Institutional repository containing data, high-performance analytical pipelines (DuckDB / Python), and an interactive web platform developed under the agreement between WRI Brasil and FUNAPE, with technical execution by LAPIG/UFG in cooperation with Embrapa Cenargen (Project GR000153-73077-CARGILL-BRAZIL).
Deliverable: Product 1 — Detailed Sampling Plan and Sampling Improvement Strategies (August 17, 2026).
The platform integrates 62,009 candidate samples (in 2025) across a multi-year historical series (2020 to 2025), matched via 64-dimensional L2-renormalized cosine similarity against Embrapa's 701 in situ reference polygons (~1 ha; 608 classified across 7 pasture condition categories + 93 unlabeled):
| Sample Series | Description & Stratification | Samples / Year (2025) | Provenance & Nature |
|---|---|---|---|
| 50k Series | Stratified random sampling from persistent pasture (2020–2025) | 50,000 | Proportional to 2025 vigor strata: Low (~35%), Medium (~44%), High (~21%) |
| 12k Series | MapBiomas Col. 4 S2 (5,453) + MapBiomas 85k Landsat validation (6,556) | 12,009 | Visual inspection pasture samples & interpreted validation subset |
| Total Candidates | Inclusive universe before similarity, SOM, and CBERS-4A filters | 62,009 | Scenarios: Inclusive (62,009) vs. Agreement with Col. 11 (56,952) |
| Embrapa Reference | In situ field reference polygons (~1 ha) | 701 | 608 classified across 7 categories + 93 unclassified |
- 2019: 12,483 points
- 2020: 12,386 points
- 2021: 12,281 points
- 2022: 12,149 points
- 2023: 12,080 points
- 2024: 12,034 points
- 2025: 12,009 points
Built upon a lightweight, modular static architecture (compatible with Jekyll and GitHub Pages Source: /docs) featuring browser GPU acceleration via Leaflet Canvas Markers and dynamic charts via Chart.js:
- Home Page (
/): Institutional overview, objectives, key figures, and section navigation. - Md3 Analysis (
/analise-top3/):- Unified Filter Panel: Instant switching between the 50k Series and the 12k Series, year selection (2019 to 2025), pasture vigor filtering (CVP 1, 2, 3), Md3 similarity thresholds (0.00 to 0.95), full-text search, and typology checkboxes.
- Spatial Bounding Box Selection: Tool to interactively select any custom geographic bounding box across the Cerrado, automatically updating counters, charts, and sample inspection tables.
- Synchronized Dual Charts: Chart 1 (Dynamic MapBiomas sample distribution in the active view) and Chart 2 (Fixed distribution of Embrapa's 701 ground-truth points).
- Detailed Table: Inspection of Top-3 dot products, reference FIDs, target coordinates, and assigned typologies.
- Static Deliverables: Direct downloads of technical Reports, Presentation decks, GEE Assets, Python scripts, and Open Data tables in Parquet, CSV, and JSON formats.
- SOM (Self-Organizing Maps): Section dedicated to unsupervised topological neural network clustering.
lapig-wribrasil/
│
├── 🌐 docs/ # Web Platform (Jekyll / GitHub Pages: /docs)
│ ├── _config.yml # Central Jekyll configuration (baseurl: "/lapig-wribrasil")
│ ├── _layouts/ # Base layouts (default.html, page.html)
│ ├── _includes/ # Modular HTML includes (map_viz, Produtos_Estaticos, etc.)
│ ├── assets/ # Stylesheets, JavaScript engine, and JSON/CSV datasets
│ ├── material_suplementar/ # Technical Report (PDF) & Presentation (PDF)
│ ├── analise-top3/ # Md3 Analysis & Deliverables page
│ ├── comparacao/ # Scenario comparison page
│ └── index.html # Homepage
│
├── 📊 dados/ # Data Pipelines & Processing Engine
│ ├── arquivos_base/ # Raw Sentinel-2 embeddings, Embrapa points, MapBiomas inputs
│ ├── arquivos_saida/ # Processed Top-3 Parquets (50k & 12k Series · 2019–2025)
│ └── scripts/ # Python / DuckDB processing pipelines
│
├── 📁 documentos/ # Reports, Slide Decks, QGIS Styles & Maintenance Guides
│ ├── Report_Product_1_*.pdf # Technical Report PDF
│ ├── Scaling_Ground_Truth_*.pdf # Executive Presentation PDF
│ ├── campo_md3.pptx # Field Presentation Deck
│ ├── qgis_cores.qml # QGIS layer styling palette
│ └── procedimentos_manutencao/ # Maintenance & configuration documentation
│
├── 🛠️ servidor_local.py # Local testing server (serves docs/ in real time)
├── 📄 LICENSE # Creative Commons Attribution 4.0 (CC BY 4.0)
└── 📘 README.md # Repository presentation and documentation
To preview the web platform locally in real time:
python servidor_local.pyOpen in your browser: http://localhost:8000 or http://localhost:8000/analise-top3/.
python dados/scripts/gerar_top3_pivotada_50k.py --year allpython dados/scripts/gerar_top3_pivotada_11k.py --year allpython dados/scripts/exportar_referencia_embrapa.pyGround-truth samples and spectral similarity predictions are categorized into 7 primary classes:
- Productive Pasture (
#faea40): Well-managed pasture with high forage biomass. - Pasture with Weeds (
#d8ff6c): Infestation of herbaceous/ruderal invasive plants. - Pasture with Shrubs / Woody Species (
#66c600): Significant presence of shrub and tree canopy layers. - Intermediate Pasture (
#f4b346): Moderate vegetative vigor with mixed grass and soil coverage. - Biological Degradation (
#813209): Severe degradation with prominent bare soil. - Natural Regeneration (
#0e5f0e): Areas undergoing natural recovery of native Cerrado vegetation. - Miscellaneous (
#ec2b10): Atypical, transitional, or mixed land-use features.
- LAPIG / UFG — Image Processing and Geoprocessing Laboratory (Federal University of Goiás)
- WRI Brasil — World Resources Institute Brasil
- Embrapa — Brazilian Agricultural Research Corporation
This work, including all code, datasets, documentation, and web assets, is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).