A carregar...
Publicação
Development of a wildfire risk prediction system based on Deep Learning methods and Remote Sensing
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 26.33 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Wildfires pose a significant threat to ecosystems, human life, and infrastructure, particularly in South America, where diverse climatic and environmental factors contribute to their occurrence. Climate change has intensified extreme weather conditions, leading to an increase in wildfire frequency and severity. Countries such as Brazil have experienced escalating wildfire damage, highlighting the need for predictive models that can accurately assess fire risk and support mitigation efforts. This thesis develops a wildfire risk prediction system leveraging deep learning methods and remote sensing data to address this challenge. The objectives of this research are threefold: (1) explore how Earth Observation (EO) data cubes can be leveraged to efficiently process and analyze spatiotemporal data for wildfire danger prediction; (2) assess the suitability of various deep learning algorithms for detecting, assessing, and mitigating wildfire impacts using remote sensing data; and (3) examine how the integration of static and dynamic variables can enhance the accuracy of near real-time wildfire
risk predictions. Key research questions focus on the role of EO data cubes in improving spatiotemporal analysis, identifying the most effective deep learning models for wildfire risk assessment, and evaluating the benefits of combining static and dynamic variables for enhanced predictive performance. To tackle the stochastic nature of wildfire occurrences, this study evaluates the performance of Random Forest (RF), Long Short-Term Memory networks (LSTM), Convolutional Neural Networks (CNN), and Convolutional LSTM (ConvLSTM) models. The results indicate that LSTM is the most balanced model, achieving an AUROC of 89% and an F1-score of 72%, effectively capturing temporal dependencies while maintaining a trade-off between false positives and false negatives. CNN exhibited the highest sensitivity, with an 84% recall, though at
the cost of lower precision (67%). ConvLSTM did not outperform LSTM significantly, suggesting that its advantage depends on the scale and structure of the input data. The RF model achieved the highest AUROC (91%) but struggled with sensitivity. Feature importance analysis identified human modification (gHM), land surface temperature (LST), and temperature as the most influential variables, reflecting the combined impact of anthropogenic and environmental drivers on fire occurrence. Seasonal analysis further demonstrated the influence of temperature, vegetation, and precipitation conditions on fire activity. The findings highlight the effectiveness of integrating EO data cubes and deep learning for wildfire risk prediction. The LSTM model’s ability to balance spatiotemporal dependencies makes it a strong candidate for real-time fire forecasting. This study contributes to advancing scalable, data-driven wildfire prediction models, providing valuable insights for fire management strategies in South America.
Descrição
Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial Technologies
Palavras-chave
Wildfires wildfire risk prediction system