What they dump, we find.
Pelago finds water-adjacent dump sites — illegal or unregulated dumping grounds near rivers, lakes, and coastlines — directly from Sentinel-2 satellite imagery. Using a deployable, two-stage machine-learning pipeline, Pelago screens entire regions from orbit, flags candidate sites that sit or drain toward water, and tracks them over time.
Every candidate is run through manual validation before it becomes a confirmed, monitored site. The result is a continuously updated inventory of the places where plastic and waste enter waterways — with the evidence to act on it.
Pelago ships a live web map with region views, site-level detail, and a monitoring timeline. Representative screens below.
| Screen | Caption |
|---|---|
 |
Globe overview — every validated Pelago site, plotted against population footprint and river networks. |
 |
Site detail — a single confirmed site with its detection footprint, contours, and Sentinel-2 context. |
 |
Confidence layers — pixel-classifier heatmap and patch-classifier intersections rendered as map layers. |
 |
Monitoring timeline — contour and spectral histories for sites under recurring observation. |
 |
Alert feed — new detections and expanded sites stacked into a reviewable, prioritized feed. |
If you are looking for the repository's original map exports, they live under docs/ and cover Bali, Indonesia, Java, Albania, Sri Lanka, Vietnam, the Philippines, and more.
Plastic doesn't start in the ocean — it starts on land, next to water.
| 🏞️ Rivers are highways | An estimated majority of ocean plastic enters through a small number of river systems. Whatever is dumped near a river gets carried downstream. |
| 🗑️ Dumpsites are unmonitored | Most of these sites are informal, unregulated, or simply off the map. Nobody logs them, and governments don't have an inventory. |
| 🛰️ Satellites already see it | Sentinel-2 revisits the entire land surface every few days at 10 m resolution. The imagery exists — it just isn't looked at systematically. |
| 📋 Manual surveys don't scale | Field audits and local reports are valuable but cannot cover a whole watershed, let alone a continent, at the pace of the problem. |
Pelago closes that gap by turning open satellite data into an operationally useful inventory of water-adjacent waste sites.
Pelago finds water-adjacent plastic dump sites from Sentinel-2 imagery using a two-stage machine-learning pipeline.
First, a pixel classifier scores every pixel for the spectral signature of waste. Then a patch classifier — trained on weakly-labeled examples at scale — confirms whether those pixel detections form a real dump-site pattern. The intersection of both stages produces candidate sites, which are manually validated and then monitored over time with contour and metadata analysis.
The output is an API-backed, map-first product: worldwide detection coverage, per-site history, and a validation workflow that keeps false positives out of the confirmed inventory.
flowchart TD
subgraph INGESTION["🛰️ INGESTION"]
A1["Sentinel-2<br/>10 m RGB + multispectral"]
A2["Descartes Labs Catalog"]
A3["Population-weighted<br/>tile generation"]
A1 --> A2 --> A3
end
subgraph MODEL["🧠 MODEL LAYER"]
B1["Pixel Classifier<br/>spectral waste signature"]
B2["Patch Classifier<br/>weakly-labeled ensemble over 28×28×24"]
B3["Intersection Filter<br/>pixel ∩ patch agreement"]
B1 --> B2 --> B3
end
subgraph DETECT["📍 SITE DETECTION"]
C1["Candidate Generation<br/>blob detection on scored tiles"]
C2["Manual Validation<br/>analyst imagery review"]
C3["Confirmed Sites<br/>confirmed · industrial · uncertain · negative"]
C1 --> C2 --> C3
end
subgraph META["📊 METADATA + MONITORING"]
D1["Contour Generation<br/>per-site extent over time"]
D2["Metadata Enrichment<br/>centroids · addresses via Nominatim"]
D3["Pelago API"]
D1 --> D2 --> D3
end
subgraph PRESENT["🌐 PRESENTATION"]
E1["Web Map<br/>Next.js + MapLibre"]
E2["Alerts"]
E3["Monitoring Dashboard"]
E4["Open GeoJSON / CSV Export"]
E5["API Access"]
end
A3 --> B1
B3 --> C1
C3 --> D1
D3 --> E1
D3 --> E2
D3 --> E3
D3 --> E4
D3 --> E5
style INGESTION fill:#0a1929,stroke:#00d4ff,stroke-width:2px,color:#f5f7fa
style MODEL fill:#0a1929,stroke:#ffb020,stroke-width:2px,color:#f5f7fa
style DETECT fill:#0a1929,stroke:#00e0ff,stroke-width:2px,color:#f5f7fa
style META fill:#0a1929,stroke:#ff6b6b,stroke-width:2px,color:#f5f7fa
style PRESENT fill:#0a1929,stroke:#00d4ff,stroke-width:2px,color:#f5f7fa
1. Ingestion. The pipeline pulls Sentinel-2 scenes through Descartes Labs, constrained to a population-weighted tile grid so compute goes where people and waste actually are.
