An Empirical and Dynamic Approach to Forest Fire Spread Modelling Using Remote Sensing and Machine Learning Techniques

Singh Harikesh

doi:10.25907/00962

Wildfires pose a severe threat to ecosystems, biodiversity, and human settlements, particularly in fire-prone regions such as Australia. With the increasing frequency and intensity of wildfires due to climate change, there is a pressing need for advanced predictive models that integrate empirical data, remote sensing technologies, and machine learning algorithms to enhance fire risk assessment and mitigation strategies. This research presents an empirical and dynamic approach to wildfire susceptibility and spread modelling, utilising geospatial data, machine learning frameworks, and remote sensing techniques to improve prediction accuracy and facilitate proactive wildfire management. The study systematically evaluates traditional and modern methodologies for wildfire prediction, including empirical models, and deep learning approaches. A comparative analysis of machine learning techniques—such as Artificial Neural Networks (ANN), U-Net Convolutional Neural Networks (CNNs), and hybrid frameworks—was conducted to determine their effectiveness in wildfire susceptibility mapping and fire spread forecasting. The integration of Geographic Information Systems (GIS) and remote sensing data enabled the incorporation of diverse environmental variables, including vegetation indices, topography, fuel accumulation, weather parameters, and historical fire records. The study employed Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 multispectral data to analyse fire-prone regions and post-fire vegetation recovery, demonstrating the effectiveness of Earth observation data in wildfire research. A key focus of this research is the application of hybrid modelling approaches that combine empirical fire behaviour models with machine learning-based spatial prediction techniques. The study explores the integration of Rothermel’s fire spread model with U-Net CNNs, utilising geospatial datasets to optimise fire spread simulation. The results indicate that deep learning models, particularly U-Net architectures, outperform traditional ANN-based approaches in terms of spatial accuracy and predictive reliability. The results indicate that deep learning models, particularly U-Net architectures, outperform traditional ANN-based approaches. Specifically, the ANN model achieved an AUC of 0.98, while the U-Net CNN model achieved an AUC of 0.87, demonstrating strong predictive performance. Validation of the hybrid fire spread model against historical fire perimeters achieved a Dice Similarity Index (DSI) of 0.82, confirming its reliability. The study also emphasises the role of cellular automata (CA) models in simulating fire propagation patterns under varying environmental conditions. By incorporating meteorological parameters such as wind speed, temperature, and humidity, the CA-based simulations provide a realistic assessment of fire spread dynamics across different landscape types. Furthermore, this research contributes to the post-fire vegetation regrowth assessment, utilising Sentinel-1 SAR data and Google Earth Engine (GEE) to track burned areas and monitor ecosystem recovery over time. A case study in Mooloolah River National Park, Queensland, Australia, demonstrates the applicability of SAR-based indices such as the Normalised Difference Fire Index (NDFI) and Differenced Normalised Burn Ratio (dNBR) in characterising fire severity and regrowth patterns. The findings reveal that SAR backscatter data, when combined with clustering algorithms, can effectively differentiate burned and regenerating vegetation, providing valuable insights for ecological restoration and fire management strategies. Additionally, the study explores the role of Indigenous fire management practices in mitigating wildfire risks. Drawing from the research on firestick farming, this study underscores the importance of integrating traditional ecological knowledge with modern fire prediction technologies. Indigenous-controlled burns, historically used to manage fuel loads and maintain ecological balance, offer valuable insights that can be incorporated into contemporary wildfire risk assessment models. The results of this study have significant implications for sustainable wildfire management, emphasising the need for multi-source data integration, interdisciplinary modelling frameworks, and adaptive machine learning techniques. The findings suggest that hybrid fire modelling approaches—combining empirical fire behaviour models, machine learning algorithms, and high-resolution remote sensing data—can substantially enhance the accuracy of wildfire susceptibility assessments and fire spread simulations. Moreover, the study highlights the potential of real-time wildfire monitoring systems, utilising satellite observations and cloud computing platforms such as Google Earth Engine for near-instantaneous fire detection and response. In conclusion, this research advances the field of wildfire modelling by bridging empirical methodologies with data-driven machine learning techniques. These results have direct implications for practice, providing fire managers and policymakers with a robust decision-support tool for targeted prescribed burning, resource allocation, and ecological restoration planning. The proposed approaches offer a scalable and adaptable framework for fire prediction, spread analysis, and post-fire recovery assessment, with applications extending beyond Australia to other wildfire-prone regions globally. Future research should focus on refining hybrid models by incorporating real-time meteorological data, fuel moisture indices, and drone-based observations to further enhance the predictive capabilities of wildfire management systems. By integrating remote sensing, GIS-based modelling, and deep learning techniques, this study provides a robust foundation for data-driven decision-making in wildfire risk reduction and ecological resilience planning.

An Empirical and Dynamic Approach to Forest Fire Spread Modelling Using Remote Sensing and Machine Learning Techniques

Files and links (1)

Abstract

Details

Metrics