Keras-Climate & keras-geo

Multi-backend framework for Earth observation, satellite imagery, climate modeling, and weather nowcasting in Keras 3.

Keras-Climate & keras-geo: Earth Observation in Keras 3

Stack: Keras 3, PyTorch, TensorFlow, JAX, Python
Code: github.com/anas-rz/keras-climate
Docs: anas-rz.github.io/keras-climate/


Overview

Keras-Climate is a modern, backend-agnostic framework built on Keras 3 dedicated to Earth observation, climate modeling, weather forecasting, and remote sensing. It allows scientists and ML engineers to train and evaluate geospatial models seamlessly on PyTorch, TensorFlow, or JAX.


Key Capabilities

  • Multispectral Remote Sensing: Implemented UNet and SegFormer architectures specialized for multi-band satellite data (RGB + Near-Infrared).
  • Long-Term Multivariate Climate Forecasting: Configured PatchTST for multi-variable atmospheric and climate sequences (e.g., input seq_len=336, forecast horizon pred_len=96 across 7 channels).
  • Radar & Precipitation Nowcasting: ConvLSTMNowcaster for spatiotemporal radar and weather nowcasting.
  • Geospatial Foundation Model: PrithviClassifier, integrating the NASA/IBM geospatial foundation model directly into the multi-backend Keras 3 ecosystem with automated weight porting and parity checks.
  • keras-geo: Companion library porting key TorchGeo modules into multi-backend Keras 3 with dedicated ViT weight converters.