keras_climate

A Keras 3 (TensorFlow / JAX / PyTorch backend) framework of models for climate modeling and Earth observation, plus tooling to port pretrained weights from the original reference implementations (almost always PyTorch) onto the Keras equivalents.

Every model is a plain keras.Model (or keras.Model subclass) builder function — no custom training loop is imposed, so they compose with standard model.fit(...), custom losses, distribution strategies, etc.

Package structure

keras_climate
├── remote_sensing        pixel-wise segmentation / classification of imagery
├── forecasting            multivariate time-series forecasting
├── weather                spatiotemporal nowcasting / weather models
├── foundation             large pretrained Earth-observation foundation models
├── operators              neural operator learning (PDE surrogate modeling)
├── datasets/              Keras port of torchgeo.datasets (190+ concrete datasets)
├── samplers/               Keras port of torchgeo.samplers (geospatial tile sampling)
├── transforms/             Keras port of torchgeo.transforms (band indices, SAR, ...)
├── losses/                 Keras port of torchgeo.losses (ELECTS, QR/RQ)
├── utils/layers.py         shared building blocks (attention, patch embed, conv blocks)
└── weights/                generic PyTorch -> Keras weight-porting framework
    ├── converter.py          WeightConverter, transpose inference, checkpoint loaders
    ├── mappings/              example / reusable name-mapping rule sets
    ├── pretrained.py          timm-style loaders for real, publicly-hosted checkpoints
    └── port_weights.py        CLI entry point + per-model registry

See the Models section for what each family contains, Data for the dataset/sampler/transform framework, Weight Porting for how the conversion framework works, and Pretrained Weights for which models ship a ready-to-download real checkpoint.

Runnable examples

Every example below opens directly in Google Colab, no local setup needed — see Examples for the full list, including a complete finetune a pretrained model walkthrough.

Installation

pip install keras-climate

Then install a Keras 3 backend of your choice:

pip install tensorflow      # or
pip install "jax[cpu]"      # or
pip install torch

See Quick Start for a minimal end-to-end example.