A reproducible pseudo-label generation framework for Sentinel-2 imagery using Google Earth Engine, spectral feature engineering, explainable clustering, and ensemble learning for large-scale land-cover mapping.
Remote Sensing • Earth Observation • Geospatial AI • Unsupervised Learning • Google Earth Engine • Pseudo Labeling
Large-scale Earth observation applications require enormous amounts of accurately labeled land-cover data. However, creating pixel-level annotations manually is expensive, time-consuming, and difficult to scale.
This repository presents a reproducible unsupervised land-cover labeling framework developed during Fall 2025 as part of a commercial remote sensing project.
Instead of training a supervised classifier, this work focuses on automatically generating high-quality semantic pseudo-labels directly from Sentinel-2 imagery through spectral analysis, unsupervised learning, and explainable rule-based semantic mapping.
The generated labels are designed to support downstream machine learning tasks including semantic segmentation, land-cover classification, and self-supervised representation learning.
The proposed framework integrates Google Earth Engine processing and local ensemble learning for generating semantic pseudo-labels from Sentinel-2 imagery.
The overall workflow produces both raster-level and polygon-level land-cover representations.
Figure 1. Sentinel-2 unsupervised land-cover clustering result generated using the Google Earth Engine pipeline.
This repository implements the research components developed by the author, including
- End-to-end pseudo-label generation pipeline for Sentinel-2 imagery
- Hybrid Google Earth Engine + local processing workflow
- Extraction of 17 spectral and texture features
- Multi-stage unsupervised clustering
- Explainable semantic cluster interpretation
- Spatial consistency refinement
- Ensemble pseudo-label generation
- GIS-ready raster and polygon outputs
- Fully reproducible experimental workflow
Reliable land-cover datasets remain one of the primary bottlenecks in remote sensing and Earth Observation.
Manual annotation of satellite imagery requires extensive human effort and domain expertise, limiting the scalability of supervised deep learning approaches.
This project investigates whether meaningful semantic labels can be generated automatically by combining
- spectral feature engineering
- vegetation indices
- texture descriptors
- unsupervised clustering
- explainable semantic rules
- lightweight ensemble learning
without requiring manually annotated training data.
sentinel2-unsupervised-landcover-labeling
│
├── notebooks
│ ├── 01_sentinel2_unsupervised_raster_labeling.ipynb
│ └── 02_sentinel2_ensemble_polygon_labeling.ipynb
│
├── figures
│ ├── gee_pipeline/
│ └── ensemble_pipeline/
│
├── docs
│ ├── methodology.md
│ ├── experiments.md
│ └── results.md
│
├── database/
│
│
├── README.md
├── requirements.txt
└── LICENSE
Sentinel-2 Level-2A
│
Cloud Masking
│
Image Composite
│
Feature Engineering
17 Spectral + Texture Features
│
KMeans
│
Gaussian Mixture Model
│
Cluster Evaluation
│
Explainable Semantic Mapping
│
Spatial Smoothing
│
Raster Labels
│
Polygon Generation
│
Ensemble Voting
│
Final Semantic Pseudo Labels
The framework consists of two complementary pipelines:
- Google Earth Engine raster pseudo-label generation
- Local ensemble polygon labeling
The two-stage design enables both pixel-level mapping and GIS-ready vector annotation.
Figure 2. Area of Interest characterization including geometry information, bounding box, centroid and spatial extent.
The first workflow performs large-scale raster labeling entirely inside Google Earth Engine.
Major processing stages include
- AOI definition
- Sentinel-2 compositing
- Cloud masking
- Spectral feature extraction
- Texture extraction
- KMeans clustering
- Gaussian Mixture clustering
- Cluster quality evaluation
- Explainable semantic mapping
- Spatial filtering
- GeoTIFF export
- Earth Engine Asset export
The first pipeline generates raster pseudo-labels directly within Google Earth Engine.
The workflow includes:
- Sentinel-2 compositing
- Feature engineering
- Unsupervised clustering
- Semantic interpretation
- Spatial refinement
Figure 3. Ranking of KMeans and Gaussian Mixture Model clustering configurations using internal validation criteria.
The second workflow operates locally using Sentinel-2 Cloud Optimized GeoTIFFs.
Major components include
- AOI registry
- Composite generation
- Polygon extraction
- Spectral statistics
- Ensemble clustering
- Majority voting
- GeoJSON export
- GeoPackage export
- Interactive HTML visualization
The second pipeline generates GIS-ready vector pseudo-labels by combining multiple unsupervised and lightweight supervised models.
Figure 10. Final ensemble semantic labeling map generated from Sentinel-2 imagery.
The feature engineering stage extracts 17 remote sensing descriptors, including
- NDVI
- NDWI
- NDMI
- NDBI
- NBR
- REIP
- SAVI
- Tasseled Cap Brightness
- Tasseled Cap Greenness
- Tasseled Cap Wetness
- GLCM texture descriptors
These features provide complementary spectral and spatial information for unsupervised semantic clustering.
