Utrack (AI) tracker
Warning
This tracker is currently in beta. The underlying model is not publicly available
and must be obtained separately (see Model weights below). It is only compatible
with AROME model outputs. In addition, the model frequently fails to detect a cyclone
on some inputs; in those cases it falls back to coordinates (0, 0), which should be
treated as a missing value in downstream processing.
Overview
The Utrack tracker (tracking_method: "utrack") estimates the cyclone center at each time
step using a convolutional neural network based on the U-Net architecture, as introduced
by Raynaud, et al. (2024).
The model was designed to detect the TC wind structure — including the maximum wind speed area
and the hurricane-force wind speed area — directly from AROME convective-scale NWP outputs,
without relying on heuristic rules or empirical thresholds.
The model was trained and evaluated on a dataset of 400 hand-labeled AROME forecasts over the West Indies domain, covering Atlantic hurricane seasons 2016–2018.
Required input fields
This tracker requires the following variables to be available in the dataset:
10 m zonal wind component (
u10),10 m meridional wind component (
v10),Absolute vorticity at 850 hPa (
absv).
Output
The tracker returns two 1D arrays:
cy(time): row index of the estimated center,cx(time): column index of the estimated center,
both returned as integers and packaged in an xarray.Dataset:
xr.Dataset({"cy": cy, "cx": cx})
Configuration and usage
To activate this tracker, set tracking_method: "utrack" in the YAML configuration and
provide the three parameters below.
tracking_method: "utrack"
utrack_weights_file: "/path/to/model_latest.ckpt"
utrack_use_gpu: false
utrack_batch_size: 16
Parameter |
Default |
Description |
|---|---|---|
|
(required) |
Path to the Utrack model checkpoint file. The simulation will raise a
|
|
|
Whether to run inference on GPU. Set to |
|
|
Number of time steps processed in a single forward pass. Reduce this value if GPU or CPU memory is limited. |
Dependencies
This tracker requires the utrack package to be installed. It is not part of the
standard FrameIt dependencies. If utrack is not installed, the tracker class is
still importable but will raise an error at instantiation time.
Clone the repository and install the package locally:
git clone https://git.meteo.fr/hoarauk/unet_tracker
cd unet_tracker
pip install .
Model weights
A trained model checkpoint file is required. The weights are not publicly distributed;
contact the model authors to obtain a copy. Once you have the file, set
utrack_weights_file in your configuration to its absolute path:
utrack_weights_file: "/path/to/model_latest.ckpt"