glody007/zamba-sat

# zamba-sat Congo Basin deforestation detection — data generation and labeling pipeline for the **AI in Space** hackathon (DPhi Space × Liquid AI). ## Problem The Congo Basin is the second-largest tropical rainforest on Earth and one of the planet's largest terrestrial carbon sinks. It is also under sustained pressure from logging-road expansion, slash-and-burn agriculture, artisanal mining, and infrastructure encroachment along park boundaries. Manual satellite review by rangers and analysts does not scale to the area or the revisit cadence needed to catch clearings while they are still small enough to act on. We want a small vision-language model that can run **on board a satellite** (target: Liquid `LFM2.5-VL-450M`) and emit a structured deforestation annotation directly from a pair of Sentinel-2 acquisitions, so only the relevant frames and labels need to be downlinked. That requires a labeled dataset of Congo Basin two-frame change pairs that the small model can be fine-tuned on. ## Approach Two-frame temporal change detection over fixed Congo Basin sites: - For each (location, spatial sub-tile, temporal sample) we fetch four Sentinel-2 frames via SimSat: | frame | bands | role | |---|---|---| | `rgb_t1.png` | B4-B3-B2 (red, green, blue) | natural colour, ~30 days ago | | `swir_t1.png` | B11-B8A-B4 (swir16, nir08, red) | vegetation moisture, ~30 days ago | | `rgb_t0.png` | B4-B3-B2 | natural colour, current | | `swir_t0.png` | B11-B8A-B4 | vegetation moisture, current | RGB shows what a human sees; SWIR makes vegetation moisture and bare/burnt ground pop, which is what catches recent clearings under partial haze. ### What the model sees Two real samples from `dry_season/salonga_north_drc`. Each row is one (location, sub-tile, temporal-sample) — four frames, one label. **Stable forest** (`change_pattern: stable`) — `dry_season/salonga_north_drc/s02_t00` | RGB t-1 (~30 d ago) | RGB t-0 …