Abstract

Monitoring wildfires is essential to mitigating their widespread environmental, economic, and social impacts. Recent advances in remote sensing technology, combined with the growing use of artificial intelligence, have significantly enhanced the ability to perform real-time, high-resolution fire monitoring. Motivated by the limitations of traditional methods in dealing with spatial heterogeneity and class imbalance, this study proposes two scalable, AI-based approaches built on the U-Net convolutional neural network for automated burned-area segmentation using multispectral Landsat imagery. We present two model variants: the 128-Crop approach, which processes fixed-size image patches, and the AllSizes (AS) strategy, which uses variable-sized crops to improve contextual understanding and dataset balance. Both models are trained on time-series Landsat imagery from two fire-prone regions in Chile-Biobio and Valparaiso-using pre-and post-fire composites along with high-resolution fire scar labels. The training pipeline includes preprocessing, data augmentation, and hyperparameter optimization, employing Dice Loss to address class imbalance. A quantitative evaluation on 195 representative test images shows that the AS model outperforms the 128-Crop variant, achieving a Dice Coefficient (DC) of 0.93, an Omission Error (OE) of 0.086, and a Commission Error (CE) of 0.045, while the 128-Crop model reached DC = 0.86, OE = 0.12, and CE = 0.12. Additionally, a QGIS plugin named "FireScar-Mapper-Plugin" has been developed to enable user-friendly access and integration of the models. The plugin supports batch processing of new imagery and is designed for non-programmers, enhancing the framework's scalability and applicability for large-scale wildfire monitoring and management across diverse regions. These contributions-combining a novel scalable U-Net strategy with an open-source QGIS implementation-make this study a distinctive step toward practical, automated burned-area mapping.