What is zamba-sat?

Zamba-sat is a data generation and labeling pipeline for detecting deforestation in the Congo Basin using satellite imagery, built for the AI in Space hackathon run by DPhi Space and Liquid AI. It is written in Python. The Congo Basin is the second largest tropical rainforest on Earth and a major carbon sink, but it faces ongoing pressure from logging road expansion, slash and burn agriculture, artisanal mining, and infrastructure encroachment near park boundaries. Manual satellite review does not scale to the size of the area or the revisit schedule needed to catch new clearings while they are still small enough to act on. The goal of this project is to train a compact vision language model that could eventually run directly on a satellite, targeting a small model called Liquid LFM2.5-VL-450M, so that only the relevant frames and labels need to be sent down to the ground instead of full imagery. The pipeline fetches pairs of Sentinel-2 satellite images taken roughly 90 days apart from fixed Congo Basin sites. For each location it retrieves four frames: two natural color images and two SWIR, or short wave infrared, images, one set from the past and one from the current period. SWIR reveals vegetation moisture and bare ground that is hard to tell apart in ordinary visible light, especially under haze. A larger oracle vision language model, Claude Opus, reads the four frames and returns a structured chain of thought response using labeled XML sections rather than plain JSON, describing each frame, analyzing what changed, and assigning a final label. A Pydantic schema validates each response before it is accepted. The possible labels are stable, clearing, expansion, regrowth, or cloud artifact, with the cloud artifact label used deliberately when cloud cover makes a judgment unsafe to make. The resulting labeled dataset has 72 samples across three sites in the Democratic Republic of Congo, Yangambi, Kindu, and Lusambo, and is published publicly on Hugging Face for others to use.