Analysis and classification of fire data using machine learning methods

Fires represent one of the most catastrophic types of disasters that can have profound and far-reaching consequences for human life, infrastructure, public health, and the natural environment. As societies continue to evolve under the pressures of climate change, increasing population density, urban sprawl, and accelerating industrialization, fire-related incidents have not only become more frequent, but also more destructive, unpredictable, and complex in nature. The interconnection between anthropogenic activities and environmental changes has exacerbated fire risks, thereby demanding a paradigm shift from conventional fire response mechanisms to more proactive, data-driven, and intelligent systems for prevention, early detection, and rapid mitigation. In traditional fire management systems, decision-making is predominantly reactive, relying heavily on human expertise, past experiences, and manual assessments of risk. While such approaches have served well in the past, they are no longer sufficient in today's data-rich and dynamically evolving environments. The sheer volume, variety, and velocity of fire-related data necessitate more sophisticated analytical tools capable of processing complex datasets, extracting hidden patterns, and making timely, reliable predictions. In this context, machine learning (ML)—a subfield of artificial intelligence that focuses on algorithms capable of learning from data and improving over time—has emerged as a powerful solution for enhancing fire risk analysis and emergency response planning. This study seeks to develop and evaluate a machine learning-based classification framework that can accurately analyze and predict the severity of fire incidents using real-world fire occurrence data. Specifically, the research is centered on building supervised classification models that categorize fire events into three severity-based levels: "Initial," "Partial," and "Complete." The dataset employed in this research consists of 29,135 individual fire incident records obtained from a municipal fire department in a province of Turkey, covering a comprehensive nine-year period from 2015 to 2023. The dataset is rich in features and includes spatial variables (e.g., district, neighborhood), temporal variables (e.g., date, time of day, season), causal attributes (e.g., electrical faults, human negligence, unknown causes), and impact-related metrics (e.g., damage extent, number of casualties, distance to nearest fire station). The methodological approach adopted in this study followed a structured, multi-phase process, beginning with extensive data preprocessing. This stage involved identifying and addressing missing or anomalous values, converting categorical variables into numerical representations through encoding techniques, and applying normalization where appropriate to standardize feature scales. Following preprocessing, a feature selection phase was conducted using the Chi-Square statistical test. This method was employed to determine the degree of association between independent variables and the categorical target variable (fire severity level), thereby enabling the exclusion of less relevant or redundant features and improving both model interpretability and computational efficiency. Subsequently, the study implemented three widely recognized supervised learning algorithms to perform the classification task: Decision Tree (specifically, the Fine Tree variant), Gaussian Naive Bayes (GNB), and Linear Support Vector Machines (SVM). Each of these models offers distinct advantages and challenges in terms of computational complexity, interpretability, and classification performance. To ensure the robustness and generalizability of the results, a 5-fold cross-validation approach was employed during model training. This method involves dividing the dataset into five equally sized subsets and iteratively using four subsets for training and one for validation, thus minimizing the risk of overfitting and providing a more reliable assessment of each model's performance. The classification performance of the models was evaluated using a comprehensive set of metrics, including overall accuracy, precision, recall, F1-score, confusion matrix analysis, and the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve. Among the models tested, the Linear SVM classifier demonstrated the highest classification accuracy, achieving a rate of 61,4%, closely followed by the Decision Tree model at 61,2%. The GNB model, although slightly less accurate with a score of 58,5%, exhibited superior computational efficiency and fast inference times. The AUC analysis further reinforced the superior performance of the Linear SVM model, which attained an AUC value of 0,7511. Notably, the model was particularly effective at correctly identifying fire incidents within the "Partial" severity class, demonstrating strong discriminative power for this category. However, the performance of all models was less satisfactory when classifying "Complete" severity fires. This shortfall is likely attributed to an imbalanced class distribution in the dataset, where the number of "Complete" fire cases was significantly lower compared to the other categories. Additionally, overlapping feature characteristics between severity levels may have contributed to classification ambiguity. These limitations highlight the importance of addressing class imbalance through advanced sampling methods such as Synthetic Minority Over-sampling Technique (SMOTE) or by collecting more granular and representative data for underrepresented fire types. Beyond numerical performance, the interpretability of the models was also a critical consideration in this study. The Decision Tree model, for instance, offered easily understandable, rule-based decision pathways that can be directly interpreted and utilized by emergency planners, fire department personnel, and municipal decision-makers. On the other hand, while the GNB model did not achieve the highest accuracy, its simplicity and computational speed make it a favorable choice for deployment in environments with limited processing capabilities or real-time operational requirements. This research offers several original contributions to the existing body of literature on fire risk prediction and machine learning applications in disaster management. First, the study introduces a large-scale, real-world fire classification framework that integrates both urban and rural fire data—an important yet often overlooked dimension in current fire-related studies, which typically focus on forest fires or specific building types. Second, by systematically comparing three machine learning algorithms under uniform conditions, the research provides valuable insights into model selection for similar use cases. Third, the application of the Chi-Square feature selection method effectively demonstrates how careful attribute selection can enhance model performance and reduce unnecessary computational overhead. In addition to model development and evaluation, the study delves into exploratory data analysis to uncover meaningful temporal and spatial patterns within the fire dataset. The findings indicate that fire incidents tend to peak during the summer and autumn seasons, with industrial areas and densely populated urban neighborhoods showing higher fire frequencies. Furthermore, human-related causes such as electrical failures, careless behavior, and unsafe disposal of flammable materials emerged as predominant factors in fire ignition. These insights align with national and international fire statistics and reinforce the need for more targeted fire prevention strategies that consider both behavioral and environmental risk factors. One of the key lessons drawn from this research is the importance of high-quality, comprehensive datasets in machine learning-driven fire management. The challenges encountered due to class imbalance and feature overlap underscore the necessity for future studies to incorporate more diverse and balanced data sources. Enhancing the dataset with external variables—such as meteorological conditions (e.g., temperature, humidity, wind speed), land use data, topographical information, and remote sensing imagery—could significantly improve the granularity and predictive accuracy of classification models. Ultimately, this thesis illustrates that machine learning offers a powerful and scalable approach for enhancing fire risk assessment and supporting data-informed decision-making in emergency management. By integrating these models into operational workflows, local governments, fire brigades, and public safety agencies can proactively allocate resources, anticipate periods and regions of elevated fire risk, and implement timely interventions to reduce casualties and property loss. The findings of this study not only contribute to the academic discourse on artificial intelligence applications in public safety but also offer practical pathways toward the development of intelligent, adaptive fire management systems tailored to the evolving needs of modern societies.