The pretrained Prithvi-EO-1.0-100M parameter model is finetuned to detect Burn Scars on HLS data from the HLS Burn Scar Scenes dataset. This dataset includes input tiles of 512x512x6, where 512 is the height and width and 6 is the number of bands. The bands are:
Blue
Green
Red
Narrow NIR
SWIR 1
SWIR 2
It is important to point out that the HLS Burn Scar Scenes dataset includes a single timestep, while the Prithvi-100m was pretrained with three timesteps. The difference highlights the flexibility of this model to adapt to different downstream tasks and requirements.
Configuration used for fine-tuning is available through config
).
Results
The experiment conducted by running the mmseg stack for 50 epochs using the above config led to an IoU of 0.73 on the burn scar class and 0.96 overall accuracy. It is noteworthy that this leads to a reasonably good model, but further developement will most likely improve performance.
Inference and demo
The github repo includes an inference script that allows to run the burn scar model for inference on HLS images. These inputs have to be in geotiff format, including the channels described above (Blue, Green, Red, Narrow NIR, SWIR, SWIR 2) in reflectance units [0-1]. There is also a demo that leverages the same code here.
Feedback
Your feedback is invaluable to us. If you have any feedback about the model, please feel free to share it with us. You can do this by submitting issues on our open-source repository, hls-foundation-os, on GitHub.
Citation
If this model helped your research, please cite Prithvi-100M-burn-scar in your publications. Here is an example BibTeX entry:
@misc{Prithvi-100M-burn-scar,
author = {Roy, Sujit and Phillips, Christopher and Jakubik, Johannes and Fraccaro, Paolo and Ankur, Kumar and Avery, Ryan and Ji, Wei and Zadrozny, Bianca and Ramachandran, Rahul},
doi = {10.57967/hf/0953},
month = aug,
title = {{Prithvi 100M burn scar}},
url = {https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M-burn-scar},
year = {2023}
}
