Results of Variable Importance of RF Classifier in GEE

Question

I create a RF module in GEE like this:

//Random Forest Result Function Module

exports.getRfResults = function(image, title, bands, trainPoint, testPoint, area, label) {
  
  //Create Sample Points for Training
  var sampleTraining = image.select(bands).sampleRegions({
    collection: trainPoint,
    properties: label,
    scale: 10
  });
  var sampleTrainingTitle = 'sample traingin for ' + title + ' : ';
  print(sampleTrainingTitle, sampleTraining.first());

  //Create Sample Points for Testing
  var sampleTesting = image.select(bands).sampleRegions({
    collection: testPoint,
    properties: label,
    scale: 10
  });
  var sampleTestingTitle = 'sample testing for ' + title + ' : ';
  print(sampleTestingTitle, sampleTesting.first());

  // Classifier for 500 trees
  var classifier = ee.Classifier.smileRandomForest(500,2).setOutputMode('PROBABILITY').train({
    features: sampleTraining,
    classProperty: label[0],
    inputProperties: bands
  });
  var dict = classifier.explain();
  var explainTitle = 'Explain ' + title + ' :';
  print(explainTitle, dict);

  //Variable Importance of RF Classifier
  var variable_importance = ee.Feature(null, ee.Dictionary(dict).get('importance'));

  // Chart of Variable Importance of RF Classifier
  var chartTitle = 'Random Forest Variable Importance for trees for ' + title;
  var chart =
    ui.Chart.feature.byProperty(variable_importance)
      .setChartType('ColumnChart')
      .setOptions({
        title: chartTitle,
        legend: {position: 'none'},
        hAxis: {title: 'Bands'},
        vAxis: {title: 'Importance'}
      });
  print(chart);

  //RF Classifier's Confusion Matrix and Some Values
  var confMatrix = classifier.confusionMatrix();
  var confMatrixTitle = 'Confusion Matrix for '+ title + ' : ';
  print(confMatrixTitle, confMatrix);

  var oaTitle = 'Overall Accuracy for ' + title + ' : ';
  var kappaTitle = 'Kappa for ' + title + ' : ';
  var orderTitle = 'Order for ' + title + ' : ';
  var caTitle = 'Consumers Accuracy for ' + title + ' : ';
  var paTitle = 'Producers Accuracy for ' + title + ' : ';

  print(oaTitle, confMatrix.accuracy());
  print(caTitle, confMatrix.consumersAccuracy());
  print(kappaTitle, confMatrix.kappa());
  print(orderTitle, confMatrix.order());
  print(paTitle, confMatrix.producersAccuracy());

  //var classified = image.classify(classifier);
  //Map.addLayer(classified, {min:0, max:1, palette: ['red', 'green']}, 'classification');

  //RF Classifier's Validation Error Matrix and Some Values
  var tested = sampleTesting.classify(classifier);

  var testErrorMatrix = tested.errorMatrix('isTea', 'classification');
  var vemTitle = 'Validation Error Matrix for ' + title +' : ';
  print(vemTitle, testErrorMatrix);

  var voaTitle = 'Validation Overall Accuracy for ' + title + ' : ';
  var vkappaTitle = 'Validation Kappa for ' + title + ' : ';
  var vorderTitle = 'Validation Order for ' + title + ' : ';
  var vcaTitle = 'Validation Consumers Accuracy for ' + title + ' : ';
  var vpaTitle = 'Validation Producers Accuracy for ' + title + ' : ';

  print(voaTitle, testErrorMatrix.accuracy());
  print(vcaTitle, testErrorMatrix.consumersAccuracy());
  print(vkappaTitle, testErrorMatrix.kappa());
  print(vorderTitle, testErrorMatrix.order());
  print(vpaTitle, testErrorMatrix.producersAccuracy());

  //Classified Image
  var classified = image.classify(classifier);
  Map.addLayer(classified, {min:0, max:1, palette: ['red', 'green']}, 'classification');

  //Pixels Numbers of Classes
  var teaMask = classified.select('classification').eq(1);
  var nonTeaMask = classified.select('classification').eq(0);

  var teaPxNum = ee.Array(classified.updateMask(teaMask).reduceRegion(ee.Reducer.toList(), area, 10).get('classification')).length();
  print('tea Classified PxNum',teaPxNum);

  var nonTeaPxNum = ee.Array(classified.updateMask(nonTeaMask).reduceRegion(ee.Reducer.toList(), area, 10).get('classification')).length();
  print('nonTea Classified PxNum',nonTeaPxNum);

};

But the problem is, which features contribute most to the change between classes? When I explain the classifier with ee.Classifier.explain() method, the output is like this for Sentinel-2;

importance: Object (4 properties)

B2: 17086.790105776927

B3: 17961.828026936393

B4: 17386.157666003204

B8: 20509.20284219093

numberOfTrees: 500

outOfBagErrorEstimate: 0.2574935769340565

trees: List (149 elements)

How do I calculate the importance value and by which method? Because most values I have seen are normalized difference values in the range of 0 to 1. So what is the meaning of these values?

The application for sentinel-2: https://code.earthengine.google.com/79c90e29348b698bae8d094a026584fd

Answer 1

It is the sum of decrease in Gini impurity index over all trees in the forest. From the comments in the code:

 /**
 * Variable importance. Every time a split of a node is made on variable
 * the (GINI, information gain, etc.) impurity criterion for the two
 * descendent nodes is less than the parent node. Adding up the decreases
 * for each individual variable over all trees in the forest gives a fast
 * variable importance that is often very consistent with the permutation
 * importance measure.
 */