Henrique Moura, imec University of Antwerp
Phil Reiter, imec University of Antwerp
Phil Reiter, Senior Researcher, imec University of Antwerp
Wildfires are destructive natural disasters that severely impact ecosystems, wildlife, and air quality. They emit large quantities of greenhouse gases, thus accelerating climate change. In the U.S., wildfires have become more frequent and severe, costing billions in property damage, infrastructure losses, and firefighting expenses. In 2022, 7.2 million acres caught fire costing around 2.5 billion dollars for Federal Firefighting, i.e., direct firefighting costs. These costs exceeded $89 billion in lost economic output for the U.S., in 2024 alone, and will cost more than 466, 000 jobs. Europe also faces rising wildfire threats due to hotter, drier summers, with an estimated 30,000 wildfires yearly, in particular in southern regions like Greece, Portugal, Spain, and Italy. Early detection of wildfires can mitigate the environmental, economic, and health impacts.
Satellites play a crucial role in quick detection, providing broader, continuous monitoring compared to traditional methods like ground-based sensors and aerial patrols, which are labor-intensive and limited in range. Satellite systems, including NASA’s MODIS and ESA’s Sentinel-2, detect fires through visible and infrared light, allowing observation of fire hotspots even through thick smoke. These systems can operate continuously across day and night, and cover vast areas, enabling authorities to manage evacuation and health advisories promptly. Deep learning in computer vision has revolutionized object detection and image segmentation, enabling precise partitioning of images into meaningful regions or objects with superhuman capabilities. These models excel at learning complex patterns and features from images, even in challenging scenarios like occlusions or varying lighting conditions. Their ability to generalize across diverse domains makes them a cornerstone of modern computer vision systems. Thus, we argue that satellite-based deep-learning models can provide an advanced tool for wildfire detection. These models can even run onboard the satellites, identifying thermal anomalies in large-scale images. However, the success of these models is challenged by the imbalanced nature of satellite data – where fire spots cover a small fraction of an image relative to the background – making training a model a hard process. The imbalance creates difficulties in accurately detecting wildfires, with models often misclassifying or missing fire pixels.
To address this, we propose balancing datasets using data-level techniques (e.g., resampling) and algorithm-level techniques (e.g., adjusting the loss function, performing class weight adjustment, or adding attention blocks to the neural network). We experimented with random oversampling methods, i.e., increasing the representation of minority classes by duplicating existing instances. But also use class weight adjustment to make the model more sensitive to underrepresented classes (fire-affected pixels). In this paper, we show that data augmentation is important for training a deep learning model, and with data imbalance using oversampling also helps obtain better results. We also tested the importance of the frequency bands used to train the model. Our results show that IR bands must be considered for fire detection from satellite images since they have greater thermal sensitivity and can propagate through smoke. Loss function optimization is critical for improving model accuracy on imbalanced datasets, allowing control over the trade-offs between false positives and negatives. Also, attention mechanisms can be crucial as they focus on learning the most relevant features, allowing models to prioritize minority class instances without requiring significant data augmentation or resampling. The proposed approach includes a workflow pipeline for wildfire detection that adjusts for dataset imbalances and uses specialized loss functions and attention modules, enhancing detection reliability in satellite imagery. We experimented with different attention blocks and loss functions. Our results show that using an attention block and an appropriate loss function helps improve the model’s convergence time and the final performance in terms of the final F1 score while keeping the precision and recall of the model balanced. Our best model obtains an F1 score of 0.9786 on the test set, while the accuracy reached 0.999999 on a dataset with less than 0.5% positive cases. The combination of techniques used in the paper holds promise for advancing wildfire monitoring, enabling authorities to act swiftly in reducing wildfire impacts and protecting at-risk regions and populations.