lfmc.md

November 3, 2025 ยท View on GitHub

Live Fuel Moisture Content (LFMC) Mapping

OlmoEarth-v1-FT-LFMC-Base is a model fine-tuned from OlmoEarth-v1-Base for predicting the live fuel moisture content of woody vegetation from Sentinel-2 and Sentinel-1 satellite images.

Here are relevant links for fine-tuning and applying the model per the documentation in the main README:

Model Details

The model inputs twelve timesteps of satellite image data, with one Sentinel-1 RTC mosaic and one Sentinel-2 L2A mosaic per 14-day period.

At each pixel, it regresses the LFMC of woody vegetation.

It achieves a mean squared error of 580.6 on our test set.

Training Data

The model is trained on the Globe-LFMC 2.0 dataset by Marta Yebra et al. We use the subset of the data in the continental US.

Each sample in the dataset specifies a longitude, latitude, timestamp, fuel type, and LFMC value. We only use the woody fuel type subset for this model. For each sample, we create an rslearn window centered at the sample's longitude/latitude and with time range equal to the 168 days ending at the sample's timestamp. We use rslearn to obtain twelve Sentinel-2 and Sentinel-1 images during that time range, with one per 14-day period.

We split the dataset into train, val, and test splits spatially, where 128x128 pixel grid cells are assigned via hash to train (70%), val (20%), or test (10%).

The training data is released under CC0.

Inference

Inference is documented in the main README. The prediction request geometry should have start timestamp equal to the timestamp for which you want to make the LFMC prediction (e.g., the current timestamp). The end timestamp won't be used and can be set arbitrarily, e.g. set equal to the start timestamp.

Fine-tuning

Fine-tuning is documented in the main README.