Array Transformations

Earth Engine supports array transformations such as transpose, inverse and pseudo-inverse. As an example, consider an ordinary least squares (OLS) regression of a time series of images. In the following example, an image with bands for predictors and a response is converted to an array image, then “solved” to obtain least squares coefficients estimates three ways. First, assemble the image data and convert to arrays:

Code Editor (JavaScript)

// Scales and masks Landsat 8 surface reflectance images.
function prepSrL8(image) {
  // Develop masks for unwanted pixels (fill, cloud, cloud shadow).
  var qaMask = image.select('QA_PIXEL').bitwiseAnd(parseInt('11111', 2)).eq(0);
  var saturationMask = image.select('QA_RADSAT').eq(0);

  // Apply the scaling factors to the appropriate bands.
  var opticalBands = image.select('SR_B.').multiply(0.0000275).add(-0.2);
  var thermalBands = image.select('ST_B.*').multiply(0.00341802).add(149.0);

  // Replace the original bands with the scaled ones and apply the masks.
  return image.addBands(opticalBands, null, true)
      .addBands(thermalBands, null, true)
      .updateMask(qaMask)
      .updateMask(saturationMask);
}

// Load a Landsat 8 surface reflectance image collection.
var collection = ee.ImageCollection('LANDSAT/LC08/C02/T1_L2')
  // Filter to get only two years of data.
  .filterDate('2019-04-01', '2021-04-01')
  // Filter to get only imagery at a point of interest.
  .filterBounds(ee.Geometry.Point(-122.08709, 36.9732))
  // Prepare images by mapping the prepSrL8 function over the collection.
  .map(prepSrL8)
  // Select NIR and red bands only.
  .select(['SR_B5', 'SR_B4'])
  // Sort the collection in chronological order.
  .sort('system:time_start', true);

// This function computes the predictors and the response from the input.
var makeVariables = function(image) {
  // Compute time of the image in fractional years relative to the Epoch.
  var year = ee.Image(image.date().difference(ee.Date('1970-01-01'), 'year'));
  // Compute the season in radians, one cycle per year.
  var season = year.multiply(2 * Math.PI);
  // Return an image of the predictors followed by the response.
  return image.select()
    .addBands(ee.Image(1))                                  // 0. constant
    .addBands(year.rename('t'))                             // 1. linear trend
    .addBands(season.sin().rename('sin'))                   // 2. seasonal
    .addBands(season.cos().rename('cos'))                   // 3. seasonal
    .addBands(image.normalizedDifference().rename('NDVI'))  // 4. response
    .toFloat();
};

// Define the axes of variation in the collection array.
var imageAxis = 0;
var bandAxis = 1;

// Convert the collection to an array.
var array = collection.map(makeVariables).toArray();

// Check the length of the image axis (number of images).
var arrayLength = array.arrayLength(imageAxis);
// Update the mask to ensure that the number of images is greater than or
// equal to the number of predictors (the linear model is solvable).
array = array.updateMask(arrayLength.gt(4));

// Get slices of the array according to positions along the band axis.
var predictors = array.arraySlice(bandAxis, 0, 4);
var response = array.arraySlice(bandAxis, 4);

Note that arraySlice() returns all the images in the time series for the range of indices specified along the bandAxis (the 1-axis). At this point, matrix algebra can be used to solve for the OLS coefficients:

Code Editor (JavaScript)

// Compute coefficients the hard way.
var coefficients1 = predictors.arrayTranspose().matrixMultiply(predictors)
  .matrixInverse().matrixMultiply(predictors.arrayTranspose())
    .matrixMultiply(response);

Although this method works, it is inefficient and makes for difficult to read code. A better way is to use the pseudoInverse() method (matrixPseudoInverse() for an array image):

Code Editor (JavaScript)

// Compute coefficients the easy way.
var coefficients2 = predictors.matrixPseudoInverse()
  .matrixMultiply(response);

From a readability and computational efficiency perspective, the best way to get the OLS coefficients is solve() (matrixSolve() for an array image). The solve() function determines how to best solve the system from characteristics of the inputs, using the pseudo-inverse for overdetermined systems, the inverse for square matrices and special techniques for nearly singular matrices:

Code Editor (JavaScript)

// Compute coefficients the easiest way.
var coefficients3 = predictors.matrixSolve(response);

To get a multi-band image, project the array image into a lower dimensional space, then flatten it:

Code Editor (JavaScript)

// Turn the results into a multi-band image.
var coefficientsImage = coefficients3
  // Get rid of the extra dimensions.
  .arrayProject([0])
  .arrayFlatten([
    ['constant', 'trend', 'sin', 'cos']
]);

Examine the outputs of the three methods and observe that the resultant matrix of coefficients is the same regardless of the solver. That solve() is flexible and efficient makes it a good choice for general purpose linear modeling.