This project generates high-resolution monthly maps of Live Fuel Moisture Content (LFMC) for Maui County, Hawaii. LFMC measures water content in living vegetation as a percentage — below 80% indicates critical wildfire danger. The 2023 Lahaina fire, which killed 100+ people, occurred when vegetation was critically dry.
We reproduce Johnson et al. (2025) by fine-tuning the Galileo-Tiny foundation model on CONUS Globe-LFMC data, then apply the trained model zero-shot to Maui (the same approach used for the 2025 LA Palisades/Eaton fire case studies in that paper). There are no Hawaii samples in Globe-LFMC 2.0 — the contribution of this project is generating and analyzing pre/post-Lahaina LFMC maps using Galileo's learned representations.
| Resource | Link |
|---|---|
| Johnson et al. (2025) paper | https://arxiv.org/abs/2506.20132 |
| Official AllenAI LFMC pipeline | https://github.com/allenai/lfmc |
| Galileo foundation model | https://github.com/nasaharvest/galileo |
| Galileo weights (HuggingFace) | https://huggingface.co/nasaharvest/galileo |
| Globe-LFMC 2.0 dataset | https://doi.org/10.1038/s41597-024-03159-6 |
| Globe-LFMC 2.0 data (figshare) | https://doi.org/10.6084/m9.figshare.24312164 |
- Galileo-Tiny: 5.3M parameter Vision Transformer pretrained on multimodal satellite data
- Regression head: Single linear layer mapping encoder embedding → LFMC %
- Training: Fine-tuned on Globe-LFMC 2.0 CONUS subset (41,214 samples, 1,031 sites)
- Target performance: RMSE ≈ 18.9, R² ≈ 0.72 (Johnson et al. 2025 Table 1)
Each sample = one 1km × 1km GeoTIFF with 12 monthly composites of:
- Sentinel-2 L1C (10 bands, optical/NDVI)
- Sentinel-1 (VV, VH SAR backscatter)
- ERA5-Land (temperature, precipitation, ET)
- TerraClimate (climate water balance)
- VIIRS (night lights)
- SRTM (elevation, slope)
- DynamicWorld + WorldCereal (land cover)
- LandScan (population, static)
Maui County (20.5°–21.1°N, 156.7°–155.9°W) is tiled into overlapping 32×32 pixel patches (320m × 320m at 10m resolution), processed by the CONUS-trained model, then stitched into a wall-to-wall GeoTIFF via overlap averaging to remove edge artifacts.
maui-lfmc/
├── src/
│ ├── data/
│ │ └── download_tifs.py # GEE satellite download (training + inference)
│ ├── model/
│ │ └── train.py # Fine-tune Galileo on Globe-LFMC (wraps allenai/lfmc)
│ └── inference/
│ └── map_generator.py # Generate monthly LFMC maps for Maui County
├── requirements.txt
└── README.md
conda create -n lfmc python=3.11
conda activate lfmc
conda install -c conda-forge gdal rasterio geopandas
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txtgit clone --recurse-submodules https://github.com/allenai/lfmc.git allenai-lfmc
pip install -e allenai-lfmc/submodules/galileo
pip install -e allenai-lfmcearthengine authenticateDownload Globe-LFMC-2.0.xlsx from https://doi.org/10.6084/m9.figshare.24312164,
then run:
cd allenai-lfmc
python -m lfmc.main.create_csvThis creates data/labels/lfmc_data_conus.csv (~90K CONUS samples).
# Small test batch first (recommended)
python -m src.data.download_tifs \
--labels allenai-lfmc/data/labels/lfmc_data_conus.csv \
--output data/tifs/ \
--project YOUR_GEE_PROJECT \
--limit 100
# Full dataset (runs overnight)
python -m src.data.download_tifs \
--labels allenai-lfmc/data/labels/lfmc_data_conus.csv \
--output data/tifs/ \
--project YOUR_GEE_PROJECTpython -m src.model.train \
--galileo-config-dir allenai-lfmc/submodules/galileo/data \
--data-dir data/tifs/ \
--h5py-dir data/h5pys/ \
--labels allenai-lfmc/data/labels/lfmc_data_conus.csv \
--output checkpoints/conus/Target: RMSE ≈ 18.9, R² ≈ 0.72.
Note:
allenai-lfmcis a separate repo cloned alongside this one. Use the absolute path toallenai-lfmc/submodules/galileo/data— relative paths can fail depending on working directory.
# August 2023 (month of Lahaina fire)
python -m src.inference.map_generator \
--checkpoint checkpoints/conus/finetuned_model.pth \
--galileo-config /path/to/allenai-lfmc/submodules/galileo/data \
--year 2023 --month 8 \
--project YOUR_GEE_PROJECT \
--output outputs/maps/
# All months for 2023
python -m src.inference.map_generator \
--checkpoint checkpoints/conus/finetuned_model.pth \
--galileo-config /path/to/allenai-lfmc/submodules/galileo/data \
--year 2023 --all-months \
--project YOUR_GEE_PROJECT \
--output outputs/maps/Output: outputs/maps/lfmc_maui_2023_08.tif — a GeoTIFF at 10m resolution covering
all of Maui County, with LFMC values in % (nodata = -9999).
August 2023 LFMC Map (month of Lahaina fire):
The orange/red zone in the central valley shows LFMC near or below the critical 80% fire danger threshold. The Lahaina fire origin (★) sits at the western coast where dry conditions met high wind exposure on August 8, 2023.
Explore all monthly maps at: maui-lfmc-web.vercel.app
Features:
- Year/month selector (2021–2026 as maps are generated)
- Click any pixel to see exact LFMC % and fire risk level
- Lahaina fire origin marker with historical context
(Figure generated once Aug 2021 / 2022 / 2024 maps complete)
(Figure generated once all 12 months complete)
Globe-LFMC 2.0 contains zero samples from Hawaii. Rather than attempting to train on Hawaii data (which doesn't exist), we follow Johnson et al.'s zero-shot transfer approach: train on CONUS, apply to new geography using Galileo's pretrained multi-modal representations. This worked for the 2025 LA fires and is appropriate for Maui since the vegetation types (chaparral-adjacent dry shrubland) and fire dynamics are similar to fire-prone CONUS regions.
The key scientific contribution is the temporal analysis: comparing LFMC maps from months leading up to August 8, 2023 to understand how vegetation drought developed before the Lahaina fire.
This project is conducted as part of a NASA Harvest internship. It builds directly on:
- Johnson et al. (2025) — LFMC methodology and Galileo application
- Tseng et al. (2025) — Galileo foundation model
- Rao et al. (2020) — Globe-LFMC 2.0 dataset