flood-detection-pair-colombia

110 paired Sentinel-2 satellite tile samples (4 PNGs per row) labeled for flood detection across 9 documented flood events in Colombia. Built for fine-tuning a small vision-language model that runs on a satellite or community ground station — see the humaid project.

What's in each sample

Each row in flood_train.jsonl / flood_eval.jsonl is a 4-image vlm_sft example. The user message contains four image content blocks (in this order) followed by a text instruction; the assistant message is the structured JSON labels.

user: [RGB-baseline] [SWIR-baseline] [RGB-current] [SWIR-current] "Label this tile pair..."
assistant: {"flood_present": ..., "flood_severity": ..., ...}

The four images are Sentinel-2 tiles, ~5 km × 5 km:

• RGB-baseline (B4-B3-B2 true color) at a pre-event timestamp (typically 14–28 days before the event)
• SWIR-baseline (B12-B8-B4 false color) at the same timestamp — water absorbs SWIR strongly so flood water reads near-black, healthy vegetation as green, dry soil as pink/magenta
• RGB-current at the timestamp to assess
• SWIR-current at the same timestamp

Why pair input

A town surrounded by chronic ciénaga (the La Mojana wetland baseline) looks identical to a flooded town in any single tile. With the baseline, the labeler can compare and decide whether water increased — well-defined. Without it, single-tile labeling on La Mojana sees ~25–37% labeler disagreement. See the project's labeling docs.

Schema (7 fields)

field	type	meaning
flood_present	bool	new water on land vs baseline
flood_severity	none / minor / moderate / severe	newly inundated land vs baseline (0% / <5% / 5–20% / >20%)
water_coverage_pct_estimate	<10% / 10-30% / 30-60% / >60%	total water in the current view
populated_area_affected	bool	new water near settlements
infrastructure_at_risk	bool	roads / bridges / buildings inundated or threatened
river_overflow_visible	bool	river out of its banks vs baseline
image_quality_limited	bool	abstention signal — clouds / partial coverage

Locations (14)

La Mojana / Depresión Momposina (8): San Jacinto del Cauca, Ayapel, San Benito Abad, Guaranda, Majagual, Caimito, Sucre cabecera, San Marcos.

Putumayo (6): Mocoa, Puerto Asís, Puerto Guzmán, Colón, Santiago, Puerto Leguízamo.

Events (9)

id	date	region	source
cara_de_gato_2021	2021-08-27	La Mojana	dike breach #1
la_mojana_peak_2022	2022-04-17	La Mojana	~590,000 ha flooded
cara_de_gato_2024	2024-05-06	La Mojana	dike breach #2
los_arrastres_2024	2024-05-08	La Mojana	second May 2024 breach
la_mojana_peak_2024	2024-06-11	La Mojana	~860,000 ha — worst peak of 2024
cara_de_gato_2025	2025-08-27	La Mojana	dike breach #3
mocoa_avalancha_2017	2017-04-01	Putumayo	avalancha torrencial, 335 fatalities
puerto_leguizamo_calamidad_2025	2025-04-10	Putumayo	calamidad pública (Río Putumayo)
putumayo_decreto_0472_2025	2025-07-23	Putumayo	departmental calamidad — 16,975 damnificados

For each (location, event) we generated up to 2 samples: (pre, event) and (pre, post).

Splits

• train: 88 samples
• eval: 22 samples (latest 20% by current-window timestamp — temporal split, not random, to avoid Sentinel-2's 5-day revisit leaking near-duplicates across train/eval)

Labeling

Labels were produced by Claude Code agents (Opus class) reading the 4-image input pair and applying a system prompt that includes:

• color semantics for RGB and SWIR false-color views
• change-detection rules (water on land in current that was dry in baseline = flooding)
• severity bands anchored on land-area percentages
• regional context (La Mojana wetland baseline, Bajo Putumayo lowland river floodplains, Mocoa avalancha mechanism)

Inter-labeler agreement (Sonnet vs Opus on subsets) is 0.66–0.68 overall. The hardest fields (flood_severity, water_coverage_pct_estimate) have agreement of 0.43–0.70 — these are subjective enums where two readings of the same image disagree at the boundaries.

Known limitations

1. Mocoa avalancha 2017 has no true pre-event Sentinel-2 acquisitions. S2 archive coverage over Mocoa was sparse before mid-2017. We use a 2017-08-15 stand-in baseline (post-recovery, terrain stable). Annotation flags this with a note field.
2. High proportion of cloud-limited tiles. ~70% of pairs have image_quality_limited=true because La Mojana wet-season cloud cover dominates Sentinel-2. The model is trained to abstain on these.
3. No human verification. Labels are model-generated. Inter-labeler agreement is the only validation signal. To turn this into operational ground truth, ~10% needs human (or remote-sensing-analyst) review.

How it was built

The full data-collection pipeline is in the humaid repo:

• finetune-flood/src/locations.ts, events.ts — declarative locations + event anchors
• finetune-flood/src/generate.ts — Sentinel-2 candidate-date sweep + cloud-aware fetcher (via SimSat)
• finetune-flood/src/label_agents.ts, label_anthropic.ts — pair-aware labeling
• finetune-flood/src/build_dataset.ts — dedupes (location, event, window), splits temporally, writes vlm_sft JSONL

Generated 2026-05-08 as part of the humaid project.

License

CC-BY-4.0 for the labels and dataset structure. The underlying Sentinel-2 imagery is from the European Copernicus programme (free and open data, terms).