File size: 2,868 Bytes
9b69290
 
 
 
 
 
 
 
 
 
 
fc7c6c4
 
 
 
 
 
32a41c8
 
 
dda58ce
9a88a25
32a41c8
 
 
 
 
 
f13bf02
a8430c0
dda58ce
32a41c8
a8430c0
32a41c8
a8430c0
dda58ce
32a41c8
 
a8430c0
dda58ce
 
a8430c0
 
99c8e62
 
 
 
a8430c0
 
 
 
 
 
 
48dd796
a8430c0
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
---
license: apache-2.0
language:
- en
tags:
- Pytorch
- mmsegmentation
- segmentation
- burn scars
- Geospatial
- Foundation model
datasets:
- ibm-nasa-geospatial/hls_burn_scars
metrics:
- accuracy
- IoU
- F1 Score
---

### Model and Inputs
The pretrained [Prithvi-100m](https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M/blob/main/README.md) parameter model is finetuned to detect Burn Scars on HLS data from the [HLS Burn Scar Scenes dataset](https://huggingface.co/datasets/ibm-nasa-geospatial/hls_burn_scars). This dataset includes input tiles of 512x512x6, where 512 is the height and width and 6 is the number of bands. The bands are:
 
1. Blue
2. Green
3. Red
4. Narrow NIR
5. SWIR 1
6. SWIR 2
![](burn_scar.png)
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.

### Code
Code for fine-tuning is available through [Github](https://github.com/NASA-IMPACT/hls-foundation-os/tree/main/configs)

Configuration used for fine-tuning is available through [config](https://github.com/NASA-IMPACT/hls-foundation-os/blob/main/configs/burn_scars.py)
).

### 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](https://huggingface.co/spaces/ibm-nasa-geospatial/Prithvi-100M-Burn-scars-demo)**.

### 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](https://github.com/NASA-IMPACT/hls-foundation-os/issues), 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}
}
```