Magnifying change: Rapid burn scar mapping with multi-resolution, multi-source satellite imagery

TL;DR

The paper tackles rapid burn scar delineation after wildfires by leveraging multi-resolution, multi-source satellite data. It extends the BAM-CD framework to BAM-MRCD, which jointly processes Sentinel-2 pre-fire imagery and MODIS bitemporal post-/pre-fire captures to produce high-resolution burn maps at 60 m GSD soon after fire activity ends, using a deep-supervision loss that combines a low-resolution and a high-resolution prediction. Across the FLOGA dataset and unseen events in Greece, Bulgaria, and Portugal, BAM-MRCD outperforms baselines and SR-CD methods, particularly for small and irregular burn scars, demonstrating robust cross-sensor fusion and practical potential for near-real-time disaster response. Limitations include difficulties with very small or elongated burns due to MODIS 500 m resolution and cloud coverage; nevertheless, the approach remains viable with VIIRS-era data and future high-frequency observations, enabling timely post-fire assessment and relief planning.

Abstract

Delineating wildfire affected areas using satellite imagery remains challenging due to irregular and spatially heterogeneous spectral changes across the electromagnetic spectrum. While recent deep learning approaches achieve high accuracy when high-resolution multispectral data are available, their applicability in operational settings, where a quick delineation of the burn scar shortly after a wildfire incident is required, is limited by the trade-off between spatial resolution and temporal revisit frequency of current satellite systems. To address this limitation, we propose a novel deep learning model, namely BAM-MRCD, which employs multi-resolution, multi-source satellite imagery (MODIS and Sentinel-2) for the timely production of detailed burnt area maps with high spatial and temporal resolution. Our model manages to detect even small scale wildfires with high accuracy, surpassing similar change detection models as well as solid baselines. All data and code are available in the GitHub repository: https://github.com/Orion-AI-Lab/BAM-MRCD.

Figures (13)

• Figure 1: Outline of our approach. A wildfire event starts at timestep $t_s$ and ends at $t_e$. Due to the inherent limitations of the satellite missions, we have daily MODIS captions over the area and sparser Sentinel-2 captions. The proposed method takes advantage of the first available MODIS imagery after the fire ends ($t_{e+1}$), as well as a pair of MODIS and Sentinel-2 captions on the same date, anytime before the fire starts (e.g. $t_{s-2}$), and produces a high resolution delineation of the burn scar. This approach alleviates the need for bitemporal high resolution imagery, thus obtaining timely and accurate burn scar maps soon after a wildfire is deemed complete.
• Figure 2: Normalized spectral responses of the Sentinel-2 and MODIS satellites.
• Figure 3: Overview of the BAM-MRCD model based on the BAM-CD architecture sdraka2024floga. Numbers inside the ResBlock, ConvBlock and AConvBlock modules indicate the number of output channels for each internal convolutional layer. ConvBlock is identical to AConvBlock, without the attention module.
• Figure 4: Example of patch grouping into small, medium and large based on the size of the burnt area.
• Figure 5: Distributions of (a) the false positive ratio and (b) the false negative ratio for BAM-MRCD and ChangeMamba (Approach 1) in the test set. Each ratio is defined by the number of false positives or false negatives over the total number of positives or negatives, respectively.