent spatial clustering patterns of vegetation, leading to compromised spatial prediction accuracy. To overcome this limitation, we introduce a Transformer-based Spatio-Temporal Estimation Framework (TSTEF) that preserves sensitivity to temporal dynamics while incorporating spatial aggregation mechanisms to achieve robust and spatiotemporally consistent FMC estimates. The framework combines spatial autocorrelation analysis with gated recurrent unit (GRU)-based temporal modeling to effectively capture spatiotemporal dependencies, and utilizes Triangular Topology Aggregation Optimization (TTAO) for hyperparameter calibration. The proposed framework was validated using Sentinel-1/2 imagery and MODIS products in California, USA, where it demonstrated: (1) outstanding performance with an average R² > 0.8, MAE < 9%, and relative RMSE of 12.35%; (2) strong agreement between estimated FMC distributions and ground observations, with wildfire burned areas significantly expanding when FMC fell below the 120% critical threshold; and (3) excellent generalization ability during cross-regional validation, achieving relative RMSE values of 20.46% in France, 25.62% in Spain, and 20.76% in Colorado. This study provides a reliable analytical framework for wildfire risk early warning and contributes meaningful insights for ecosystem management amid global environmental change.

In the United States (U.S.), aggressive suppression policies in the 20th century reduced wildfires in the short term, but accumulated fuels contributed to increased wildfire risk in the long term. Land managers are slowly reintroducing the use of fuel treatments, including prescribed (Rx) fires, to remove fuels and mitigate future wildfires. To date, efforts to systematically quantify the efficacy of fuel treatments for wildfire mitigation have been limited in spatial and temporal scope or rely on proxies, such as low-intensity wildfires. Here we use the 30-m Harmonized Landsat and Sentinel-2 (HLS) dataset to analyze timeseries of vegetation “greenness” indices, such as the Normalized Burn Ratio (NBR), with observations up to every 2-3 days. We apply a causal inference approach, difference-in-differences (DID), on HLS-derived timeseries of NBR to compare outcomes in treated and surrounding control areas in three different time phases: 1) post-treatment fuel reduction, 2) wildfire-induced burn severity, and 3) post-wildfire vegetation recovery. As a case study, we targeted 37 Rx fires that preceded and intersected the 2024 Park Fire, a large wildfire in northern California. We show statistical evidence that Rx fires reduce fuel loads (12 Rx fires), wildfire burn severity (12 Rx fires), and post-wildfire vegetation recovery (14 Rx fires). Our approach requires only the spatial footprint and timing of the fuel treatments, thus enabling regional to nationwide analyses using a large number of fuel treatments to quantify the general efficacy of fuel treatments across a variety of treatment types, fuel types, and topography.

Ecosystem extent has been selected as an indicator of biodiversity under the Kunming-Montreal Global Biodiversity Framework. With wildfires on the rise in Canada and around the World there is a need to understand their impact on ecosystem extent for Framework reporting requirements. In Canada’s Arctic and Sub-arctic the growing season is much shorter than at lower latitudes, resulting in few cloud-free optical images and making it a challenge to monitor the impacts of fire and recovery at fine temporal resolutions. Synthetic Aperture Radar (SAR), particularly winter coherence, can be an indicator of ecosystem extent in the sub-arctic because burned areas will exhibit patterns that reflect more dynamic freeze-thaw cycles in the winter period than non-burned areas. Winter coherence and phase was calculated for 39 Sentinel 1 images over the time periods of 2017-2018, 2018-2019, 2019-2020, 2020-2021 in Northern Manitoba, Canada. When relating these values to known fire areas it was found that there is a distinct difference between areas recently burned as opposed to those burned further in the past. These data will be used as predictor variables in a Random Forest ecosystem classifier with outputs of overall accuracy and Shapley feature importance assessed.

Accurate and frequently updated land cover maps are vital for various scientific communities, as well as for public and regional authorities, supporting decision-making, planning, sustainable development, and natural resources management. Regular monitoring and mapping also play a crucial role for agricultural areas, particularly considering the projected population growth and shifting dietary patterns in many of the fastest growing regions of the world, that pose significant challenges for humanity. Over the past decade, the availability of Sentinel-1 and Sentinel-2 data has significantly increased the potential for high spatial resolution land cover mapping using dense time series. However, mapping croplands and distinguishing between crop types remains a more complex task, often requiring data of higher spatial, spectral, and temporal resolution. In this context, this study aims to evaluate the synergistic use of multi-temporal data from Sentinel-1, Sentinel-2, and EnMAP data for detailed land cover and crop type mapping in agriculural regions of western Greece.

