Why does adding tileScale to GEE script does not run the supervised classification?

Question

I am performing a supervised classification with GEE on a region in Italy and I face the error: Output of image computation is too large (15 bands for 931221 pixels = 99.5 MiB > 80.0 MiB). If this is a reduction, try specifying a larger 'tileScale' parameter.

This is the part of the code that I think makes the problem because before these lines, I did not face errors:

//Set the input image composite
var input = ee.Image(mean_Summer_IC);
//Set the band combinations
var bands = ['B1', 'B2', 'B3', 'B4','B5','B6','B7', 'B8', 'B8A', 'B9' ,'B11', 'B12','NDVI', 'EVI', /*'GNDVI',*/ 'NBR'/*, 'NDII'*/];
//Create training data
var training_Supervised = input.select(bands).sampleRegions({
  collection: Training_Points,
  properties: ['land_class'],
  scale:10
});

//RandomForest classification approach
//Create the RF_classifier
var RF_classifier = ee.Classifier.smileRandomForest({
  numberOfTrees: 500,
  variablesPerSplit: 1, //null means default value. In this case the rootsquare of the number of variables
  minLeafPopulation: 1,
  bagFraction: 1,
  maxNodes: null, //null means default value. In this case "no limits" of nodes
  seed: 0,
});

//Train the classifier
var classifier01 = RF_classifier.train({
  features: training_Supervised,
  classProperty: 'land_class',
  inputProperties: bands
});

//Run the classifier
var RF_classified = input.select(bands).classify(classifier01);
print(RF_classified);

var Palette = [
  
  'aec3d4', //  Water
  'cc0013', // Residential
  'cdb33b', // Agricultural
  'd9903d', // Arbusti
  'c3aa69', // BoschiMisti
  '30eb5b', //Latifoglie
  '152106', //Conifere
  'f7e084'  //BareSoil
 ];
 

//Show classification results
Map.addLayer(RF_classified, {min: 1, max: 8, palette: Palette},'RF_classification');

// Get a confusion matrix representing resubstitution accuracy.
print('RF error matrix: ', classifier01.confusionMatrix());
print('RF accuracy: ', classifier01.confusionMatrix().accuracy());

//Show classification results
Map.addLayer(RF_classified, {min: 1, max: 8, palette: Palette},'RF_classification');

// Get a confusion matrix representing resubstitution accuracy.
print('RF error matrix: ', classifier01.confusionMatrix());
print('RF accuracy: ', classifier01.confusionMatrix().accuracy());

What I tried to do based on the error itself was to specify a tilescale (I tried values 2, 4, 8, and 16), but it did not solve the problem. Link to my script: https://code.earthengine.google.com/578e87ba2a48ce51b2892ffbbf5cdb5c?accept_repo=users%2Fessepratico%2FPratico_et_al_RemoteSensing_2021

Is there anyway to run this script? Thanks in advance.

Answer 1

There are several factors that make your code heavy, including parameters for scale, amount of bands, tilescale, size of ICs etc.

I tested and was able to run your version with scale set to 120: https://code.earthengine.google.com/e18afb397b4954c6d841b7d7d5a2238a?accept_repo=users%2Fessepratico%2FPratico_et_al_RemoteSensing_2021

Also I experimented with using less bands, and was able to run at a scale of 50 without bands 1,6,7,8A.

You might want to see where you can reduce the amount of computing necessary to run the script. If you want to operate at a scale of 10 I would suggest to divide the region in smaller subsets.