Using Public Training Datasets

DeepGEE provides built-in support for using publicly available land cover datasets as training data for deep learning classification. This eliminates the need for manual training data collection and ensures high-quality, scientifically validated training samples.

📊 Available Datasets

1. ESA WorldCover (10m resolution)

Best for: High-resolution Sentinel-2 classification

from deepgee import GEEDataDownloader

downloader = GEEDataDownloader()

# Get training samples from ESA WorldCover
samples = downloader.get_training_from_esa_worldcover(
    roi=[85.0, 20.0, 87.0, 22.0],
    year=2021,  # 2020 or 2021 available
    samples_per_class=500,
    scale=10
)

Class Mapping:

• 10: Tree cover
• 20: Shrubland
• 30: Grassland
• 40: Cropland
• 50: Built-up
• 60: Bare / sparse vegetation
• 70: Snow and ice
• 80: Permanent water bodies
• 90: Herbaceous wetland
• 95: Mangroves
• 100: Moss and lichen

Advantages:

• ✅ High quality (10m resolution)
• ✅ Recent data (2020-2021)
• ✅ 11 detailed classes
• ✅ Global coverage

Use with: Sentinel-2 imagery

---

2. Dynamic World (10m resolution)

Best for: Near real-time classification and time series

# Get training samples from Dynamic World
samples = downloader.get_training_from_dynamic_world(
    roi=[85.0, 20.0, 87.0, 22.0],
    start_date='2023-01-01',
    end_date='2023-12-31',
    samples_per_class=500,
    scale=10
)

Class Mapping:

• 0: Water
• 1: Trees
• 2: Grass
• 3: Flooded vegetation
• 4: Crops
• 5: Shrub and scrub
• 6: Built area
• 7: Bare ground
• 8: Snow and ice

Advantages:

• ✅ Near real-time (updated every 2-5 days)
• ✅ Flexible date ranges
• ✅ 9 classes
• ✅ Based on Sentinel-2

Use with: Sentinel-2 imagery

---

3. MODIS Land Cover (500m resolution)

Best for: Large areas and long time series

# Get training samples from MODIS
samples = downloader.get_training_from_modis_lc(
    roi=[85.0, 20.0, 87.0, 22.0],
    year=2020,
    samples_per_class=300,
    scale=500,
    lc_type=1  # IGBP classification
)

Classification Schemes:

• lc_type=1: IGBP (17 classes) - Default
• lc_type=2: UMD (15 classes)
• lc_type=3: LAI (8 classes)
• lc_type=4: BGC (8 classes)
• lc_type=5: PFT (11 classes)

IGBP Classes (LC_Type1):

1. Evergreen Needleleaf Forests
2. Evergreen Broadleaf Forests
3. Deciduous Needleleaf Forests
4. Deciduous Broadleaf Forests
5. Mixed Forests
6. Closed Shrublands
7. Open Shrublands
8. Woody Savannas
9. Savannas
10. Grasslands
11. Permanent Wetlands
12. Croplands
13. Urban and Built-up Lands
14. Cropland/Natural Vegetation Mosaics
15. Permanent Snow and Ice
16. Barren
17. Water Bodies

Advantages:

• ✅ Long time series (2001-present)
• ✅ Multiple classification schemes
• ✅ Global coverage
• ✅ Well-validated

Use with: Landsat or coarser imagery

---

4. Copernicus Land Cover (100m resolution)

Best for: Medium resolution with detailed forest classes

# Get training samples from Copernicus
samples = downloader.get_training_from_copernicus_lc(
    roi=[85.0, 20.0, 87.0, 22.0],
    year=2019,  # 2015-2019 available
    samples_per_class=300,
    scale=100
)

Class Mapping:

• 0: Unknown
• 20: Shrubs
• 30: Herbaceous vegetation
• 40: Cultivated and managed vegetation/agriculture
• 50: Urban / built up
• 60: Bare / sparse vegetation
• 70: Snow and ice
• 80: Permanent water bodies
• 90: Herbaceous wetland
• 100: Moss and lichen
• 111-116: Closed forest (various types)
• 121-126: Open forest (various types)
• 200: Oceans, seas

Advantages:

• ✅ 23 detailed classes
• ✅ Excellent for forest mapping
• ✅ 100m resolution
• ✅ Global coverage

Use with: Landsat imagery

---

🚀 Complete Workflow Example

import deepgee
from deepgee import GEEDataDownloader, LandCoverClassifier
import pandas as pd

# 1. Initialize
deepgee.initialize_gee(project='your-project-id')
downloader = GEEDataDownloader()

# 2. Define area
roi = [85.0, 20.0, 87.0, 22.0]

# 3. Get training samples from ESA WorldCover
training_samples = downloader.get_training_from_esa_worldcover(
    roi=roi,
    year=2021,
    samples_per_class=500
)

# 4. Download Sentinel-2 composite
composite = downloader.create_composite(
    roi, '2021-01-01', '2021-12-31', sensor='sentinel2'
)

downloader.download_image(composite, 'data.tif', roi=roi, scale=10)

# 5. Extract features
training_data = downloader.extract_training_samples(
    composite, training_samples, scale=10, output_path='training.csv'
)

