Two Scalable Approaches for Burned-Area Mapping Using U-Net and Landsat Imagery

• DOI: 10.48550/arxiv.2311.17368
• Publication Date: 2023-01-01
• Authors: Ian Mancilla-Wulff; Jaime Carrasco; Cristóbal Pais; Alejandro Miranda; Andrés Weintraub
• Keywords: Hyperparameter

Monitoring wildfires is an essential step in minimizing their impact on the planet, understanding many negative environmental, economic, and social consequences. Recent advances remote sensing technology combined with increasing application of artificial intelligence methods have improved real-time, high-resolution fire monitoring. This study explores two proposed approaches based U-Net model for automating optimizing burned-area mapping process. Denoted 128 AllSizes (AS), they are trained datasets a different class balance by cropping input images to sizes. They then applied Landsat imagery time-series data from fire-prone regions Chile. The results obtained after enhancement performance hyperparameter optimization demonstrate effectiveness both approaches. Tests 195 representative area show that dataset using AS yields better performance. More specifically, exhibited Dice Coefficient (DC) 0.93, Omission Error (OE) 0.086, Commission (CE) 0.045, while achieved DC 0.86, OE 0.12, CE 0.12. These findings should provide basis further development scalable automatic tools.