ee.FeatureCollection.errorMatrix

コレクションの 2 つの列（実際の値を含む列と予測値を含む列）を比較して、コレクションの 2D エラー行列を計算します。値は 0 から始まる小さな連続した整数であることが想定されています。行列の軸 0（行）は実際の値に対応し、軸 1（列）は予測値に対応します。

用途	戻り値
FeatureCollection.errorMatrix(actual, predicted, order)	ConfusionMatrix

引数	タイプ	詳細
これ: collection	FeatureCollection	入力コレクション。
actual	文字列	実際の値を含むプロパティの名前。
predicted	文字列	予測値を含むプロパティの名前。
order	リスト、デフォルト: null	期待値のリスト。この引数を指定しない場合、値は連続しており、0 から maxValue までの範囲に及ぶとみなされます。指定した場合、このリストに一致する値のみが使用され、行列にはこのリストに一致するディメンションと順序が設定されます。

例

コードエディタ（JavaScript）

/**
 * Classifies features in a FeatureCollection and computes an error matrix.
 */

// Combine Landsat and NLCD images using only the bands representing
// predictor variables (spectral reflectance) and target labels (land cover).
var spectral =
    ee.Image('LANDSAT/LC08/C02/T1_L2/LC08_038032_20160820').select('SR_B[1-7]');
var landcover =
    ee.Image('USGS/NLCD_RELEASES/2016_REL/2016').select('landcover');
var sampleSource = spectral.addBands(landcover);

// Sample the combined images to generate a FeatureCollection.
var sample = sampleSource.sample({
  region: spectral.geometry(),  // sample only from within Landsat image extent
  scale: 30,
  numPixels: 2000,
  geometries: true
})
// Add a random value column with uniform distribution for hold-out
// training/validation splitting.
.randomColumn({distribution: 'uniform'});
print('Sample for classifier development', sample);

// Split out ~80% of the sample for training the classifier.
var training = sample.filter('random < 0.8');
print('Training set', training);

// Train a random forest classifier.
var classifier = ee.Classifier.smileRandomForest(10).train({
  features: training,
  classProperty: landcover.bandNames().get(0),
  inputProperties: spectral.bandNames()
});

// Classify the sample.
var predictions = sample.classify(
    {classifier: classifier, outputName: 'predicted_landcover'});
print('Predictions', predictions);

// Split out the validation feature set.
var validation = predictions.filter('random >= 0.8');
print('Validation set', validation);

// Get a list of possible class values to use for error matrix axis labels.
var order = sample.aggregate_array('landcover').distinct().sort();
print('Error matrix axis labels', order);

// Compute an error matrix that compares predicted vs. expected values.
var errorMatrix = validation.errorMatrix({
  actual: landcover.bandNames().get(0),
  predicted: 'predicted_landcover',
  order: order
});
print('Error matrix', errorMatrix);

// Compute accuracy metrics from the error matrix.
print("Overall accuracy", errorMatrix.accuracy());
print("Consumer's accuracy", errorMatrix.consumersAccuracy());
print("Producer's accuracy", errorMatrix.producersAccuracy());
print("Kappa", errorMatrix.kappa());

Python の設定

Python API とインタラクティブな開発での geemap の使用については、 Python 環境のページをご覧ください。

import ee
import geemap.core as geemap

Colab（Python）

from pprint import pprint

# Classifies features in a FeatureCollection and computes an error matrix.

# Combine Landsat and NLCD images using only the bands representing
# predictor variables (spectral reflectance) and target labels (land cover).
spectral = ee.Image('LANDSAT/LC08/C02/T1_L2/LC08_038032_20160820').select(
    'SR_B[1-7]')
landcover = ee.Image('USGS/NLCD_RELEASES/2016_REL/2016').select('landcover')
sample_source = spectral.addBands(landcover)

# Sample the combined images to generate a FeatureCollection.
sample = sample_source.sample(**{
    # sample only from within Landsat image extent
    'region': spectral.geometry(),
    'scale': 30,
    'numPixels': 2000,
    'geometries': True
    })
# Add a random value column with uniform distribution for hold-out
# training/validation splitting.
sample = sample.randomColumn(**{'distribution': 'uniform'})
print('Sample for classifier development:', sample.getInfo())

# Split out ~80% of the sample for training the classifier.
training = sample.filter('random < 0.8')
print('Training set:', training.getInfo())

# Train a random forest classifier.
classifier = ee.Classifier.smileRandomForest(10).train(**{
    'features': training,
    'classProperty': landcover.bandNames().get(0),
    'inputProperties': spectral.bandNames()
    })

# Classify the sample.
predictions = sample.classify(
    **{'classifier': classifier, 'outputName': 'predicted_landcover'})
print('Predictions:', predictions.getInfo())

# Split out the validation feature set.
validation = predictions.filter('random >= 0.8')
print('Validation set:', validation.getInfo())

# Get a list of possible class values to use for error matrix axis labels.
order = sample.aggregate_array('landcover').distinct().sort()
print('Error matrix axis labels:')
pprint(order.getInfo())

# Compute an error matrix that compares predicted vs. expected values.
error_matrix = validation.errorMatrix(**{
    'actual': landcover.bandNames().get(0),
    'predicted': 'predicted_landcover',
    'order': order
    })
print('Error matrix:')
pprint(error_matrix.getInfo())

# Compute accuracy metrics from the error matrix.
print('Overall accuracy:', error_matrix.accuracy().getInfo())
print('Consumer\'s accuracy:')
pprint(error_matrix.consumersAccuracy().getInfo())
print('Producer\'s accuracy:')
pprint(error_matrix.producersAccuracy().getInfo())
print('Kappa:', error_matrix.kappa().getInfo())