2. Model layer. A pixel classifier scores the spectral signature of waste across each tile. A patch classifier, trained with weakly-labeled data and ensembled across seeds, decides whether a scored region is genuinely a dump site. Candidates require agreement between both stages — dramatically reducing false positives.
3. Site detection. Scored and filtered regions become candidate sites. Each candidate passes through analyst manual validation before being labeled a confirmed site.
4. Metadata + monitoring. Confirmed sites get contours (boundary and area over time) plus enriched metadata (centroid, address, catchment context). Everything is pushed to the Pelago API.
5. Presentation. The web application renders sites on an interactive map with confidence layers, monitoring timelines, and alert feeds — and exposes the inventory for export and programmatic access.
- 🌍 Global satellite coverage — powered by open Sentinel-2 data with a population-limited detection footprint.
- 🧠 Two-stage ML pipeline — pixel classifier → patch classifier → intersection filter for high-precision detection.
- 🎓 Weakly-supervised training — patch labels mined at scale from spectral + spatial priors rather than hand-annotated imagery.
- 🎯 Confidence-tiered detections — confirmed, industrial, uncertain, and negative categories keep the inventory honest.
- 🔭 Continuous monitoring — confirmed sites are revisited and contoured over time to track expansion and new activity.
- 🗂️ Open data export — candidate sites, validated points, and metadata as standard GeoJSON/CSV.
- 🔌 API access — a loader service pushes model outputs into the Pelago API for the web app and integrations.
- 🛰️ Sentinel-2 native — detections are grounded in real 10 m multispectral observations, not modeled estimates.
| Layer | Technology | Purpose |
|---|---|---|
| Satellite data | Copernicus Sentinel-2 | 10 m multi-spectral imagery of the full land mass |
| Cloud imagery access | Descartes Labs | Bulk scene search, download, and scalable inference |
| Modeling | Python · TensorFlow · scikit-learn | Pixel classifier, patch classifier, feature pipelines |
| Geospatial | GeoPandas · Rasterio · Shapely · PyProj | Vector/contour processing and geometry handles |
| MLOps / Deploy | Descartes Deploy endpoints |
Distributed candidate detection and contour runs |
| Backend API | Go · Pelago loader (loader/) |
Ingests model outputs, serves site data |
| Frontend | Next.js · MapLibre · TypeScript | Interactive globe, site detail, monitoring dashboard |
| Data formats | GeoJSON · CSV · HDF5 | Interchange between models, validation, and API |
| Module | Inputs | Role |
|---|---|---|
create_pixel_dataset · create_spectrogram_dataset |
Raw Sentinel-2 tiles | Assemble per-pixel training chips and temporal spectrograms |
train_pixel_classifier · train_spectrogram_classifier |
Pixel-level crops | Learn the spectral waste signature; emit a temporal pixel model |
train_patch_classifier |
Weak labels + pixel scores | Learn site-level structure; ensemble, SVM, and 1 px variants trained at scale |
Notebooks: create_pixel_dataset.ipynb · create_spectrogram_dataset.ipynb · train_pixel_classifier.ipynb · train_spectrogram_classifier.ipynb · Train Patch Classifier (Weak Labeling, Ensemble/SVM/1px/LARGE).ipynb
Trained artifacts live under models/, versioned by release (e.g., v0.0.7, v0.0.11, ensemble families).
| Module | Role |
|---|---|
generate_populated_dltiles |
Compute a population-weighted tile grid for a region |
descartes_spectrogram_run_withpop |
Deploy pixel + patch inference over the region on Descartes |
descartes_candidate_detect |
Run blob detection to extract candidate sites from scored tiles |
query_patch_classifier |
Keep only candidates that pass the patch-classifier intersection |
validate_candidate_sites |
Analyst review of candidates → confirmed / industrial / uncertain / negative |
Notebooks: generate_populated_dltiles.ipynb · descartes_spectrogram_run_withpop.ipynb · descartes_candidate_detect.ipynb · query_patch_classifier.ipynb · validate_candidate_sites.ipynb
Validated sites are stored under data/sampling_locations/ and candidate detections under data/model_outputs/candidate_sites/.
