A high-performance, model-agnostic geospatial inference pipeline for gigapixel satellite imagery. GSIP enables seamless deep learning integration on standard hardware through a memory-efficient chunking engine, flexible batch processing, and a unified CLI/GUI ecosystem.

The gsip command provides access to four specialized tools designed for the full lifecycle of Earth Observation inference:

• 🚀 gsip infer: The core high-performance engine. Processes massive Sentinel-2/1 tiles using ERF-aware tiling and dynamic memory auto-configuration to prevent OOM errors on any hardware.
• 📊 gsip suite: A powerful batch orchestrator. Run complex experiments using Cartesian product job generation (Multi-Model × Multi-Input) with hierarchical config overrides and automatic GPU cooldowns.
• 🎨 gsip studio: A native GTK4 dashboard. Features an interactive Visual Editor for batch runs, a System Info hub, and detailed Post-Run Performance Analysis (GPU/RAM usage charts).
• 🔧 gsip manage: Developer-centric CLI for extending the pipeline. Quickly scaffold and register new model adapters or output reporters.

# Dependencies
sudo apt install libcairo2-dev libgirepository-2.0-dev

# Recommended: Isolated install via pipx
pipx install .

# Or standard development install
pip install -e .

# 1. Run inference on a single tile
gsip infer model=resnet_s2 input_path=/path/to/S2_tile.SAFE output_path=./out

# 2. Launch the GUI for analysis or batch building
gsip studio

# 3. Run a batch of jobs
gsip suite --config my_batch.json

GSIP is designed to be highly extensible. You can integrate any PyTorch-based model (CNNs, ViTs, Foundation Models) by writing a Model Adapter.

👉 Check out the Extending Guide to learn how to add your own models.