💻 computer science

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

To overcome the trade-off between spatial resolution and temporal frequency in rapid wildfire monitoring, this paper proposes BAM-MRCD, a novel deep learning model that fuses multi-resolution MODIS and Sentinel-2 imagery to generate accurate, detailed burn scar maps shortly after fire incidents.

Original authors: Maria Sdraka, Dimitrios Michail, Ioannis Papoutsis

Published 2026-01-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maria Sdraka, Dimitrios Michail, Ioannis Papoutsis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Too Fast vs. Too Blurry" Dilemma

Imagine you are trying to draw a map of a forest fire the moment it stops burning. You need two things to do this perfectly:

A super-detailed camera (like a high-end DSLR) to see the exact edges of the burnt trees.
A camera that never blinks (like a security camera that takes a picture every minute) to catch the fire the second it goes out.

The problem is that in the world of satellites, you usually can't have both.

The "Security Camera" (MODIS): This satellite flies over the Earth every day. It sees the whole picture quickly, but the image is blurry (like a low-resolution phone photo). It's great for knowing where the fire is, but bad at showing the fine details of the burnt edges.
The "Super-Detailed Camera" (Sentinel-2): This satellite takes incredibly sharp, high-definition photos. But it only flies over a specific spot once every few days. If you wait for it to show up after a fire, the smoke might have cleared, or the damage might have been assessed too late for emergency crews.

The Solution: The "Smart Detective" (BAM-MRCD)

The authors created a new AI model called BAM-MRCD. Think of this model as a brilliant detective who solves a crime by combining two different types of witness testimony.

Instead of waiting for the perfect photo, the detective uses a clever trick:

The "Before" Photo: They grab a super-sharp photo of the forest before the fire happened (from the detailed camera).
The "After" Photo: They grab a blurry photo of the forest immediately after the fire (from the daily camera).
The "Before" Blurry Photo: They also grab a blurry photo of the forest from the same day as the sharp "before" photo, just to help the detective understand what the blurry camera usually sees.

The AI then compares these three images. It uses the sharp "before" photo to know what the trees and ground should look like, and the blurry "after" photo to see what changed. It essentially "fills in the gaps" of the blurry photo using the sharp memory of the landscape.

How It Works (The "Parallel Kitchen" Analogy)

Most old AI models tried to fix the blurry photo first (like trying to sharpen a pixelated image) and then find the fire. The authors say this is like trying to sharpen a blurry photo of a cake and then guessing which part is burnt. It often creates "artifacts" (weird digital noise) that confuse the AI.

Instead, BAM-MRCD works like a kitchen with two chefs working in parallel:

Chef A looks at the daily, blurry photos to spot the big changes (the "smoke" and the general burnt area).
Chef B looks at the sharp, pre-fire photo to remember the fine details (the exact shape of the trees and roads).
The Head Chef (The AI) takes notes from both chefs at the same time. It doesn't try to fix the blurry photo; it just uses the sharp memory to guide the blurry observation.

This allows the model to produce a high-definition map of the burnt area the very next day after the fire, without waiting for the high-resolution satellite to return.

What They Found (The Results)

The team tested this "Smart Detective" on hundreds of real wildfires in Greece. Here is what happened:

It's Faster: It can map a fire the day after it happens, whereas waiting for the sharp satellite might take 3–5 days.
It's Smarter: It found small fires that other models missed. Other models were like "sledgehammers" that only saw big fires; this model is a "scalpel" that can see small patches of burnt grass too.
It's Accurate: It drew the borders of the burnt areas very precisely, avoiding false alarms (like confusing a harvested cornfield with a burnt forest).

The Limitations (Where the Detective Gets Stumped)

Even the best detective has blind spots. The paper admits the model struggles in two specific situations:

Tiny Fires: If a fire is so small that it fits inside just one or two pixels of the blurry daily camera, the AI can't see it. It's like trying to see a single ant on a blurry photo of a football field.
Clouds and Smoke: If the sky is covered in thick clouds or smoke when the daily camera flies over, the AI can't see the ground. It has to wait until the sky clears.

The Bottom Line

This paper introduces a new way to use two different types of satellite data together. By combining the "daily but blurry" view with the "sharp but rare" view, they created an AI that can map wildfire damage almost instantly. This helps emergency teams plan their recovery efforts much faster, potentially saving ecosystems and communities from further damage.

Technical Summary: Magnifying Change: Rapid Burn Scar Mapping with Multi-Resolution, Multi-Source Satellite Imagery

Problem Formulation
The paper addresses the critical challenge of rapidly delineating wildfire burn scars using satellite imagery. While deep learning models achieve high accuracy with high-resolution data (e.g., Sentinel-2), their operational utility is limited by the trade-off between spatial resolution and temporal revisit frequency. High-resolution satellites often have revisit times of several days, delaying post-fire assessment. Conversely, high-temporal resolution sensors (e.g., MODIS) offer daily coverage but at a coarse spatial resolution (500m), which is insufficient for detailed perimeter mapping, especially for smaller fires.