| Module | Role |
|---|---|
generate_metadata |
Enrich confirmed sites with centroids, addresses, and context |
descartes_contour_run |
Generate per-site contours on Descartes (boundary and area over time) |
loader |
Push enriched model outputs into the Pelago API for the web app |
Notebooks: generate_metadata.ipynb · descartes_contour_run.ipynb
Contours and metadata are served through the API rather than stored in the repository — see loader/README.md for the ingestion pipeline.
Pelago's pipeline has been exercised across South and Southeast Asia, the Mediterranean, Africa, and South America. Numbers below reflect the current operating inventory:
| Metric | Value |
|---|---|
| Confirmed positive sites | 4,700+ validated across operating regions |
| Labeled validation dataset | 19,900+ features (positive / negative / industrial / uncertain) |
| Largest tracked sites | Top 100 largest sites by area, maintained in-site inventory |
| Regional exports | 38+ interactive maps under docs/ |
| Spatial resolution | 10 m (Sentinel-2) |
| Detection footprint | Population-weighted global tile grid |
| Validation workflow | Analyst-reviewed, confidence-tiered |
| Refresh cadence | Recurring on each Sentinel-2 revisit cycle |
The paper describes the approach in detail and includes quantitative validation of the pixel and patch classifiers.
git clone https://github.com/Flowthread/Pelago.git
cd Pelago
python -m venv env
source env/bin/activate
pip install -r requirements.txtPython ≥ 3.7 is supported.
Imports are relative to the repository root, which must be on PYTHONPATH:
export PYTHONPATH=/path/to/pelago:$PYTHONPATHThe bulk-processing pipeline runs on Descartes Labs:
descarteslabs auth loginFollow the link to enter your email and password. If needed, see the Descartes authentication docs.
# Train or use a released model (models/)
python -m scripts.deploy_nn_v0 # deploy pixel inference
python -m scripts.candidate_detect # generate candidate sites
python -m scripts.deploy_query_patch # run patch-classifier intersection
python -m scripts.contour_gen # generate contours for confirmed sites
# Push validated sites to the API
cd loader && python load.pyEach stage is also available as a runnable pipeline notebook under notebooks/ and produces standard GeoJSON/CSV outputs under data/.
Pelago/
├── api/ # Backend API (Go)
├── assets/ # Brand assets, logo, pipeline diagrams
├── data/
│ ├── boundaries/ # Region boundary files
│ ├── model_outputs/ # Candidate sites, contours, patch results
│ ├── sampling_locations/ # Validated site inventories (GeoJSON)
│ └── site_metadata/ # Site-level metadata & largest-sites export
├── docs/
│ ├── screenshots/ # Product screens for the README gallery
│ └── *.html # Regional map exports
├── frontend/ # Next.js + MapLibre web application
├── loader/ # Model-output → API loader (Go + Python)
├── models/ # Versioned model artifacts
├── notebooks/ # Runnables for each pipeline stage
├── paper/plos/ # Manuscript source (LaTeX/PDF)
├── scripts/ # Core Python modules (dl_utils, nn_predict, ...)
├── validation/ # Candidate validation tooling
├── .github/workflows/ # CI (lint, build, test on ./api)
├── requirements.txt
└── README.md
- Two-stage pixel → patch detection pipeline
- Operational candidate generation on Descartes Labs
- Analyst validation workflow and confidence-tiered inventory
- Contour generation and site-level monitoring
- API ingestion loader and open GeoJSON/CSV export
- Region expansions: South & Southeast Asia, Mediterranean, Africa, South America
- Expanding detection to new river basins and coastal watersheds
- Adding higher-cadence revisit monitoring for high-risk sites
- Integrating additional Sentinel-2 band products for finer material discrimination
- Publishing a public API endpoint for alert subscriptions
The Pelago software is released under the MIT License (see LICENSE).
- Aug–Sep 2026 — Built for NextStep Hacks 2026 (Earth Forward)
- Submission — September 2026
Flowthread — GitHub Built for NextStep Hacks 2026.
Built for NextStep Hacks 2026 — Earth Forward 🌱