geosave-engine

GeoSave Engine is a local-first product for building geospatial AI workflows end to end. It standardizes the full path from data acquisition, environment setup, model training, and prediction to serving-ready outputs, so teams do not need to reinvent a different workflow for every project.

It generates a ready-to-use boilerplate and applies proven best practices out of the box, including access to state-of-the-art models and multiple training methods with minimal coding. Instead of building model pipelines from scratch, users can focus on dataset creation and preprocessing, then run the resulting pipeline on fresh satellite data directly from their own machine.

Visit official Documentation : https://weedkat.github.io/geosave-engine/

Features

• Geospatial data pipeline — GeoAnchor/GeoTile/GeoStack model location+time, fetched pixels, and multi-layer samples. Pull from a live STAC catalog (Copernicus, Planetary Computer, Element84, or any self-hosted endpoint) or local GeoTIFF, derive layers (cloud masks, NDVI, labels), save to disk as .zarr stores or stream straight into prediction with no disk round trip.
• Training, config-only — SemanticSegmentationTask + SemanticSegmentationDataModule cover plain supervised segmentation entirely from a LightningCLI YAML config, no Python to write. A pipeline's own per-sample context (e.g. a Prithvi/Clay encoder's real acquisition time/location) wires straight in via one config field. Drop to a hand-written LightningModule when you need full control.
• Pretrained model registry — encoders (DINOv3, Prithvi, Prithvi-TL, Clay), decoders (DPT, UNet), heads, selected by registry key, chained together automatically, no manual import wiring or hand-glued forward pass.
• Sensor-aware band metadata — wavelength/GSD/mean/std per sensor (Sentinel-2, Landsat, MODIS, more), feeding model config directly (Clay's wavelength conditioning, normalization stats) — a geodata concern, not hardcoded into any model.
• MLflow model registry integration — geosave upload rebuilds a trained model from its checkpoint and registers it, ready to serve.
• Editable scaffolding, not a framework lock-in — geosave create hands you real, editable files. No required base class your code has to obey to keep working.

Installations

Requires Python 3.12+.

pip install geosave-engine
# or
uv add geosave-engine

Want the rolling dev build (rebuilt on every push to main) instead of the latest stable tag:

pip install --pre --index-url https://test.pypi.org/simple/ \
  --extra-index-url https://pypi.org/simple/ geosave-engine

Working on GeoSave Engine itself (clone + uv sync), or installing an exact dev build off a GitHub release — see docs/guide/installation.md.

Quick Start

uv run geosave create -d my-project
cd my-project
# fill in .env with your CDSE (or other STAC provider) credentials

Then follow docs/guide/workflow.md for the full step-by-step — explore a pipeline, build a dataset, train, register.

Generated Workspace

my-project/
├── artifacts/     # checkpoints, logs, saved configs (created by training)
├── configs/       # LightningCLI YAML configs
├── data/          # ingested layers land here
├── logs/
├── modules/       # your pipeline (Path A); data module + lightning module too, if Path B
├── predictions/
├── .env           # CDSE credentials, filled in with placeholders
├── geosave.toml   # workspace identity (task/method/catalog), read by the CLI
└── main.py        # LightningCLI entry point — do not need to touch this

Development Workflow

TO BE ADDED LATER