GeeFlow

GeeFlow is a library developed by Google DeepMind for generating and processing large-scale geospatial datasets using Google Earth Engine (GEE). It provides utilities, configurations, and pipeline launch scripts specifically tailored for creating geospatial datasets. The primary focus of GeeFlow is to support geospatial AI research, enabling the creation of datasets that conform to the TFDS format for direct use with tf.data.Dataset pipelines. It is designed for research and exploration rather than as a production-ready utility.

This library can be utilized for creating datasets of varying scales, both supervised and unsupervised, which are ready for ingestion into geospatial AI model training, complete with precomputed statistics. GeeFlow also facilitates the creation of inference maps, capable of reaching global scales at any desired resolution. It supports any type of geospatial satellite and remote sensing data, along with label data available from Google Earth Engine, offering arbitrary spatial and temporal resolution and data sampling.

While GeeFlow excels at data generation, it intentionally remains out of scope for model training and inference, interactive GEE visualization, or serving as a datasets repository. For these purposes, it recommends integrating with other specialized frameworks and tools such as Jeo for Jax/Flax, TorchGeo for PyTorch, or geemap for Python-based GEE analysis. GeeFlow prioritizes reproducibility, versioning, scalability, and efficiency in data generation.