💻 computer science

Low-Rank Adaptation of Geospatial Foundation Models for Wildfire Mapping Using Sentinel-2 Data

This study demonstrates that Low-Rank Adaptation (LoRA) enables state-of-the-art geospatial foundation models, particularly Prithvi-v2, to achieve superior cross-domain generalization and accuracy in wildfire burned-area mapping using Sentinel-2 data while updating less than 1% of parameters.

Original authors: Ali Shibli, Andrea Nascetti, Yifang Ban

Published 2026-05-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ali Shibli, Andrea Nascetti, Yifang Ban

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

Imagine you have a super-smart, all-knowing librarian who has read every book, map, and satellite photo ever taken of the Earth. This librarian is a "Geospatial Foundation Model." They know what forests, deserts, and cities look like in incredible detail.

However, there's a problem: if you ask this librarian to specifically find wildfire scars (burned areas) in a new part of the world, they might get confused. They are so used to general knowledge that they might mistake a dry field for a fire, or miss a small patch of burn because they haven't seen that specific type of fire before.

This paper is about teaching this super-librarian how to spot wildfires quickly and efficiently without making them forget everything they already know or requiring a massive, expensive brain surgery.

The Challenge: The "Brain Surgery" Problem

Usually, to teach an AI a new specific task (like finding fire), scientists do "Full Fine-Tuning." Think of this as retraining the librarian's entire brain. You make them re-read every single book they've ever read, but this time focusing only on fire.

The downside: It takes a huge amount of time, money, and computer power. Plus, if you train them too hard on one specific region, they might forget how to handle other regions (like confusing a Canadian forest fire with an Australian one).

The Solution: The "Sticky Note" Method (LoRA)

The researchers tried a smarter approach called Low-Rank Adaptation (LoRA).

The Analogy: Instead of rewriting the librarian's entire brain, you just give them a tiny, specialized notebook (a "sticky note" system) to carry around.
The librarian keeps their original, massive knowledge base frozen (untouched).
They only write new, specific rules for spotting wildfires in this tiny notebook.
The Magic: This notebook is so small that it updates less than 1% of the librarian's total knowledge, yet it helps them perform the specific task almost perfectly.

What They Did

The team tested three different "super-librarians" (AI models named TerraMind, DINOv3, and Prithvi-v2) using satellite photos from Sentinel-2 (a camera in space that takes pictures of Earth).

The Data: They looked at 3,820 wildfires across the US and Canada from 2017 to 2023.
The Test: They tried to teach these models to find burned areas in places and years they had never seen before (a "spatiotemporal" test).
The Comparison: They compared three teaching methods:
- Full Fine-Tuning: Rewriting the whole brain.
- Decoder-Only: Only training the part that draws the final map, leaving the brain alone.
- LoRA: Using the tiny "sticky note" notebook.

The Results

The Winner: The LoRA method (the sticky notes) won every time. It was the most accurate at finding wildfires in new places and new years.
The Champion Model: Among the three librarians, Prithvi-v2 was the best at the job, especially when using the LoRA method. It improved its accuracy significantly more than the others.
Efficiency: LoRA achieved these high scores while changing less than 1% of the model's parameters. It's like getting a PhD in wildfire spotting by reading just a few pages of a new book, rather than re-reading the whole library.

Why This Matters

The paper shows that you don't need to rebuild the entire AI engine to make it good at a specific job. By using LoRA, we can take powerful, pre-trained models and adapt them to map wildfires across huge areas (like the whole US and Canada) quickly and cheaply.

The researchers also noted that this method is great for real-world use: you can keep one giant "base" model on a server and just swap out the tiny "LoRA notebooks" depending on whether you are looking at a forest in Canada or a desert in the US, without needing to download a whole new massive model every time.

In short: They found a way to turn a general-purpose Earth expert into a wildfire specialist using a tiny, efficient "cheat sheet," making it faster and cheaper to map fire damage from space.