The authors define a specific "Burnt Area Mapping" (BAM) problem distinct from generic change detection. Unlike urban change detection which relies on sharp geometric edges, burn scars exhibit heterogeneous spectral changes across the electromagnetic spectrum (particularly in red, near-infrared, and shortwave infrared bands) and variable severity. The core objective is to produce a high-resolution (60m) binary burn scar map immediately after a fire event ( $t_e + 1$ ) using:

Pre-fire High-Spatial/Low-Temporal (HSLT) imagery: Sentinel-2 ( $I_{t1}$ ).
Post-fire Low-Spatial/High-Temporal (LSHT) imagery: MODIS ( $I_{t2}$ ).
Pre-fire LSHT imagery: MODIS ( $I_{t1}$ ) to facilitate change detection.

This formulation (Approach 3) aims to bypass the need for bitemporal high-resolution imagery, which is often unavailable immediately after a disaster.

Methodology: BAM-MRCD
The authors propose BAM-MRCD (Burnt Area Mapping - MultiResolution Change Detection), a novel deep learning architecture that avoids the error propagation common in traditional two-step approaches (super-resolution followed by change detection).

Architecture: The model employs a parallel processing strategy using two distinct branches:
1. BAM-CD Branch: A Siamese (or pseudo-Siamese) DeepResUnet architecture that processes bitemporal MODIS imagery ( $I_{t1}$ and $I_{t2}$ ) to detect changes at the native 500m resolution.
2. UNet Branch: A standard UNet-like model that processes the pre-fire Sentinel-2 imagery ( $I_{t1}$ ) to extract high-frequency spatial details.
Feature Fusion: Features from the BAM-CD encoders are concatenated and passed through a decoder with "squeeze-and-excitation" attention mechanisms to isolate relevant change features. These features are then fused with the high-resolution features from the UNet branch and passed to a final decoder for the 60m prediction.
Deep Supervision: The model utilizes a compound loss function. In addition to the final high-resolution loss, an auxiliary low-resolution loss is applied to the output of the BAM-CD branch. This guides the network to learn the change detection task at the MODIS scale before refining the details at the Sentinel-2 scale, effectively distilling the complex multi-resolution problem into simpler subtasks.
Data: The model is trained on the FLOGA dataset, containing 326 wildfire events in Greece (2017–2021) with manual annotations. The dataset includes paired Sentinel-2 and MODIS imagery. The authors note the class imbalance issue and employ a spatial split strategy to prevent data leakage.

Key Contributions

Problem Definition: The paper formulates burn scar mapping as a unique multi-resolution, multi-source change detection task, distinct from standard change detection due to the spectral nature of fire and the lack of strict geometric boundaries.
Novel Architecture: BAM-MRCD is the first proposed model to exploit multi-resolution pre-change and low-resolution post-change inputs in a parallel fashion without relying on a dedicated super-resolution module, thereby avoiding the introduction of artifacts.
Evaluation Metric: Recognizing that standard metrics like F1-score and IoU are saturated by large overlaps and fail to capture performance on small-scale fires, the authors propose a multi-scale IoU metric. This metric groups events by size (small, medium, large) to provide a more granular assessment of a model's robustness across varying fire scales.

Results

Performance: BAM-MRCD outperforms state-of-the-art baselines (including FC-EF-Diff, SNUNet-CD, ChangeFormer, MM-Trans, and DCILNet) across all metrics, particularly in the multi-scale IoU categories. It achieves an F1-score of 92.06% and successfully retrieves 36 out of 53 test events.
Comparison with Baselines:
- Conventional change detection models (Approach 1: MODIS-MODIS) struggle with small fires due to resolution limits.
- Models attempting direct multi-resolution input (Approach 2: Sentinel-2/MODIS) often fail to align features effectively or suffer from instability.
- SR-based approaches (e.g., DCILNet) underperform on small events, likely due to artifacts generated by the super-resolution module.
Qualitative Analysis: The model demonstrates superior ability to delineate small burn scars (< 90 ha) that are missed by other models. It also shows high precision in preserving complex boundaries and avoiding false positives on land covers with similar spectral signatures (e.g., harvested crops).
Generalization: The model was tested on unseen wildfire events in Greece, Bulgaria, and Portugal from the 2025 fire season. It successfully provided delineations days before the next available cloud-free Sentinel-2 image, demonstrating operational potential.

Limitations and Claims
The authors modestly acknowledge specific limitations:

Small/Linear Features: The model struggles with burn scars that are extremely narrow or oblong, where the feature width is smaller than the MODIS pixel size, leading to insufficient spectral signal.
Atmospheric Conditions: The reliance on multispectral imagery means performance is hindered by cloud or smoke coverage, though the high temporal frequency of MODIS mitigates this compared to high-resolution sensors.
Sensor Retirement: The authors note the planned retirement of MODIS satellites (late 2026/early 2027) but suggest the approach remains applicable as it can be adapted to the VIIRS constellation, which shares similar characteristics.

Significance
The paper claims that BAM-MRCD represents a significant step toward next-day burn scar mapping in operational settings. By leveraging the complementary strengths of MODIS (temporal frequency) and Sentinel-2 (spatial detail), the method enables rapid, high-resolution damage assessment crucial for immediate relief planning and ecosystem restoration. The work aims to propel research in this emerging field by providing a robust framework for handling multi-resolution remote sensing data in disaster management.