""" GLCM Texture Feature Template for Google Earth Engine Gray-Level Co-occurrence Matrix texture features for land cover classification Usage: Copy and adapt these functions for your Session 2 notebook """ import ee def add_glcm_texture(image, bands=['B8'], radius=1): """ Add GLCM texture features to an image Parameters: ----------- image : ee.Image Input Sentinel-2 image bands : list Bands to compute texture from (default: NIR band B8) radius : int Neighborhood radius in pixels (default: 1 = 3x3 window) Returns: -------- ee.Image with added texture bands Texture Features: - asm: Angular Second Moment (uniformity) - contrast: Local variation - corr: Correlation between pixel pairs - var: Variance - idm: Inverse Difference Moment (homogeneity) - savg: Sum Average - svar: Sum Variance - sent: Sum Entropy - ent: Entropy (randomness) - dvar: Difference Variance - dent: Difference Entropy - imcorr1, imcorr2: Information Measures of Correlation """ # Select band for texture calculation gray = image.select(bands) # Compute GLCM with specified radius # size = (radius * 2) + 1 glcm = gray.glcmTexture(size=radius * 2 + 1) # Return image with GLCM bands added return image.addBands(glcm) def add_selected_glcm(image, bands=['B8'], features=['contrast', 'ent', 'corr']): """ Add only selected GLCM texture features (faster computation) Parameters: ----------- image : ee.Image Input Sentinel-2 image bands : list Bands to compute texture from features : list Which texture features to keep Options: 'contrast', 'corr', 'ent', 'var', 'idm', 'asm' Returns: -------- ee.Image with selected texture bands """ # Compute all GLCM features glcm = add_glcm_texture(image, bands=bands) # Build list of band names to select texture_bands = [] for band in bands: for feature in features: texture_bands.append(f'{band}_{feature}') # Select only requested features selected = glcm.select(texture_bands) return image.addBands(selected) def glcm_for_classification(image, nir_band='B8', red_band='B4'): """ Add optimized GLCM features for land cover classification Uses NIR for vegetation texture and Red for overall contrast Parameters: ----------- image : ee.Image Input Sentinel-2 image nir_band : str Near-infrared band name (default: 'B8') red_band : str Red band name (default: 'B4') Returns: -------- ee.Image with classification-optimized texture features """ # NIR texture (good for vegetation structure) nir_texture = image.select(nir_band).glcmTexture(size=3) nir_contrast = nir_texture.select(f'{nir_band}_contrast').rename('nir_texture_contrast') nir_entropy = nir_texture.select(f'{nir_band}_ent').rename('nir_texture_entropy') nir_corr = nir_texture.select(f'{nir_band}_corr').rename('nir_texture_corr') # Red texture (good for overall image structure) red_texture = image.select(red_band).glcmTexture(size=3) red_contrast = red_texture.select(f'{red_band}_contrast').rename('red_texture_contrast') # Combine selected features texture_features = ee.Image.cat([ nir_contrast, nir_entropy, nir_corr, red_contrast ]) return image.addBands(texture_features) # Example usage in GEE workflow def example_usage(): """ Example: Add texture features to Sentinel-2 image """ # Define area of interest (Palawan example) aoi = ee.Geometry.Rectangle([118.5, 9.5, 119.5, 10.5]) # Load Sentinel-2 image s2 = ee.ImageCollection('COPERNICUS/S2_SR') \ .filterBounds(aoi) \ .filterDate('2024-01-01', '2024-04-30') \ .filter(ee.Filter.lt('CLOUDY_PIXEL_PERCENTAGE', 20)) \ .median() # Method 1: Add all GLCM features (slow but comprehensive) # image_with_all_texture = add_glcm_texture(s2, bands=['B8'], radius=1) # Method 2: Add selected features (faster) image_with_selected = add_selected_glcm( s2, bands=['B8'], features=['contrast', 'ent', 'corr'] ) # Method 3: Add optimized features for classification (recommended) image_with_optimized = glcm_for_classification(s2) # Print band names to verify print('Optimized texture bands:', image_with_optimized.bandNames().getInfo()) return image_with_optimized # Tips for using GLCM in land cover classification: """ GLCM TIPS & BEST PRACTICES: 1. COMPUTATIONAL COST: - GLCM is computationally expensive - Use smaller windows (3x3 or 5x5) for faster processing - Select only needed features (contrast, entropy, correlation) - Avoid computing on large areas at once 2. WHICH FEATURES TO USE: - Contrast: Distinguishes forest from agriculture - Entropy: Captures urban heterogeneity - Correlation: Good for textured surfaces (forest canopy) - Variance: Similar to contrast but different scale 3. WHICH BANDS: - NIR (B8): Best for vegetation texture - Red (B4): Good for overall structure - SWIR (B11): Useful for urban/bare soil - Avoid using too many bands (slows computation) 4. WINDOW SIZE: - 3x3 (radius=1): Good for 10m Sentinel-2 - 5x5 (radius=2): Captures more context but slower - 7x7 (radius=3): Too large for 10m data usually 5. WHEN TO USE: - Separating primary vs secondary forest (canopy texture) - Urban vs bare soil (heterogeneity) - Mangroves vs water (structural complexity) 6. WHEN NOT TO USE: - Spectral differences are sufficient - Time constraints (GLCM is slow) - Very heterogeneous landscapes (texture less meaningful) 7. OPTIMIZATION: - Compute texture on median composite, not each image - Use .aside() to print progress during development - Consider pre-computing and exporting for large areas - Use .reproject() carefully to avoid memory issues 8. TROUBLESHOOTING: - "Computation timed out": Reduce area or window size - "Memory limit exceeded": Process in tiles - "Too slow": Select fewer features or smaller radius - Results look strange: Check band scaling (0-1 vs 0-10000) """ # Advanced: Multi-scale texture def multiscale_glcm(image, band='B8', radii=[1, 2]): """ Compute GLCM at multiple scales Useful for capturing texture at different spatial frequencies WARNING: Very computationally expensive! """ texture_features = [] for radius in radii: glcm = image.select(band).glcmTexture(size=radius * 2 + 1) # Select key features and rename with scale contrast = glcm.select(f'{band}_contrast').rename(f'texture_contrast_r{radius}') entropy = glcm.select(f'{band}_ent').rename(f'texture_entropy_r{radius}') texture_features.extend([contrast, entropy]) return image.addBands(texture_features)