|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
tags: |
|
- Pytorch |
|
- mmsegmentation |
|
- segmentation |
|
- Flood mapping |
|
- Sentinel-2 |
|
- Geospatial |
|
- Foundation model |
|
metrics: |
|
- accuracy |
|
- IoU |
|
--- |
|
### Model and Inputs |
|
The pretrained [Prithvi-100m](https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M/blob/main/README.md) model is finetuned to segment the extend of floods on Sentinel-2 images from the [Sen1Floods11 dataset](https://github.com/cloudtostreet/Sen1Floods11). |
|
|
|
The dataset consists of 446 labeled 512x512 chips that span all 14 biomes, 357 ecoregions, and 6 continents of the world across 11 flood events. The benchmark associated to Sen1Floods11 provides results for fully convolutional neural networks trained in various input/labeled data setups, considering Sentinel-1 and Sentinel-2 imagery. |
|
|
|
We extract the following bands for flood mapping: |
|
|
|
1. Blue |
|
2. Green |
|
3. Red |
|
4. Narrow NIR |
|
5. SWIR 1 |
|
6. SWIR 2 |
|
|
|
Labels represent no water (class 0), water/flood (class 1), and no data/clouds (class 2). |
|
|
|
The Prithvi-100m model was initially pretrained using a sequence length of 3 timesteps. Based on the characteristics of this benchmark dataset, we focus on single-timestamp segmentation. This demonstrates that our model can be utilized with an arbitrary number of timestamps during finetuning. |
|
|
|
 |
|
|
|
### Code |
|
|
|
The code for this finetuning is available through [github](https://github.com/NASA-IMPACT/hls-foundation-os/). |
|
|
|
The configuration used for finetuning is available through this [config](https://github.com/NASA-IMPACT/hls-foundation-os/blob/main/fine-tuning-examples/configs/sen1floods11.py). |
|
|
|
### Results |
|
|
|
Finetuning the geospatial foundation model for 100 epochs leads to the following performance on out-of-sample test data: |
|
|
|
| **Classes** | **IoU**| **Acc**| |
|
|:------------------:|:------:|:------:| |
|
| No water | 96.90% | 98.11% | |
|
| Water/Flood | 80.46% | 90.54% | |
|
|
|
| **aAcc** |**mIoU**|**mAcc**| |
|
|:------------------:|:------:|:------:| |
|
| 97.25% | 88.68% | 94.37% | |
|
|
|
|
|
The performance of the model has been further validated on an unseen, holdout flood event in Bolivia. The results are consistent with the performance on the test set: |
|
|
|
|
|
| **Classes** | **IoU**| **Acc**| |
|
|:------------------:|:------:|:------:| |
|
| No water | 95.37% | 97.39% | |
|
| Water/Flood | 77.95% | 88.74% | |
|
|
|
| **aAcc** |**mIoU**|**mAcc**| |
|
|:------------------:|:------:|:------:| |
|
| 96.02% | 86.66% | 93.07% | |
|
|
|
Finetuning took ~1 hour on an NVIDIA V100. |
|
|
|
|
|
### Inference |
|
The github repo includes an inference script that allows running the flood mapping model for inference on Sentinel-2 images. These inputs have to be geotiff format, including 6 bands for a single time-step described above (Blue, Green, Red, Narrow NIR, SWIR, SWIR 2) in order. There is also a **demo** that leverages the same code **[here](https://huggingface.co/spaces/ibm-nasa-geospatial/Prithvi-100M-sen1floods11-demo)**. |