# 6. Train model
samples = pd.read_csv('training.csv')
feature_cols = ['B2', 'B3', 'B4', 'B8', 'B11', 'B12',
                'NDVI', 'EVI', 'NDWI', 'NDBI', 'NBR', 'NDMI', 'NDBaI']

X = samples[feature_cols].values
y = samples['class'].values

classifier = LandCoverClassifier(n_classes=11)
classifier.build_model(input_shape=(len(feature_cols),))

X_train, X_test, y_train, y_test = classifier.prepare_data(X, y)
history = classifier.train(X_train, y_train, epochs=100)

# 7. Evaluate
results = classifier.evaluate(X_test, y_test)
print(f"Accuracy: {results['accuracy']:.2%}")
print(f"Kappa: {results['kappa']:.3f}")

---

📋 Dataset Comparison

Dataset	Resolution	Classes	Time Range	Best For
ESA WorldCover	10m	11	2020-2021	High-res Sentinel-2
Dynamic World	10m	9	2015-present	Real-time, time series
MODIS	500m	17	2001-present	Large areas, long series
Copernicus	100m	23	2015-2019	Forest mapping

---

💡 Best Practices

Choosing the Right Dataset

For Sentinel-2 classification:

# Use ESA WorldCover (best quality)
samples = downloader.get_training_from_esa_worldcover(roi, year=2021)

# OR Dynamic World (for specific dates)
samples = downloader.get_training_from_dynamic_world(
    roi, '2023-01-01', '2023-12-31'
)

For Landsat classification:

# Use MODIS (for large areas)
samples = downloader.get_training_from_modis_lc(roi, year=2020)

# OR Copernicus (for detailed forest classes)
samples = downloader.get_training_from_copernicus_lc(roi, year=2019)

Sample Size Guidelines

# Small study area (<100 km²)
samples_per_class = 300

# Medium study area (100-1000 km²)
samples_per_class = 500

# Large study area (>1000 km²)
samples_per_class = 1000

Scale Matching

Always match the scale of training data to your imagery:

# Sentinel-2 (10m)
samples = downloader.get_training_from_esa_worldcover(roi, scale=10)
composite = downloader.create_composite(roi, sensor='sentinel2')
downloader.download_image(composite, 'data.tif', scale=10)

# Landsat (30m)
samples = downloader.get_training_from_modis_lc(roi, scale=30)
composite = downloader.create_composite(roi, sensor='landsat8')
downloader.download_image(composite, 'data.tif', scale=30)

---

🎯 Advantages of Public Datasets

1. No Manual Digitization

• ✅ Save time - no need to manually create training polygons
• ✅ Immediate availability
• ✅ Global coverage

2. High Quality

• ✅ Scientifically validated
• ✅ Consistent methodology
• ✅ Expert-reviewed

3. Reproducibility

• ✅ Same training data for everyone
• ✅ Standardized classes
• ✅ Documented methodology

4. Scalability

• ✅ Works anywhere in the world
• ✅ Easy to generate thousands of samples
• ✅ Automated workflow

---

⚠️ Limitations & Considerations

Temporal Mismatch

If your imagery date doesn't match the reference dataset:

# Your imagery: 2023
# ESA WorldCover: 2021

# Solution: Use Dynamic World for matching dates
samples = downloader.get_training_from_dynamic_world(
    roi, '2023-01-01', '2023-12-31'
)

Class Mapping

Different datasets have different class schemes:

# ESA WorldCover: 11 classes (10, 20, 30, ...)
# Dynamic World: 9 classes (0, 1, 2, ...)
# MODIS: 17 classes (1, 2, 3, ...)

# Always check class values in your training data!
import pandas as pd
samples = pd.read_csv('training.csv')
print(samples['class'].unique())

Resolution Mismatch

Match dataset resolution to your imagery:

# ✓ Good: Sentinel-2 (10m) + ESA WorldCover (10m)
# ✓ Good: Landsat (30m) + MODIS (500m resampled to 30m)
# ✗ Poor: Sentinel-2 (10m) + MODIS (500m) - too coarse

---

📚 References

ESA WorldCover

• Zanaga, D., et al. (2021). ESA WorldCover 10 m 2020 v100.
• https://doi.org/10.5281/zenodo.5571936

Dynamic World

• Brown, C.F., et al. (2022). Dynamic World, Near real-time global 10 m land use land cover mapping.
• Scientific Data, 9(1), 251.

MODIS Land Cover

• Friedl, M., & Sulla-Menashe, D. (2019). MCD12Q1 MODIS/Terra+Aqua Land Cover Type Yearly L3 Global 500m SIN Grid V006.
• https://doi.org/10.5067/MODIS/MCD12Q1.006

Copernicus Land Cover

• Buchhorn, M., et al. (2020). Copernicus Global Land Service: Land Cover 100m: collection 3: epoch 2019: Globe.
• https://doi.org/10.5281/zenodo.3939050

---

🎓 Next Steps

1. Try the example: Run examples/public_dataset_training.py
2. Compare datasets: Test different datasets for your area
3. Optimize: Adjust sample sizes and parameters
4. Validate: Check accuracy against ground truth

---

Copyright © 2026 Pulakesh Pradhan