Classification of buildings using USGS Landsat 8 images on Google Earth Engine

Question

I am trying to classify built-up and non-built-up areas in the district of Ambala in India within the time period 2014-17 using a simple classifier model I had created.

The model uses 20,000 points from this fusion table. The classifier takes these bands: ['B1','B2', 'B3', 'B4', 'B5', 'B6', 'B7', 'B8', 'B10', 'B11','NDBI','NDVI'], trains a CART model on 3000 points (1200 built up+1800 non built up). I had also tried to account for cloud cover by sorting and filtering according to my value of CLOUD_COVER. In the final result I get a mask s.t white is for built up areas and black for non-built up. The problem is there is a clear discrepancy in my results for 4 consecutive years, 2014-2017. Here are my classification results for the 4 years.

Now , the white portion unnaturally increases from 2014-15, and then subsequently decreases from 2015-2016 and 2017. Such kind of pattern in building areas is not possible, and the weird part is that I tried the same code for different districts and the pattern of discrepancy is the same among all. I wish to know what could be the cause of this issue? Is it some kind of problem with my code or the LANDSAT 8 images for 2015? And how can I solve this?

Here is my code (for reference):

//Loading India image, the extracting data for Haryana (a state in India) and then subsequently Ambala (a district in Haryana) 
var bands = ['B1','B2', 'B3', 'B4', 'B5', 'B6', 'B7', 'B8', 'B10', 'B11'];
var india = ee.FeatureCollection('ft:1UDdgOCf8DoRJ9bVm-UVbR6CqxtkJToLQjTFd0r0Z','geometry').filter(ee.Filter.eq('Name','India')).geometry();
var india_image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA').filterBounds(india).filterDate('2014-02-01','2014-09-01').sort('CLOUD_COVER').limit(500).mosaic();
var fc = ee.FeatureCollection('ft:1PA2zwArj8EsplrX9eMxJ2H_TICyyx855KPnbJhC1','geometry')
var district = fc.filter(ee.Filter.eq('name','Ambala'));
var haryana_image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA').filterBounds(district).filterDate('2014-02-01','2014-04-01').sort('CLOUD_COVER').limit(4).mosaic();
var input = haryana_image;
input = addBands(input.select(bands));
india_image = addBands(india_image);

//Loading the points from the fusion table and training the classifier
var ft = ee.FeatureCollection('ft:1fWY4IyYiV-BA5HsAKi2V9LdoQgsbFtKK2BoQiHb0');
var ft_builtup = ft.filter(ee.Filter.eq('class',1)).limit(1200);
var ft_nonbuiltup = ft.filter(ee.Filter.eq('class',2)).limit(1800);
ft = ft_builtup.merge(ft_nonbuiltup);
var new_bands = ['B1','B2', 'B3', 'B4', 'B5', 'B6', 'B7', 'B8', 'B10', 'B11','NDBI','NDVI'];
function addBands(image){
  var ndvi = image.normalizedDifference(['B4', 'B3']).rename('NDVI');
  var ndbi = image.normalizedDifference(['B5', 'B4']).rename('NDBI');
  var ndwi = image.normalizedDifference(['B6', 'B6']).rename('NDWI');
  return image.addBands(ndvi).addBands(ndbi).addBands(ndwi);
}

// Load a Landsat 8 image to be used for prediction.
var training = india_image.sampleRegions(ft,['class'],30);
var trained = ee.Classifier.cart().train(training, 'class', new_bands);

input = input.clip(district);
input = input.classify(trained);
input = input.expression('LC==1?1:0',{'LC':input.select('classification')});
Map.addLayer(input.clip(district));

Answer 1

The issue is clouds. You're trying to select only less cloudy scenes, but you're not doing anything to remove the cloudy pixels within those scenes. You are going to need a way to remove cloudy pixels or to only select the less-cloudy pixels. There are several ways you might do this:

You can map a function that masks cloudy pixels over your image collection. You can write your own masking function, but personally I like to use Rodrigo Principe's cloud masking module. Once you have masked the pixels, you can create a composite by selecting the pixel from the least cloudy scene, taking a median of all available pixels, or by calculating the cloud score of each pixel and using that as the quality band for a quality mosaic.

You can also use the built-in ee.Algorithms.Landsat.simpleComposite method, which takes a collection of raw Landsat images, calculates cloud score, takes the least cloudy pixel, and returns a TOA adjusted image. The following code display a 2015 image of your study area using these different methods.

// Load Rodrigo Principe's cloud masking module
var cloud_masks = require('users/fitoprincipe/geetools:cloud_masks');
var maskClouds = cloud_masks.landsatTOA();

//Loading India image, the extracting data for Haryana (a state in India) and then subsequently Ambala (a district in Haryana) 
var bands = ['B1','B2', 'B3', 'B4', 'B5', 'B6', 'B7', 'B8', 'B10', 'B11'];
var india = ee.FeatureCollection('ft:1UDdgOCf8DoRJ9bVm-UVbR6CqxtkJToLQjTFd0r0Z','geometry').filter(ee.Filter.eq('Name','India')).geometry();
var india_image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA').filterBounds(india).filterDate('2014-02-01','2014-09-01').sort('CLOUD_COVER').limit(500).mosaic();
var fc = ee.FeatureCollection('ft:1PA2zwArj8EsplrX9eMxJ2H_TICyyx855KPnbJhC1','geometry')
var district = fc.filter(ee.Filter.eq('name','Ambala'));
var haryana_image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA')
    .filterBounds(district)
    .filterDate('2014-02-01','2014-04-01')
    .sort('CLOUD_COVER')
    .limit(20)
    .map(maskClouds)
    .mosaic();
var input = haryana_image;
input = addBands(input.select(bands));
india_image = addBands(india_image);

