Search

The destructive potential of wildfires has been exacerbated by climate change, causing their frequencies and intensities to continuously increase globally. In this context, increasing wildfire activity across the globe has become an urgent issue with enormous ecological and social impacts. Wildfires have consumed important areas and forest resources, as a result, fire management expenditures have increased and thousands of homes and many lives have been lost. Moreover, they have significantly impacted biodiversity and greenhouse gas emissions on a global scale.The current incidents across the globe highlight the need for preemptive policy measures to reduce the risk of fire occurrence, managing the land in an effective way to protect natural forests, agricultural areas, and human lives. These concepts are included in what is known as FireSmart Forest Management (FSFM). This paradigm considers opportunities in three dimensions: i) decrease of the fire behavior potential of the landscape, ii) reduction of the potential for fire ignitions, and iii) increase in the fire suppression capability.This dissertation aims at advancing the theory, practice, and large-scale implementation of complex data-driven decision making and machine learning models in the context of landscape management under wildfire risk, integrating Operations Research, Computer Science, and Data Science techniques. We focus our efforts on the understanding, evaluation, and development of effective prevention and mitigation policies, with the potential of being implemented practice, as well as exploring and developing new FSFM techniques.We divided our study into three main aspects: Simulation, Decision-Making, and Machine Learning. In Chapter 1, we focus on the development and evaluation of an accurate, flexible, and efficient wildfire simulation model that can be integrated with data-driven decision-making models. Empirical results on thousands of simulations show the high performance of the model compared to existing solutions, highlighting its accuracy with real-life instances. We then focus our efforts on its generalization in Chapter 2, seeking to adjust its main parameters to mimic the fire spread behavior observed in different regions of the world where no empirical models are available. This, exploiting historical information for training purposes using derivative-free optimization techniques to adjust the parameters of the model, allowing us to capture current wildfire dynamics. Experiments performed on datasets located in different regions of the world show the potential of the proposed method.Second, in Chapters 3 and 4, we explore the integration of this model with landscape planning decision-making models to derive robust fuel treatment policies to mitigate expected losses due to future wildfire events, generating fire resilient landscapes. We study and compare complex network algorithms to develop a mathematical model denoted Downstream Protection Value, capturing the importance of the different components of the land to provide a natural prioritization of where mitigation actions should be implemented. An optimization framework incorporating multiple variables to analyze the inherent trade-offs involved in the planning process is developed, providing practitioners and researchers with an open-source decision support system implementation involving multiple and potentially, opposing objectives. We evaluate the performance of the proposed mathematical model compared to existing solutions, highlighting its superior performance with thousands of experiments involving uncertainty for landscapes located in North America. Several extensions are discussed providing future research directions in the field. We further expand this framework in Chapter 5, incorporating wildfire suppression strategies derived from a novel multi-agent decision-making model. In this application, a group of agents is deployed to the field once an ignition or active fire is detected with the aim of containing or stopping it as soon as possible. The sequential and temporal dimensions of the problem become a challenge to apply traditional modeling techniques. We develop a deep reinforcement learning algorithm focused on exploiting the collaboration and coordination between independent agents. Extensive computational results demonstrate the impact of including local rewards and the concept of sub-groups of agents in the context of centralized training and decentralized execution algorithms, leading to more complex and effective collaboration strategies between agents belonging to the same group and the environment in general.As part of our decision support system extensions, we build end-to-end machine learning models to understand the wildfire phenomenon from a large-scale perspective. For this, in Chapter 6, we explore the impact of different landscape compositions on the risk of wildfire ignition using remote sensing data to support challenging landscape planning decisions. Using a custom convolutional neural network model integrated with state-of-the-art visualization techniques, we highlight the main areas of interest for the deep learning model, focusing our efforts on the interpretability of the results. This, with the aim of opening the artificial intelligence black-box to fully understand the rationale behind the results and the different risk levels associated with characteristic spatial patterns observed in the land. We validate our results with previous studies using similar datasets, noting how the proposed model significantly surpasses their predictive performance while providing insights about the learning process of the model. Finally, in Chapter 7, we develop a global study of the main characteristics and drivers of wildfire regimes consolidating observations for almost two decades of wildfires. We classify and delineate regions of the world sharing similar fire activity as well as identifying their driving factors to support national or regional wildfire prevention/mitigation policies using a variety of machine learning techniques. To the best of our knowledge, this represents the first study that defines fire regimes spatially at a global scale bridging existing knowledge gaps between global and regional fire studies.Our results represent an attempt at improving the integration of multiple disciplines in the context of effective data-driven decision-making under natural hazards uncertainty. Several challenges are still open. We hope that this research can serve as a motivation to expand the field's perspective with high-impact applied projects involving mathematical, ecological, economic, data, and social sciences.

In practical applications, it is common that wildfire simulators do not correctly predict the evolution of the fire scar. Usually, this is caused due to multiple factors including inaccuracy in the input data such as land cover classification, moisture, improperly represented local winds, cumulative errors in the fire growth simulation model, high level of discontinuity/heterogeneity within the landscape, among many others. Therefore in practice, it is necessary to adjust the propagation of the fire to obtain better results, either to support suppression activities or to improve the performance of the simulator considering new default parameters for future events, best representing the current fire spread growth phenomenon. In this article, we address this problem through a new methodology using Derivative-Free Optimization (DFO) algorithms for adjusting the Rate of Spread (ROS) factors in a fire simulation growth model called Cell2Fire. To achieve this, we solve an error minimization optimization problem that captures the difference between the simulated and observed fire, which involves the evaluation of the simulator output in each iteration as part of a DFO framework, allowing us to find the best possible factors for each fuel present on the landscape. Numerical results for different objective functions are shown and discussed, including a performance comparison of alternative DFO algorithms., 8 figures, 35 pages