Urban localization for autonomous driving requires accurate 6-DoF vehicle pose despite GNSS multipath, occlusions, and rapidly changing visibility. We fuse LiDAR, IMU, and GNSS in an error-state Kalman filter against a high-resolution (HR) map, aiming (i) to reduce LiDAR load without degrading accuracy and (ii) to improve robustness in building-sparse areas such as open junctions. The reference trajectory and HR map stem from a dedicated urban measurement campaign; Monte-Carlo simulations use ray-cast LiDAR, synthesized IMU, and GNSS tied to this trajectory so that only sensor noise is varied. A geometry-aware farthest-point sampling scheme prioritizes points informative for building/ground planes and pole-like structures, while an extended functional model introduces poles as additional vertical constraints. A retained-point rate of 10 % preserves trajectory-wide millimetrelevel and sub-milliradian accuracy, meeting in theory automotive requirements. Filter runtime is reduced by about 82 % relative

to the full LiDAR data. Compared with plane-only variants, the planes+poles configuration yields statistically significant but globally modest improvements in longitudinal, lateral, and yaw accuracy. More importantly, a sliding-window analysis reveals that it markedly stabilizes pose in plane-sparse junctions. Overall, the results suggest that task-aware subsampling preserves trajectory-wide performance while pole constraints add local robustness in challenging urban scenes; validation with real sensor logs remains necessary to confirm these accuracy margins, but the proposed filtering scheme shows promising potential for practical deployment.

Sensor data are increasingly used for monitoring and observation of spatiotemporal phenomena for diverse applications such as in flood management, urban traffic, air quality control, forest fire management, etc. Real time modelling and representation of such evolving phenomena is fundamental for efficient and timeliness decision-making processes. In the context of multisensory systems, where two phenomena (e.g.: air pollution index and windy condition) can both be sensed by networked sensors, analysing the relationship that hold between them is a major issue for decision making. Knowing if the pollution extent is expanding or contracting around a given spot or if it is within a windy zone can help in adopting more appropriate strategies. Sensing system equipped with rule-based reasoning engine to infer on spatiotemporal changes or topological relationship that holds between sensed phenomena with broad boundaries over time will provide decision-maker with adequate and non-ambiguous information. In this paper spatial changes and topological relationship about fuzzy-crisp object modelling the geometry of vague shape phenomena are conceptualized using an Extended Fuzzy Spatiotemporal Change Pattern (FESTCP) and a 5x5 Intersection model (I5x5M) respectively. The rule-based reasoning engine proposed in this paper is based on this conceptualisation. To evaluate our method, a simulated case study of air pollution in Quebec City is carried out. The results reveal that the proposed method captures well the spatiotemporal evolution of a given air pollution episode that served for an on-the-fly decision-making process in real life situations.

Reliable lane-level localization remains difficult for autonomous vehicles (AVs) when Global Navigation Satellite System (GNSS) observations are degraded by blockage, multipath, and non-line-of-sight reception in urban environments. This paper presents PointLoc, an Inertial Navigation System (INS)-centric locator aided by GNSS, monocular visual-inertial odometry (VIO), and High-Definition (HD) Vector Maps. The proposed method is formulated as an INS-centric error-state extended Kalman filter (EKF), in which the INS provides persistent state propagation, while GNSS, VIO, and map matching are incorporated as aiding updates according to their availability and reliability. This design preserves a unified position, velocity, and attitude solution and enables graceful degradation when some aiding sources become unavailable. The method is validated through real-vehicle experiments in Taichung Shuinan and Tainan Shalun under mixed GNSS conditions. The results show that PointLoc achieves the best overall full-route performance in Taichung Shuinan and remains broadly comparable to GNSS/INS/VIO, while still outperforming GNSS/INS, in Tainan Shalun. In the mapped GNSS-denied segment of Taichung Shuinan, PointLoc effectively suppresses vertical drift and substantially improves three-dimensional positioning. The mapped-road analysis further shows that the INS-centric design avoids the planar instability observed in a vision-centric benchmark and provides a more continuous localization solution.