Technical Summary: Low-Rank Adaptation of Geospatial Foundation Models for Wildfire Mapping

Problem Statement
Accurate burned-area (BA) mapping is critical for damage assessment, emissions modeling, and understanding fire–climate interactions. While deep learning (e.g., U-Net variants) has improved upon traditional spectral index methods, existing approaches often suffer from limited geographic and temporal generalization, frequently focusing on single countries or specific fire seasons. Furthermore, the emerging class of Geospatial Foundation Models (GFMs), pre-trained on massive satellite imagery collections, presents a new opportunity. However, adapting these large models (hundreds of millions of parameters) to specific downstream tasks like wildfire mapping via full fine-tuning is computationally expensive and prone to overfitting under domain shifts. Conversely, decoder-only fine-tuning may fail to capture necessary domain-specific variations in the encoder. There is a lack of understanding regarding the most efficient adaptation strategy for GFMs in large-scale, cross-domain wildfire mapping.

Methodology
The study treats burned-area mapping as a bi-temporal change detection problem using Sentinel-2 data (B4, B8, B12 bands) from the United States and Canada (2017–2023). The dataset comprises 3,820 wildfire events, split into a source domain (2017–2020, diverse biomes) and a target domain (2021–2023, specifically Boreal/Taiga and Tundra biomes) to test spatiotemporal generalization.

The proposed architecture utilizes a shared Transformer-based encoder with a UPerNet decoder. The core investigation compares three adaptation strategies for three state-of-the-art GFMs: TerraMind, DINOv3, and Prithvi-v2:

Full Fine-Tuning: Updating all parameters.
Decoder-Only Fine-Tuning: Freezing the encoder and updating only the decoder.
Low-Rank Adaptation (LoRA): Freezing the original encoder weights and injecting trainable low-rank matrices ( $\Delta W = BA$ ) into specific layers (self-attention, projection, and MLP layers for Prithvi-v2).

The LoRA implementation freezes the entire backbone, updating only the adapters, a pyramidal neck, the decoder, and the classification head. The models are trained using a class-balanced cross-entropy loss (3:1 weight for burned vs. unburned) to address class imbalance.

Key Results
Experiments were conducted across 3,820 wildfire events with a focus on cross-domain generalization (temporal and biome shifts). The results indicate:

Superiority of LoRA: Across all three foundation models, LoRA achieved the highest Intersection over Union (IoU) and F1 scores, outperforming both full fine-tuning and decoder-only fine-tuning.
Performance Gains:
- Prithvi-v2: Showed the most significant improvement with LoRA, achieving an IoU of 78.78 and F1 of 88.13. This represented a substantial gain of +6.8 IoU and +4.4 F1 over full fine-tuning.
- TerraMind & DINOv3: Also saw improvements with LoRA (e.g., TerraMind IoU increased from 70.52 to 75.59).
Parameter Efficiency: LoRA updated less than 1% of the encoder parameters (0.51% for TerraMind/DINOv3 and 1.03% for Prithvi-v2). Even when considering the full encoder-decoder network, the trainable parameter count remained significantly lower than full fine-tuning (e.g., 6.79% for Prithvi-v2 full network vs. 100% in full fine-tuning).
Generalization: Decoder-only fine-tuning consistently outperformed full fine-tuning, suggesting a regularization effect where updating all backbone parameters leads to overfitting. LoRA further enhanced this robustness, reducing false positives along fire perimeters and false negatives in unburned regions.

Significance and Claims
The authors claim that this study demonstrates that geospatial foundation models, when adapted using lightweight parameter-efficient methods like LoRA, offer a robust and scalable solution for large-scale burned-area mapping. Key claims include:

Efficiency vs. Accuracy Trade-off: LoRA provides the strongest cross-domain generalization while updating less than 1% of parameters, offering a favorable trade-off between accuracy and computational cost.
Domain-Specific Pre-training: The results suggest that domain-specific pre-training (as seen in Prithvi-v2) yields stronger representations for wildfire regions than generic vision models, and LoRA efficiently isolates fire-induced spectral changes within these high-capacity models.
Operational Scalability: The approach allows for a single frozen foundation model to be deployed on a central server, with lightweight LoRA adapters dynamically swapped to process imagery from different geographic or seasonal fire regimes without loading entirely separate models.
Global Potential: While the study is geographically focused on North America, the successful transfer across distinct biomes and years suggests the learned representations are not strictly localized, implying the approach is globally scalable with minimal target-domain data for new regions.

The study concludes that parameter-efficient fine-tuning is a practical choice for operational cross-domain wildfire monitoring, avoiding the computational burdens and generalization pitfalls of full fine-tuning.