Ljiljana Šerić

After ignition, fire behavior is governed by a complex interaction of fuel, weather and topography. In practice, fire behavior is labeled as wind, topography or fuel-driven depending on the dominant driver. These labels are typically assigned through expert judgment and post-fire analysis rather than real-time classification. In this paper, we propose a model based on the ConvNeXtV2 convolutional neural network (CNN), capable of detecting the dominant wildfire driver directly from high-resolution surveillance images. The model is trained using 16 labeled sequences with targeted augmentation, class-balanced sampling and loss weighting to counter strong class imbalance and heavy repetition within sequences. The network shows good results in both fuel- and wind-driven situations but struggles with topography-driven wildfires, even when trained on full-resolution images. Despite the small dataset, our approach illustrates how modern CNNs can complement expert predictions.

Eruptive fire behavior is one of the most hazardous forms of extreme wildfire dynamics and has caused numerous firefighter fatalities. Early identification of high-risk conditions and locations is essential for improving operational safety. The Croatian Advanced Wildfire Surveillance System includes an Eruptive Fire Early Warning System that provides real-time assessment of eruptive fire potential. While earlier versions classified risk based on slope, aspect, meteorological conditions, and vegetation, the new version introduces an additional overlay layer incorporating canyon geometry as a key factor in accelerating fire spread. The proposed operational approach integrates experimentally derived canyon geometry thresholds with terrain analysis from Digital Elevation Model (DEM)-derived GIS layers in order to identify locations with increased eruptive fire potential.

Wildfire behavior can roughly be described as wind-, fuel-, or topography-driven, but these labels usually rest on expert judgment and post-fire analysis rather than on simple, reusable rules. Here, we build a small, interpretable classifier that predicts the dominant fire behavior driver: fuel-driven (plume/convection dominated), wind-driven, or topography-driven, from basic environmental information. The classifier is built using ERA5 10 m wind and relative humidity, summary elevation metrics, and fuel descriptors for a 14-event dataset of coastal Croatian wildfires. We compare the performance of multinomial logistic regression, random forest, and decision tree-based classifiers, focusing on agreement between their coefficients, feature importances, and splits rather than on formal optimization. All three models converge on a simple common rule set. Relative humidity, mean elevation, and elevation range emerge as the main axes of variation, consistent with basic fire behavior physics and published fire type schemes. Despite the small dataset, the classifier formalizes expert intuition in a transparent way and offers a template for scaling larger datasets, where it could evolve into a quick diagnostic of the dominant spread driver for ongoing fires.

Satellite remote sensing increasingly supports water quality monitoring, yet the temporal transferability of machine learning (ML) models remains insufficiently tested, particularly in coastal shallow lakes subject to hydrological variability. This study evaluates the predictive robustness of satellite-based ML models for electrical conductivity (EC), turbidity (TUR), water temperature (WT), and dissolved oxygen (DO) in Vrana Lake, Croatia. A total of 409 in situ measurements collected during 2023–2024 and 2025 were paired with Sentinel-2 and Landsat 8–9 imagery. Pearson, Spearman, and Kendall correlation analyses were applied for parameter-specific band selection using original, inverse, quadratic, and logarithmic feature transformations. Seventeen regression algorithms were evaluated under six training–testing split strategies, including strict temporal projection. WT exhibited high robustness (R² ≈ 0.90 under temporal projection) due to its strong dependence on thermal bands, while DO achieved moderate temporal stability (R² = 0.51) using log-transformed predictors. EC and TUR demonstrated substantial performance degradation under temporal separation (R² = 0.14 and −4.62, respectively), reflecting sensitivity to distribution shifts. For parameters showing sufficient stability, interpretable band-based retrieval equations were derived using the most strongly correlated spectral predictors. These findings highlight the importance of temporally structured validation and demonstrate that model complexity does not guarantee operational robustness in shallow, dynamically evolving lake systems.

