|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
tags: |
|
- Pytorch |
|
- mmsegmentation |
|
- segmentation |
|
- Crop Classification |
|
- Multi Temporal |
|
- Geospatial |
|
- Foundation model |
|
datasets: |
|
- ibm-nasa-geospatial/hls_burn_scars |
|
metrics: |
|
- accuracy |
|
- IoU |
|
--- |
|
### Model and Inputs |
|
The pretrained [Prithvi-100m](https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M/blob/main/README.md) parameter model is finetuned to classify crop and other land cover types based off HLS data and CDL labels from the [multi_temporal_crop_classification dataset](https://huggingface.co/datasets/ibm-nasa-geospatial/multi-temporal-crop-classification). |
|
|
|
This dataset includes input chips of 224x224x18, where 224 is the height and width and 18 is combined with 6 bands of 3 time-steps. The bands are: |
|
|
|
1. Blue |
|
2. Green |
|
3. Red |
|
4. Narrow NIR |
|
5. SWIR 1 |
|
6. SWIR 2 |
|
|
|
Labels are from CDL(Crop Data Layer) and classified into 13 classes. |
|
|
|
 |
|
|
|
The Prithvi-100m model was initially pretrained using a sequence length of 3 timesteps. For this task, we leverage the capacity for multi-temporal data input, which has been integrated from the foundational pretrained model. This adaptation allows us to achieve more generalized finetuning outcomes. |
|
|
|
### Code |
|
Code for Finetuning is available through [github](https://github.com/NASA-IMPACT/hls-foundation-os/tree/main/fine-tuning-examples) |
|
|
|
Configuration used for finetuning is available through [config](https://github.com/NASA-IMPACT/hls-foundation-os/blob/main/fine-tuning-examples/configs/multi_temporal_crop_classification.py). |
|
|
|
### Results |
|
The experiment by running the mmseg stack for 80 epochs using the above config led to the following result: |
|
|
|
| **Classes** | **IoU**| **Acc**| |
|
|:------------------:|:------:|:------:| |
|
| Natural Vegetation | 0.3362 | 39.06% | |
|
| Forest | 0.4362 | 65.88% | |
|
| Corn | 0.4574 | 54.53% | |
|
| Soybeans | 0.4682 | 62.25% | |
|
| Wetlands | 0.3246 | 45.62% | |
|
| Developed/Barren | 0.3077 | 49.1% | |
|
| Open Water | 0.6181 | 90.04% | |
|
| Winter Wheat | 0.4497 | 66.75% | |
|
| Alfalfa | 0.2518 | 63.97% | |
|
|Fallow/Idle Cropland| 0.328 | 54.46% | |
|
| Cotton | 0.2679 | 66.37% | |
|
| Sorghum | 0.2741 | 75.91% | |
|
| Other | 0.2803 | 39.76% | |
|
|
|
|**aAcc**|**mIoU**|**mAcc**| |
|
|:------:|:------:|:------:| |
|
| 54.32% | 0.3692 | 59.51% | |
|
|
|
It is important to acknowledge that the CDL (Crop Data Layer) labels employed in this process are known to contain noise and are not entirely precise, thereby influencing the model's performance. Fine-tuning the model with more accurate labels is expected to further enhance its overall effectiveness, leading to improved results. |
|
|
|
|
|
### Inference and demo |
|
There is an inference script that allows to run the hls-cdl crop classification model for inference on HLS images. These input have to be geotiff format, including 18 bands for 3 time-step, and each time-step includes the channels described above (Blue, Green, Red, Narrow NIR, SWIR, SWIR 2) in order. |
|
|
|
|
|
|
|
|
