Sentinel-2 Crop Classification with SVM in Google Earth Engine

Why Sentinel-2 + SVM for Crop Mapping?

Sentinel-2 provides free 10 m multispectral imagery with a 5-day revisit time — ideal for capturing the phenological dynamics that distinguish crop types. Support Vector Machines remain competitive with deep learning methods when training data is limited, which is often the case in data-scarce regions like Sudan.

In my FAO project (2023–2024), SVM on Sentinel-2 time-series achieved 94.2% overall accuracy for wheat classification across the Gezira Scheme — outperforming Random Forest by 2.3% with far less training data.

Building the Image Collection

The key is capturing multiple phenological stages — sowing, peak vegetation, and senescence:

// Google Earth Engine script
var gezira = ee.FeatureCollection('users/osmangeomatics/gezira_boundary');
var startDate = '2023-10-01';
var endDate = '2024-04-30';

// Load Sentinel-2 Surface Reflectance with cloud masking
function maskS2clouds(image) {
  var qa = image.select('QA60');
  var cloudBitMask = 1 << 10;
  var cirrusBitMask = 1 << 11;
  var mask = qa.bitwiseAnd(cloudBitMask).eq(0)
    .and(qa.bitwiseAnd(cirrusBitMask).eq(0));
  return image.updateMask(mask)
    .divide(10000)
    .copyProperties(image, ['system:time_start']);
}

var s2 = ee.ImageCollection('COPERNICUS/S2_SR_HARMONIZED')
  .filterDate(startDate, endDate)
  .filterBounds(gezira)
  .filter(ee.Filter.lt('CLOUDY_PIXEL_PERCENTAGE', 20))
  .map(maskS2clouds);

print('Images in collection:', s2.size());

Engineering Spectral Features

Rather than using individual bands, compute vegetation indices per dekad and stack them:

function addIndices(image) {
  var ndvi = image.normalizedDifference(['B8', 'B4']).rename('NDVI');
  var lswi = image.normalizedDifference(['B8', 'B11']).rename('LSWI');  // moisture
  var evi = image.expression(
    '2.5 * ((NIR - RED) / (NIR + 6 * RED - 7.5 * BLUE + 1))',
    { NIR: image.select('B8'), RED: image.select('B4'), BLUE: image.select('B2') }
  ).rename('EVI');
  var ndwi = image.normalizedDifference(['B3', 'B8']).rename('NDWI');
  return image.addBands([ndvi, lswi, evi, ndwi]);
}

// Compute monthly composites (median) across the growing season
var months = ee.List.sequence(10, 4); // Oct–Apr
var monthlyComposites = months.map(function(m) {
  var month = ee.Number(m);
  var year = month.lte(12) ? 2023 : 2024;
  return s2.filter(ee.Filter.calendarRange(month, month, 'month'))
    .map(addIndices)
    .median()
    .set('month', month);
});

// Stack all monthly composites into a single multi-band image
var featureStack = ee.ImageCollection(monthlyComposites)
  .toBands()
  .clip(gezira);

print('Total feature bands:', featureStack.bandNames().size());
// Result: 7 months × 4 indices = 28 features + original bands

Training the SVM Classifier

// Load training samples (field-collected ground truth)
var trainingPoints = ee.FeatureCollection('users/osmangeomatics/gezira_training_2024');

// Sample the feature stack at training locations
var training = featureStack.sampleRegions({
  collection: trainingPoints,
  properties: ['class_id'],  // 0=wheat, 1=other crops, 2=fallow, 3=water
  scale: 10,
  geometries: true,
});

// Split 70/30 for training and validation
var withRandom = training.randomColumn('random');
var trainSet = withRandom.filter(ee.Filter.lt('random', 0.7));
var testSet  = withRandom.filter(ee.Filter.gte('random', 0.7));

// Train SVM with RBF kernel
var classifier = ee.Classifier.libsvm({
  kernelType: 'RBF',
  gamma: 0.5,
  cost: 10,
}).train({
  features: trainSet,
  classProperty: 'class_id',
  inputProperties: featureStack.bandNames(),
});

// Classify the full image
var classified = featureStack.classify(classifier);

Accuracy Assessment

// Apply classifier to test set
var testClassified = testSet.classify(classifier);

// Confusion matrix
var confMatrix = testClassified.errorMatrix('class_id', 'classification');
print('Overall Accuracy:', confMatrix.accuracy());
print('Kappa Coefficient:', confMatrix.kappa());
print('Confusion Matrix:', confMatrix);

// Per-class accuracy
print('Producers Accuracy:', confMatrix.producersAccuracy());
print('Users Accuracy:', confMatrix.consumersAccuracy());

Our results on the Gezira dataset:

| Class | Producer's Acc. | User's Acc. | |-------|----------------|-------------| | Wheat | 96.1% | 93.8% | | Other Crops | 91.4% | 94.2% | | Fallow | 95.7% | 96.1% | | Water | 99.2% | 99.0% | | Overall | 94.2% | — |

Exporting Results

// Export classified map to Google Drive
Export.image.toDrive({
  image: classified.toByte(),
  description: 'Gezira_Wheat_Map_2024',
  folder: 'GEE_exports',
  region: gezira.geometry(),
  scale: 10,
  crs: 'EPSG:32636',  // UTM Zone 36N
  maxPixels: 1e13,
});

Lessons Learned

1. LSWI is critical for wheat detection in irrigated areas — it captures soil moisture differences between irrigated wheat and dry fallow much better than NDVI alone
2. Dekadal composites > monthly for capturing the fast phenological window of short-season wheat (~120 days)
3. Cost parameter tuning matters significantly — test Cost values of 1, 10, 100 before settling
4. Minimum 50 samples per class for reliable SVM training in GEE; below this, accuracy degrades rapidly

I've open-sourced the complete GEE script and the accuracy assessment tool (GeoAccuRate) that implements the Olofsson et al. (2014) methodology for unbiased area estimation.

Feel free to contact me if you're working on crop mapping in arid/semi-arid irrigated systems.