Coastal lakes require monitoring approaches that capture spatial and temporal variability beyond the limits of conventional in situ measurements. In this study, a SIGMaL framework (Satellite–In situ–GIS-multicriteria decision analysis (MCDA)–Machine Learning (ML)) was developed, a unified methodology that integrates in situ monitoring, GIS MCDA-derived water quality index (WQI), satellite imagery, and ML models for comprehensive coastal lake water quality assessment. A WQI, derived from a 12-month series of in situ measurements and environmental parameters, was used alongside four physicochemical parameters measured by a multiparameter probe. First, satellite reflectance from each sensor was used to train a set of nine regression models for modelling electrical conductivity (EC), turbidity, water temperature (WT), and dissolved oxygen (DO). Second, convolutional neural networks (CNNs) with spectral and temporal inputs were trained to classify WQI classes, enabling a cross-sensor evaluation of their suitability for lake water quality monitoring. Third, the trained CNNs were applied to generate WQI maps for a subsequent 12-month period without in situ data. Across all analyses, WQI-based models provided more stable and accurate models than those trained on raw parameters. Sentinel-2 achieved the most consistent WQI performance (AUC ≈ 1.00, R² ≈ 0.84), PlanetScope captured fine-scale spatial detail (R² ≈ 0.77), while Landsat 8–9 was most effective for WT but less reliable for multi-class WQI discrimination. Sentinel-2 is recommended as the primary satellite sensor for WQI mapping within the SIGMaL framework. These findings demonstrate the advantages of WQI-based modelling and highlight the potential of ML–remote sensing integration to support coastal lake water quality monitoring.

Recent advances in Deep Learning (DL) have enabled the integration of diverse biomedical data for disease prediction and risk stratification. Building on this progress, this study proposes a robust framework for Multi Label Classification (MLC) of comorbidities in patients with Obstructive Sleep Apnea (OSA) using diverse physiological and clinical data sources. A one-dimensional Convolutional Neural Network (1D-CNN) was developed to process and fuse time series signals, oxygen saturation (SpO2), derived SpO2 features, and nasal airflow (FP0), with demographic and physiological parameters, enabling the identification of key comorbidities such as arterial hypertension, diabetes mellitus, and asthma/COPD. The proposed model was trained and fine tuned using the Optuna hyperparameter optimization framework, addressing class imbalance through weighted loss adjustments. Its performance was comprehensively assessed using multi label evaluation metrics, including macro/micro F1-score, AUC-ROC, AUC-PR, subset and partial accuracy, Hamming loss, and Multi Label Confusion Matrix (MLCM) . The 1D-CNN achieved superior predictive performance compared to traditional Machine Learning (ML) classifiers with macro AUC-ROC = 0.7305 and AUC-PR = 0.7497. The model demonstrated consistent behavior across age, gender, and BMI groups, indicating strong generalization and minimal demographic bias. The results confirm that SpO2 and airflow signals inherently encode comorbidity specific physiological patterns, enabling efficient and scalable screening of OSA related comorbidities without the need for full polysomnography. Although the study is limited by data set size, it provides a methodological basis for the application of multi label DL models in clinical decision support systems. Future research should focus on the expansion of multicenter datasets, thereby improving model interpretability and potential clinical adoption.

Fires represent a serious threat to humans, property, and the environment. Timely detection and effective extinguishing are essential to reduce harmful consequences. This paper focuses on the analysis of contemporary methods, technologies, datasets and approaches for fire detection, prevention and suppression. Fire detection methods are divided into those applicable to indoor and outdoor settings, with special emphasis on challenges and solutions in each category. The methods reviewed in this paper include those based on image analysis, which include advanced image processing and machine learning techniques, enabling systems to recognize visual signs of fire. Furthermore, a review of publicly available datasets was conducted to provide a better understanding of currently existing fire analysis and detection methods. Next, an evaluation of sensor-based fire detection approaches was carried out, focusing on their capability to identify a wide range of combustion products. Fire suppression methods, including traditional practices and contemporary technologies such as robotic and drone systems, are reviewed, with particular emphasis on extinguishing fires caused by lithium-ion batteries, which require a unique approach due to their specific characteristics. The goal of this paper is to outline the current state of research and existing technologies, identify their advantages and disadvantages, and highlight potential areas for further research and development.