As far as I understand your question you're looking for some methods for feature selection and reduction before actually building your model, is this correct?

In case this is what you want, a basic approach is to compute the Gini index for each feature given a dataset, which basically evaluates the importance of a feature (https://medium.com/analytics-steps/understanding-the-gini-index-and-information-gain-in-decision-trees-ab4720518ba8).

Another very basic method is simply to compute the variance for all features and remove all features below a certain threshold.

For a more general overview see here: https://en.wikipedia.org/wiki/Feature_selection

There are also some Python packages for this which are part of scikit: https://scikit-learn.org/stable/modules/feature_selection.html

Not sure about an equivalent in R though.

In case this sin't what you're looking you can try to further specify what you need.

If you use a random forest algorithm and wanna see feature importance, you just print out the following line.

var classifier = ee.Classifier.smileRandomForest(tree).train({

features: trainingPartition,

classProperty: 'landcover',

inputProperties: bands,

});

print(classifier.explain()); // this explain the feature importance.

Hope this helps you out.

Duong Cao Phan thank you for your suggestion! I also made it work and adjusted your code a bit. I also included a table export by casting the dictionary in a ee.Feature. There is also a bar plot included.

//Get the feature importance

// Get the band names from the input dataset (image)

var bandNames = image.bandNames();

print(bandNames, 'band names');

// Set up the classifier

var classifier_importance = ee.Classifier.smileRandomForest(500).train({ //random forest with 500 trees

features: training,

classProperty: 'Class',

inputProperties: bandNames,

});

// get a dictionary with the feature importance of the classifier

var dict_featImportance = classifier_importance.explain();

print(dict_featImportance, 'dict_featImportance');

// Cast into a ee.Feature and a ee.FeatureCollection for exporting the importance as CSV

var variable_importance = ee.Feature(null, ee.Dictionary(dict_featImportance).get('importance'));

var variable_importance01 = ee.FeatureCollection(variable_importance);

// Export the FeatureCollection as CSV

Export.table.toDrive({

collection: variable_importance01,

description: 'variable_importance_subset01',

fileFormat: 'CSV'

});

// Plot the importance of each band in a bar plot

var chart =

ui.Chart.feature.byProperty(variable_importance)

.setChartType('ColumnChart')

.setOptions({

title: 'Random Forest Variable Importance',

legend: {position: 'none'},

hAxis: {title: 'Bands'},

vAxis: {title: 'Importance'}

});

print(chart);

I hope this is useful for some users. Im not sure about the format of the feature collection though. Maybe that could be improved.

Best,

Marcel

Marcel Mohr Thanks for posting this. It's helpful and the scripts worked. Unfortunately, I got different variable importances if I did random forest classification in R by using the identical tree number. Did you try using R and any thoughts on this? Thanks in advance!

Luwen Wan, I think it might be due to different default settings of the Mtry parameter in GEE and R. It might a little bit higher in R. Im not sure but its worth to check.

Marcel Mohr Thanks, dear Marcel! Yeah, I checked the Mtry as it's one of the user-defined parameters along with the number of trees. I got a higher OOB estimate error in R (higher than 0.3) while it reported only about 0.19 in GEE. Unfortunately, I am not sure how to change the number of Mtry in GEE, the default is the squre root of the number of input variables.

Dear Luwen, this entry from the GEE API may help:

ee.Classifier.smileRandomForest(numberOfTrees, variablesPerSplit, minLeafPopulation, bagFraction, maxNodes, seed)⨯

Creates an empty Random Forest classifier.

Arguments:

numberOfTrees (Integer):

The number of decision trees to create.

variablesPerSplit (Integer, default: null):

The number of variables per split. If unspecified, uses the square root of the number of variables.

minLeafPopulation (Integer, default: 1):

Only create nodes whose training set contains at least this many points.

bagFraction (Float, default: 0.5):

The fraction of input to bag per tree.

maxNodes (Integer, default: null):

The maximum number of leaf nodes in each tree. If unspecified, defaults to no limit.

seed (Integer, default: 0):

The randomization seed.

Returns: Classifier

the second argument is the Mtry parameter (variablesPerSplit) and can be set as stated above.

Kind regards, Marcel

Marcel Mohr Dear Marcel. Thank you so much for your continuous response and help. Yeah, I noticed these in the ee.Classifier.smileRandomForest Docs as well and it helps. Thanks again!