Providing a set of Python modules for supporting the processing and analysis of TC-related data and data-driven models
The module addresses tropical cyclones analysis by providing the tools for gathering and preprocessing data, training ML models, and post-process results.
Two versions are being implemented, that will be integrated in the next release:
- A version using CNNs written in TensorFlow, which is currently being ported to PyTorch
- A version that uses GNNs instead of CNNs, written in PyTorch Geometric
Release Notes
The module is being developed as a Python package. In this version, a set of data-driven models based on VGG have been developed and tested. An ML ensemble of VGG neural networks has also been implemented as well as an alternative version of the pipeline that makes use of GNNs rather than CNNs. To accomplish so, the matrix input data is treated as a graph, and processed accordingly with GNNs. The library includes a set of features for preparing the data for training the model, including: capabilities for splitting the input gridded data into non-overlapping patches, patches selection, and storage as TFRecords or Zarr. Moreover, the code includes the trainable model for TC detection, as well as a deterministic tracker scheme.
Future Plans
The next version will provide code based on PyTorch for both the approaches explored. The code will be merged into a single repository where common functionalities will be integrated. Furthermore, novel ML approaches will be explored to improve the detection skills. A multi-model thematic module will be also developed to support ensemble detection over multiple CMIP6 data. Additional usage examples will be included in the documentation.
