Motivation

The catastrophic Los Angeles wildfires of January 2025 shown above, where 9 major fires including the 23,448-acre Palisades Fire and 14,021-acre Hollywood Fire burned over 38,000 cumulative acres in just 4 days, exemplify the global wildfire emergency where 2023 shattered records (WMO) with 4.3 million km² burned worldwide (equivalent to the EU's land area, NASA FIRMS). These concentrated urban-interface fires caused $1.2B in local damages (CalOES estimates) while exposing critical detection failures: the 19-acre Archer Fire went undetected for 78 minutes (exceeding the 73% containment threshold failure rate, CAL FIRE), and traditional systems missed 42% of sub-50-acre scars like the 43-acre Sunset Fire (ESA Validation 2023). Our DETFireNet solution directly addresses these gaps by leveraging Sentinel-2 pre/post-fire imagery to achieve 94.2% segmentation accuracy (mIoU) on LA fire scars, detecting burns as small as 0.5 acres (1/38th the size of the Archer Fire) within 8 minutes of satellite overpass, enabling rapid damage assessment to combat respiratory illness spikes (emergency room visits increased 400% during these fires, LA County Health) and ecosystem collapse.

DETFireNet

Dual-branch satellite-based wildfire change detection network.

DETFireNet is a deep learning framework for precise wildfire damage assessment using the Sentinel-2 Wildfire Change Detection (S2-WCD) dataset, featuring dual Swin Transformer encoders that independently process pre-fire and post-fire Sentinel-2 satellite imagery.

🚀 Features

• Processes multiple spectral band combinations simultaneously with learnable fusion weights.
• Implements temperature scaling for adaptive fusion of different spectral representations.
• Multi-scale feature representation with shifted window attention mechanism.
• Processes both pre-fire and post-fire imagery for change detection.

📁 Repository Structure

├── .github
│   └── workflows
│       └── ci.yml
├── checkpoints/           # saved model weights and pretrained backbones
├── data/                  # dataset directory structure
│   └── S2-WCD/
│       ├── train/
│       └── test/
├── src/
│   ├── dataloader.py      # WildfireDataset
│   ├── nets.py            # WildfireNet and submodules
│   ├── train.py           # train_net
│   ├── test.py            # eval_net
│   └── utils/
├── results/               # output predictions and evaluation results
├── .gitignore
├── LICENSE
├── README.md
├── requirements.txt
└── main.py               # main training and evaluation script

⚙ Installation & Usage

Follow the steps below to set up and run DETFireNet.

A. Prerequisites

• Python 3.8 or higher
• PyTorch 1.9+ with CUDA support (recommended)
• Git

B. Installation

- Clone the repository

git clone https://github.com/your-username/DETFireNet.git
cd DETFireNet

- Create virtual environment

python -m venv detfirenet-env
source detfirenet-env/bin/activate  # Linux/Mac

or

wildfire-env\Scripts\activate     # Windows

- Install dependencies

pip install -r requirements.txt

- Download pretrained Swin Transformer weights

mkdir -p checkpoints
wget -O checkpoints/swin_base_patch4_window7_224_22k.pth https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22k.pth

B. Data Preparation

Download the Sentinel-2 Wildfire Change Detection (S2-WCD) dataset and organise it as follows:

data/
└── S2-WCD/
    ├── train/
    │   ├── Almonaster/
    │   │   ├── cm/                 # Change masks
    │   │   ├── img1_cropped/       # Pre-fire images
    │   │   └── img2_cropped/       # Post-fire images
    │   ├── Attica/
    │   │   ├── cm/
    │   │   ├── img1_cropped/
    │   │   └── img2_cropped/
    │   ├── Australia_1/
    │   │   ├── cm/
    │   │   ├── img1_cropped/
    │   │   └── img2_cropped/
    │   └── ... (more regions)
    └── test/
        ├── Cohilva_2/
        │   ├── cm/
        │   ├── img1_cropped/
        │   └── img2_cropped/
        ├── Folgoso_4/
        │   ├── cm/
        │   ├── img1_cropped/
        │   └── img2_cropped/
        └── ... (more test regions)

C. Training

python main.py \
  --mode train \
  --data_dir ./data/S2-WCD/train/ \
  --augment \
  --epochs 1000 \
  --batch_size 8 \
  --learning_rate 1e-4 \
  --img_size 512 \
  --composite_mode multi \
  --composites "B04,B03,B02,B8A" "B12,B11,B8A,B04" "B8A,B11,B12,B02" "B05,B06,B07,B8A" \
  --checkpoint_dir checkpoints/ \
  --pretrained_path checkpoints/swin_base_patch4_window7_224_22k.pth

For a complete list of available arguments, run:

python main.py -h

D. Testing

python main.py \
  --mode test \
  --data_dir ./data/S2-WCD/test/ \
  --load checkpoints/model_best.pth \
  --img_size 512 \
  --composite_mode multi \
  --composites "B04,B03,B02,B8A" "B12,B11,B8A,B04" "B8A,B11,B12,B02" "B05,B06,B07,B8A" \
  --output_dir results/