Figure 4. Spectral index statistics extracted from Sentinel-2 composites across multiple AOIs.
| Algorithm | Purpose |
|---|---|
| KMeans | Spectral clustering |
| Gaussian Mixture Model | Probabilistic clustering |
| Ensemble Voting | Final pseudo-label generation |
Raw cluster IDs are transformed into meaningful semantic land-cover classes using interpretable spectral decision rules based on
- NDVI
- NDWI
- Vegetation Score
This improves both interpretability and reproducibility of the generated labels.
Figure 5. Explainable semantic mapping of spectral clusters into meaningful land-cover categories using NDVI/NDWI-based rules.
Study Area
≈ 938 km²
- Non-Vegetation / Water
- Sparse Vegetation
- Dense Vegetation
Figure 6. Raw KMeans clustering output with pixel statistics and class area estimation.
Figure 7. Spatially refined KMeans result after 3×3 focal mode filtering.
| Class | Area (ha) |
|---|---|
| Non-Vegetation | 59,661.65 |
| Sparse Vegetation | 23,451.85 |
| Dense Vegetation | 5,703.01 |
| Class | Area (ha) |
|---|---|
| Non-Vegetation | 53,264.46 |
| Sparse Vegetation | 30,231.68 |
| Dense Vegetation | 5,558.65 |
Figure 8. Split visualization comparing Sentinel-2 background imagery and GMM classification output.
Figure 9. Final three-class semantic map generated using GMM clustering.
Both clustering approaches produced consistent semantic distributions, with non-vegetated regions dominating the study area and dense vegetation accounting for approximately 6% of the mapped surface.
Final semantic classes
| Class | Area (ha) |
|---|---|
| Non-tree Vegetation | 2,097.80 |
| Non-Vegetation | 232.07 |
| Tree Vegetation | 167.46 |
| Unknown | 26.31 |
| Metric | Value |
|---|---|
| Silhouette Score | 0.585 |
| Davies–Bouldin Index | 0.542 |
Figure 11. PCA projection showing separation between semantic vegetation clusters.
PCA visualization demonstrated good separation between semantic vegetation classes after ensemble refinement.
Representative Sentinel-2 samples from the generated semantic land-cover classes are presented below.
The visualization includes three semantic categories:
- Non-Vegetation
- Non-Tree Vegetation
- Tree Vegetation
Examples of non-vegetated surfaces including bare soil, built-up areas, and water-related regions.
Figure 12. Representative examples of the Non-Vegetation semantic class generated from Sentinel-2 imagery.
Examples of herbaceous vegetation, agricultural fields, and low-height vegetation regions.
Figure 13. Representative examples of the Non-Tree Vegetation semantic class generated from Sentinel-2 imagery.
Examples of dense vegetation areas and tree-dominated regions identified by the ensemble labeling framework.
Figure 14. Representative examples of the Tree Vegetation semantic class generated from Sentinel-2 imagery.
The framework exports
- Google Earth Engine Assets
- GeoTIFF
- GeoJSON
- GeoPackage (GPKG)
- Interactive HTML Maps
ready for visualization in QGIS, ArcGIS, or downstream machine learning pipelines.
- Python
- Google Earth Engine
- Sentinel-2 Level-2A
- NumPy
- Pandas
- Scikit-learn
- Rasterio
- GeoPandas
- Folium
- Matplotlib
- Jupyter Notebook
Clone the repository
git clone https://github.com/hannah-fathi/sentinel2-unsupervised-landcover-labeling.gitInstall dependencies
pip install -r requirements.txtRun the notebooks sequentially
01 → Google Earth Engine Raster Labeling
02 → Local Ensemble Polygon Labeling
- The current framework focuses on vegetation-oriented land-cover classes.
- Semantic mapping relies on handcrafted spectral decision rules.
- Proprietary datasets from the commercial project are excluded.
- Experimental evaluation is limited to selected Sentinel-2 AOIs.
- Temporal land-cover monitoring
- Self-supervised representation learning
- Active learning for pseudo-label refinement
- Integration with semantic segmentation networks
- Foundation Models for Earth Observation
- Diffusion-based pseudo-label enhancement
| Period | Milestone |
|---|---|
| Fall 2025 | Development of the unsupervised labeling framework within a commercial remote sensing project |
| 2026 | Repository refactoring, documentation, and public research release |
This repository documents the research components implemented by the author as part of a larger commercial remote sensing project.
Confidential datasets, proprietary assets, and organization-specific implementation details have been intentionally excluded.
Only reproducible research components are included.
Hannah Fathi
Research Interests
- Remote Sensing
- Geospatial Artificial Intelligence
- Computer Vision
- Earth Observation
- Self-Supervised Learning
- Foundation Models
@misc{fathi2026sentinel2,
author = {Hannah Fathi},
title = {Sentinel-2 Unsupervised Land Cover Labeling},
year = {2026},
publisher = {GitHub},
howpublished = {\url{https://github.com/hannah-fathi/sentinel2-unsupervised-landcover-labeling}}
}This project is released under the MIT License.