Different min and max NDVI values with different approaches using Google Earth Engine

Question

I'm trying to calculate the vegetation fraction for land surface temperature calculations of a single Landsat 8 image. I need to find the min and max NDVI of the scene.

I've masked the image and calculated NDVI with two different methods:

var ndvi1 = image.normalizedDifference(['B5','B4']);
var ndvi2 = nir.subtract(red).divide(nir.add(red));

But when I calculate the min and max NDVI values, the results are different for these two methods.

ndvi1 min: -0.33634447455585664
  ndvi1 max: 1
  
  ndvi2 min: -94.81818181818181
  ndvi2 max: 63.333333333333336

What is the source of this difference? How should I revise my code for the accurate calculation of min and max NDVI values?

Here is my code:

Code Editor script link

var raw2 = ee.Image('LANDSAT/LC08/C01/T1_SR/LC08_174034_20161117');

var vizParams1 = {
  bands: ['B4', 'B3', 'B2'],
  min: 0,
  max: 3000,
  gamma: 1.4,
};
Map.addLayer(raw2, vizParams1, 'raw2');

// Cloud mask of a single Landsat 8 image.
{
  // Bits 3 and 5 are cloud shadow and cloud, respectively.
  var cloudShadowBitMask = ee.Number(2).pow(3).int();
  var cloudsBitMask = ee.Number(2).pow(5).int();
  var snowBitMask = ee.Number(2).pow(4).int();
  var waterBitMask = ee.Number(2).pow(2).int();
  // Get the QA band.
  var qa = raw2.select('pixel_qa');
  // Both flags should be set to zero, indicating clear conditions.
  var mask = qa.bitwiseAnd(cloudShadowBitMask).eq(0).and(
            qa.bitwiseAnd(cloudsBitMask).eq(0)).and(
            qa.bitwiseAnd(snowBitMask).eq(0)).and(
            qa.bitwiseAnd(waterBitMask).eq(0));
  var image =
      // Scale the data to reflectance and temperature.
      //.select(['B5', 'B7'], ['NIR', 'SWIR']).multiply(0.0001)
      raw2.updateMask(mask);
}

var vizParams2 = {
  bands: ['B4', 'B3', 'B2'],
  min: 0,
  max: 3000,
  gamma: 1.4,
};
Map.addLayer(image, vizParams2, 'Landsat');

print(image,'Landsat');

var nir = image.select(['B5']);
var red = image.select(['B4']);
var ndvi = nir.subtract(red).divide(nir.add(red));
//.rename('NDVI');

var NDVI = image.normalizedDifference(['B5','B4']);

var ndviParams = {min: -1, max: 1, palette: ['blue', 'white', 'green']};

Map.addLayer(NDVI, ndviParams, 'NDVI image');

Map.addLayer(ndvi, ndviParams, 'ndvi image');

//Calculating min and max values:
var min = ee.Number(NDVI.reduceRegion({
     reducer: ee.Reducer.min(),
     scale: 30,
     maxPixels: 1e9
     }).values().get(0));
  print(min, 'NDVImin');
var max = ee.Number(NDVI.reduceRegion({
      reducer: ee.Reducer.max(),
     scale: 30,
     maxPixels: 1e9
     }).values().get(0));
  print(max, 'NDVImax');

var min = ee.Number(ndvi.reduceRegion({
     reducer: ee.Reducer.min(),
     scale: 30,
     maxPixels: 1e9
     }).values().get(0));
  print(min, 'ndvi-min');
var max = ee.Number(ndvi.reduceRegion({
      reducer: ee.Reducer.max(),
     scale: 30,
     maxPixels: 1e9
     }).values().get(0));
  print(max, 'ndvi-max');

Justin Braaten · Answer

The normalizedDifference() function clamps inputs to be non-negative, i.e., negative values are first set to 0 before evaluating. For most applications using normalizedDifference() is the best practice. The same behavior can be achieved with your band math method by adding .max(0) to the lines that define the red and nir variables.

The fact that there are differences in the result of min and max calculation indicate that the Landsat RED and NIR bands have some negative values (however, reflectance should not be negative, so it may be some artifact of USGS processing).

I was curious to know how different the results of various methods were, so I did an analysis where NDVI is calculated three different ways and then subtracted one from the other and counted the number of pixels that had unequal NDVI values. It turns out that the two "manual" expression methods have 947 pixels that are different (0.002749 percent of valid image pixels) - a tiny fraction.

Here is the analysis script for posterity:

Code Editor script

var image = ee.Image('LANDSAT/LC08/C01/T1_SR/LC08_174034_20161117');

var cloudShadowBitMask = ee.Number(2).pow(3).int();
var cloudsBitMask = ee.Number(2).pow(5).int();
var snowBitMask = ee.Number(2).pow(4).int();
var waterBitMask = ee.Number(2).pow(2).int();
var qa = image.select('pixel_qa');
var mask = qa.bitwiseAnd(cloudShadowBitMask).eq(0).and(
  qa.bitwiseAnd(cloudsBitMask).eq(0)).and(
  qa.bitwiseAnd(snowBitMask).eq(0)).and(
  qa.bitwiseAnd(waterBitMask).eq(0));
image = image.updateMask(mask);

var nir = image.select(['B5']);
var red = image.select(['B4']);
var ndvi1 = nir.subtract(red).divide(nir.add(red)).rename('NDVI');
var ndvi2 = image.normalizedDifference(['B5','B4']).rename('NDVI');
var ndvi3 = image.expression(
  '((NIR - RED) / (NIR + RED))',
  {
    'NIR': image.select('B5'),
    'RED': image.select('B4')
  })
  .rename('NDVI');

function calcMinMax(img, title) {
  var stats = img.reduceRegion({
     reducer: ee.Reducer.minMax(),
     scale: 30,
     maxPixels: 1e13});
  print(title+':',
    ee.String('Min: ').cat(ee.Algorithms.String(stats.get('NDVI_min'))),
    ee.String('Max: ').cat(ee.Algorithms.String(stats.get('NDVI_max'))));
}

calcMinMax(ndvi1, 'chained expression');
calcMinMax(ndvi2, '.normalizedDifference()');
calcMinMax(ndvi3, '.expression()');

var nPixels = ndvi1.select(0)//.mask()
  .reduceRegion({
    reducer: ee.Reducer.count(),
    scale: 30,
    maxPixels: 1e13});

var nDifPixels = ndvi1.subtract(ndvi2)
  .neq(0)
  .selfMask()
  .reduceRegion({
    reducer: ee.Reducer.count(),
    scale: 30,
    maxPixels: 1e13});

print('n valid image pixels:', nPixels.get('NDVI'));
print('n valid pixels neq 0 (dif):', nDifPixels.get('NDVI'));
print('percent valid pixels neq 0 (dif)',
  ee.Number(nDifPixels.get('NDVI'))
  .divide(ee.Number(nPixels.get('NDVI')))
  .multiply(100));

Different min and max NDVI values with different approaches using Google Earth Engine

One Answer

Add your own answers!

Ask a Question