Detection of Upper-Level Troughs and Ridges Using Deep Learning — Checkpoints
Application in the Mediterranean
Production and cross-validation checkpoints accompanying Detection of Upper-Level Troughs and Ridges Using Deep Learning – Application in the Mediterranean by Ofir Ariel, Omer Sela, Hadas Saaroni, and Baruch Ziv.
The models detect synoptic-scale trough and ridge axes from ERA5 500 hPa geopotential height and horizontal wind on a regular 1° Mediterranean-domain grid.
Project resources
| Resource | Link |
|---|---|
| Paper | EGUsphere preprint |
| Interactive project page | Explore model predictions and seasonal climatology |
| Source code | Sela-Omer/upper-level-trough-ridge-detection |
| Dataset | Omer-Sela/upper-level-trough-ridge-detection-data |
| Model checkpoints | Omer-Sela/upper-level-trough-ridge-detection-models |
Recommended checkpoints
Use the all-data production checkpoint for ordinary inference:
| Task | Weights | Configuration |
|---|---|---|
| Trough | production/trough/model.safetensors |
production/trough/config.json |
| Ridge | production/ridge/model.safetensors |
production/ridge/config.json |
The directories cross_validation/<task>/fold_0 through fold_9 contain the held-out-fold checkpoints used to compute the reported evaluation results.
Repository structure
production/
trough/{model.safetensors,config.json,provenance.json}
ridge/{model.safetensors,config.json,provenance.json}
cross_validation/
trough/fold_0 ... fold_9/
ridge/fold_0 ... fold_9/
manifest.json
Every model.safetensors file contains model tensors only. The adjacent config.json records the architecture, preprocessing, task, monthly wind normalization, and extraction configuration. provenance.json records source and artifact checksums.
Inference
Install the package from the source repository, then run the production trough detector on a released scene:
ultr infer --task trough --sample-id 20180101T0000 --output prediction.json
For direct artifact access:
from huggingface_hub import hf_hub_download
weights = hf_hub_download(
repo_id="Omer-Sela/upper-level-trough-ridge-detection-models",
filename="production/trough/model.safetensors",
)
config = hf_hub_download(
repo_id="Omer-Sela/upper-level-trough-ridge-detection-models",
filename="production/trough/config.json",
)
Use the revision argument or the code CLI's --model-revision option to pin an immutable Hub commit or release tag.
Evaluation
Out-of-fold evaluation applies each scene's held-out checkpoint and the task-specific M1 spline extraction described in the paper.
| Task | Scenes | F1 | Completeness | Chamfer | TP / FP / FN |
|---|---|---|---|---|---|
| Trough | 600 | 0.84 | 0.87 | 1.04 | 2171 / 404 / 408 |
| Ridge | 200 | 0.75 | 0.84 | 1.20 | 562 / 241 / 121 |
Production checkpoints are retrained on all 600 trough or 200 ridge rows and are not used to calculate the cross-validation table.
Intended use and limitations
These research models identify synoptic-scale upper-level trough and ridge axes in the Mediterranean-domain regime represented by the training dataset. They are not operational forecast products. Performance has not been established for other geographic domains, pressure levels, grid resolutions, or substantially different atmospheric distributions.
Citation
If you use these checkpoints, please cite the accompanying paper.
License
The model weights are released under the Apache License 2.0. The associated dataset is licensed separately under CC BY 4.0.