Conventional laser scanning techniques (such as in a Terrestrial Laser Scanner or Mobile mapping), whether used in a static or mobile mode require the object of interest to remain stationary during the scanning stage. Any motion of the object during scanning results in the apparent distortions in the resulting point cloud. The authors in Goel and Lohani (2014b) proposed a motion correction technique to estimate the 3D geometry of a moving object, utilizing a fusion of inertial and GNSS (Global Navigation Satellite Systems) sensors and transformation of the resulting point cloud to an object body coordinate system (OBCS). This paper aims to carry out the experimental validation and performance analysis of the motion correction method. Field experiments are designed and conducted in three phases to verify the correctness of the method. Through this, the paper aims to uncover insights into the performance of the motion correction algorithm and provide the first experimental validation of the proposed technique.

Wearable sensors are essential for gait analysis outside of traditional laboratory environments. However, selection of the right sensor technology involves several trade-offs. Inertial Measurement Units (IMUs) offer high temporal resolution which are ideal for detecting gait events but they suffer from drift. Ultra-Wideband (UWB) provides stable spatial data, but are less precise for detecting event timing. This paper presents a comparative study of three distinct foot-mounted sensor methodologies for heel strike detection and cadence estimation: (1) IMU-Only approach, (2) UWB-Only approach, and (3) a multi-sensor IMU+UWB fusion approach. Each method is evaluated against a camera-based ground truth system using data from four subjects. Results show the IMU-Only method is inconsistent, with moderate event precision (Avg. F1: 0.798), temporal accuracy (Avg. MAE: 47.99 ms), and subject-dependent cadence accuracy (Avg. Acc: 89.59%). The UWB-Only method provides robust event detection (Avg. F1: 0.811) with similar temporal error (Avg. MAE: 49.0 ms) but is exceptionally accurate for cadence estimation (Avg. Acc: 96.94%). The IMU+UWB fusion approach achieves the highest event precision (Avg. Precision: 0.835) and the best temporal accuracy (Avg. MAE: 46.51 ms), while also maintaining robust cadence accuracy (Avg. Acc: 95.62%). In conclusion, while the UWB-Only method is ideal for cadence-only applications, the IMU+UWB fusion approach provides the best overall balance of high event precision, superior temporal accuracy, and reliable cadence estimation.

Accurately registering building footprints from heterogeneous datasets with LiDAR data remains a critical challenge in urban mapping and 3D reconstruction. The objective of this work is to register source data, defined as 2D cadastral vector footprints from structured, regularized, or manually-verified datasets to target building footprints derived from classified aerial LiDAR. LiDAR provides direct 3D information with precise footprint positioning and high spatial resolution, enabling a geometrically reliable representation of dense 3D structures. Conversely, source datasets are not always up-to-date, and may exhibit geometric distortions such as translational offsets, rotational deviations, or local deformations, yet they remain valuable due to their structured organization and metadata content. To enhance geometric fidelity while preserving semantic structure, we propose a practical framework for non-rigid polygon registration that adjusts the geometry of cadastral footprints toward LiDAR-derived targets. The framework consists of two core components: (1) establishing correspondences between source and target polygons, and (2) minimizing a robust distance function that governs the registration process. Three deformation models are introduced: a rigid model allowing translations only, a semi-rigid model allowing deformations while keeping the overall structure of source footprints, and a non-rigid model allowing rotations. We evaluate our method by aligning real cadastral datasets to aerial LiDAR data. The results confirm the effectiveness and robustness of the proposed framework in the context of 2D polygonal cadastral data. This work thus represents the first practical solution for non-rigid polygon registration in this domain.

This paper presents a mask-aware multi-stage depth enhancement framework for digital documentation of historical timber surfaces using RGB–Stereo-IR fusion. Accurate geometric recording of aged wood features such as wooden knots remains challenging due to uneven illumination and weak texture. The proposed pipeline, which aims to stabilise depth geometry under uneven illumination and low-texture surface conditions, integrates object detection, instance segmentation and confidence-guided depth refinement across three stages: (A) TV(total variation)-regularized mask-aware refinement, (B) confidence-weighted multi-view fusion, and (C) patch-based stereo reconstruction. Experiments on historical timber beams under varying illumination demonstrate improved depth completeness and geometric consistency, achieving a residual standard deviation below 0.6 mm in bright scenes and stable reconstruction in low-light conditions. The framework offers a practical solution for depth reconstruction of cultural heritage timber, supporting more reliable feature detection and analysis.