Repository of DUTH

Fire spreading prediction algorithm using hybrid quantum neural networks for vegetation classification (Master thesis)

Τσιόπλος, Νικόλαος/ Tsioplos, Nikolaos/ Quantum Computing and Quantum Technologies/ Κβαντική Υπολογιστική και Κβαντικές Τεχνολογίες

Wildfires represent one of the most destructive natural hazards, threatening ecosystems, infrastructure, and human life. Accurate fire-spread modeling remains a major challenge, requiring precise characterization of vegetation, terrain, and weather conditions. This thesis explores the integration of Hybrid Quantum-Classical Neural Networks (HQCNNs) and Classical Neural Networks (NNs) within a Cellular Automaton (CA) framework to simulate wildfire spread, using the 2018 Mati wildfire in Greece as a case study. A comprehensive dataset was developed by combining Sentinel-1 radar (VV, VH) and Sentinel-2 optical imagery (Red, Green, Blue, NIR) with derived vegetation indices such as NDVI, NDWI, BAEI, and BSI, along with elevation and meteorological data including wind speed and direction. Three network architectures—a Classical NN, an Estimator HQCNN, and a Sampler HQCNN—were trained under identical experimental conditions to classify vegetation and non-vegetation areas. Classical NN achieved the highest accuracy (99.2%) and balanced F1-scores, producing the most reliable results for downstream fire simulations. The Estimator HQCNN performed competitively, reaching 98.3% accuracy but showing sensitivity to circuit design and convergence stability, while the Sampler HQCNN underperformed (below 60% accuracy), reflecting the current limitations of quantum sam-pling. When integrated into the Cellular Automaton model, differences in classification accuracy and stability directly influenced the realism of fire propagation. Classical NN produced smooth burned-area perimeters closely matching the observed fire extent, the Es-timator HQCNN introduced slight over-propagation at the edges, and the Sampler HQCNN generated irregular spread patterns. These outcomes demonstrate how small variations in classification can significantly affect fire-spread dynamics. Overall, the results confirm that classical neural networks remain the most reliable solution for operational fire modeling, while hybrid quantum-classical methods, particularly estimator-based approaches, show promising potential for future environmental applications such as quantum hardware and optimization algorithms continue to advance.