Lab 13 - Improving your Classifications

Assignment 1: Collect More Training Points

For the supervised classification, try collecting more points for each class.

• Collect points across the entire composite, not just one location
• Look for pixels of the same class that show variability
• For water: collect pixels in parts of rivers that vary in color
• For developed: collect pixels from different rooftops and materials

Assignment 2: Add Predictor Bands

Usually, the more spectral information you feed the classifier, the easier it is to separate classes.

Add NDVI or NDWI as a new predictor band:

// Calculate NDVI and add to the image
var ndvi = Landsat.normalizedDifference(['SR_B5', 'SR_B4']).rename('NDVI');
var imageWithNDVI = Landsat.addBands(ndvi);

// Update your inputProperties to include NDVI
var inputProperties = ['SR_B2', 'SR_B3', 'SR_B4', 'SR_B5', 'SR_B6', 'SR_B7', 'NDVI'];

Questions to answer in your comments:

• Does adding NDVI help the classification?
• Check developed areas classified as herbaceous (or vice versa)

Assignment 3: Tune Random Forest Trees

Use more trees in the Random Forest classifier:

// Try different numbers of trees
var classifier100 = ee.Classifier.smileRandomForest(100).train({...});
var classifier200 = ee.Classifier.smileRandomForest(200).train({...});

Questions to answer:

• Do you see improvements compared to 50 trees?
• How does computation time change?

Assignment 4: Increase Sample Size (Unsupervised)

Increase the number of samples extracted from the composite:

var training = Landsat.sample({
    region: Landsat.geometry(),
    scale: 30,
    numPixels: 5000,  // Increased from 1000
    tileScale: 8
});

Does increasing samples improve the unsupervised result?

Assignment 5: Increase K Clusters

Increase the number of k clusters:

// Try 10 clusters instead of 4
var clusterer = ee.Clusterer.wekaKMeans(10).train(training);

Questions to answer:

• Does the classified map result in meaningful classes?
• Can multiple clusters be merged into your original 4 classes?

Assignment 6: Test Other Clustering Algorithms

Try other options under the ee.Clusterer object:

// Try X-Means (automatically determines optimal k)
var clusterer = ee.Clusterer.wekaXMeans({
    minClusters: 2,
    maxClusters: 10
}).train(training);

// Try Cascade K-Means
var clusterer = ee.Clusterer.wekaCascadeKMeans({
    minClusters: 2,
    maxClusters: 10
}).train(training);

How do results differ from standard k-means?