Deep learning-based burned area mapping of California wildfires using Sentinel-2 and Landsat-8 imagery enhanced with super-resolution techniques

Abstract

The increasing frequency of wildfires under a changing climate has led to extensive ecosystem destruction, highlighting the need for reliable burned area assessment using satellite imagery. Single-satellite data are constrained by observation gaps and interference from smoke and clouds, whereas multi-satellite data fusion can mitigate these limitations. Nonetheless, the fusion techniques still encounter challenges such as spatial information loss from resolution differences and cross-satellite domain mismatch. This study presents a burned area mapping framework that integrates super-resolution (SR) with transfer learning to address spatial and domain gaps in multi-satellite data. Specifically, Landsat-8 imagery is super-resolved to 7.5 m resolution, and the pretrained model is subsequently fine-tuned using Sentinel-2 imagery. Additionally, we investigated spectral index expansion to improve detection performance. Applied to California wildfires from 2013 to 2023, the input band expansion achieved a Burned Intersection over Union (IoU) of 0.821 when multiple spectral indices were incorporated. The SR-based transfer learning approach further improved performance with a Burned IoU of 0.850, representing a 4.55 % enhancement over the original Landsat-8 pre-trained model (0.813). Validation on 2025 wildfires demonstrated consistent performance and model robustness. These results highlight that the proposed SR-based approach offers benefits beyond simple resolution enhancement in multi-satellite data fusion. Furthermore, they confirm that spectral index-based input expansion contributes to accuracy improvements. The proposed framework is expected to support disaster response and environmental monitoring applications.