//Loading the points from the fusion table and training the classifier
var ft = ee.FeatureCollection('ft:1fWY4IyYiV-BA5HsAKi2V9LdoQgsbFtKK2BoQiHb0');
var ft_builtup = ft.filter(ee.Filter.eq('class',1)).limit(1200);
var ft_nonbuiltup = ft.filter(ee.Filter.eq('class',2)).limit(1800);
ft = ft_builtup.merge(ft_nonbuiltup);
var new_bands = ['B1','B2', 'B3', 'B4', 'B5', 'B6', 'B7', 'B8', 'B10', 'B11','NDBI','NDVI'];
function addBands(image){
  var ndvi = image.normalizedDifference(['B4', 'B3']).rename('NDVI');
  var ndbi = image.normalizedDifference(['B5', 'B4']).rename('NDBI');
  var ndwi = image.normalizedDifference(['B6', 'B6']).rename('NDWI');
  return image.addBands(ndvi).addBands(ndbi).addBands(ndwi);
}


// Load a Landsat 8 image to be used for prediction.
var training = india_image.sampleRegions(ft,['class'],30);
var trained = ee.Classifier.cart().train(training, 'class', new_bands);


input = input.clip(district);
input = input.classify(trained);
input = input.expression('LC==1?1:0',{'LC':input.select('classification')});

// Display input images with and without cloud masking.

// Display the input images for 2014 - 2017
var unmaskedImages = ee.List.sequence(2014, 2017).map(function(year) {
    var startDate = ee.Date.fromYMD(year, 2, 1);
    var endDate = ee.Date.fromYMD(year, 4, 1);

    var image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA')
        .filterBounds(district)
        .filterDate(startDate, endDate)
        .sort('CLOUD_COVER')
        .limit(4)
        .mosaic()
        .clip(district);
    return image;
});

// Mask images before mosaicking
var maskedImages = ee.List.sequence(2014, 2017).map(function(year) {
    var startDate = ee.Date.fromYMD(year, 2, 1);
    var endDate = ee.Date.fromYMD(year, 4, 1);

    var image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA')
        .filterBounds(district)
        .filterDate(startDate, endDate)
        .sort('CLOUD_COVER')
        .map(maskClouds)
        // .limit(4)
        .mosaic()
        .clip(district);
    return image;
});

// Calculate simple cloud score and use that for quality mosaic
var qualityMosaickedImages = ee.List.sequence(2014, 2017).map(function(year) {
    var startDate = ee.Date.fromYMD(year, 2, 1);
    var endDate = ee.Date.fromYMD(year, 4, 1);

    var image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA')
        .filterBounds(district)
        .filterDate(startDate, endDate)
        .map(function(image) {
            return image.addBands(ee.Image(100).subtract(ee.Algorithms.Landsat.simpleCloudScore(image)))
        })
        .qualityMosaic('cloud')
        .clip(district);
    return image;
});

// Use built in simple composite
var simpleCompositeImages = ee.List.sequence(2014, 2017).map(function(year) {
    var startDate = ee.Date.fromYMD(year, 2, 1);
    var endDate = ee.Date.fromYMD(year, 4, 1);

    var images = ee.ImageCollection('LANDSAT/LC08/C01/T1')
        .filterBounds(district)
        .filterDate(startDate, endDate)

    var image = ee.Algorithms.Landsat.simpleComposite(images)
        .clip(district);
    return image;
});

// Take the median value
var medianImages = ee.List.sequence(2014, 2017).map(function(year) {
    var startDate = ee.Date.fromYMD(year, 2, 1);
    var endDate = ee.Date.fromYMD(year, 4, 1);

    var image = ee.ImageCollection('LANDSAT/LC08/C01/T1_TOA')
        .filterBounds(district)
        .filterDate(startDate, endDate)
        .sort('CLOUD_COVER')
        // .limit(4)
        .map(maskClouds)
        .median()
        .clip(district);
    return image;
});

var vis = {
    bands: 'B4,B3,B2',
    gain: 500.0,
    // max: 3000.0,
    // min: 0,
};

for (var i = 0; i < 4; i++) {
    var year = 2014 + i;
    var shouldDisplay = i === 1;
    Map.addLayer(ee.Image(medianImages.get(i)), vis, 'Median: ' + year, shouldDisplay);
}

for (var i = 0; i < 4; i++) {
    var year = 2014 + i;
    var shouldDisplay = i === 1;
    var simpleVis = {
        bands: 'B4,B3,B2',
        // gain: 500.0,
        max: 100,
        min: 0,
    };
    Map.addLayer(ee.Image(simpleCompositeImages.get(i)), simpleVis, 'Simple Composite: ' + year, shouldDisplay);
}

for (var i = 0; i < 4; i++) {
    var year = 2014 + i;
    var shouldDisplay = i === 1;
    Map.addLayer(ee.Image(qualityMosaickedImages.get(i)), vis, 'Quality Mosaicked: ' + year, shouldDisplay);
}

for (var i = 0; i < 4; i++) {
    var year = 2014 + i;
    var shouldDisplay = i === 1;
    Map.addLayer(ee.Image(maskedImages.get(i)), vis, 'Masked: ' + year, shouldDisplay);
}

for (var i = 0; i < 4; i++) {
    var year = 2014 + i;
    var shouldDisplay = i === 1;
    Map.addLayer(ee.Image(unmaskedImages.get(i)), vis, 'Non-masked: ' + year, shouldDisplay);
}

https://code.earthengine.google.com/4b5e29abed47496cec1327c5962036a8