GBDX

Automated Land Cover Classification (LULC)

This task performs unsupervised land use land cover classification on the GBDX platform. There are six classes: vegetation, water, bare soil, clouds, shadows, and unclassified.

GBDX Registered Name: lulc
Provider: GBDX
Inputs: This task requires an atmospherically compensated WorldView-2 or WorldView-3 multispectral image
Outputs: RGB
Compatible bands & sensors: WorldView 2 or WorldView 3 multi-spectral imagery (8-band optical and VNIR data sets) that has been atmospherically compensated by the Advanced Image Preprocessor task.

Table of Contents

Section Description
Overview Detailed description
Imagery Examples Before and after examples
Quickstart Get started with a Python-based quickstart tutorial
Task Runtime Benchmark runtimes for the algorithm
Input Options Required and optional task inputs
Outputs Task outputs and example contents
Known Issues Issues users should be aware of

Overview

The underlying algorithm classifies each pixel in the input image using spectral fitting to known material spectral signatures and certain class-specific shape and size filters. The input image must be an atmospherically compensated WorldView-2 or WorldView-3 multispectral image.

By default, the task produces an RGB image where each class is color coded.

Class Color Description
vegetation green [0,255,0] All types of vegetation (healthy chlorophyll content)
water blue [0,0,128] All types of water, including murky/impure water
bare soil brown [128,64,0] All types of soils, excluding rocks and stone
clouds light blue [128,255,255] All types of clouds excluding smoke
shadows purple [164,74,164] Shadows
unclassified gray [128,128,128] Unclassified (equivalent to man-made materials, rock, stone)

The task can also produce a mask for selected classes, where pixels corresponding to the selected classes are white and all remaining pixels are black.

No data zones in the input image are colored black in the output image.

Imagery Example

This example shows the rgb output from the Automated Land Cover Classification

This example shows the rgb output from the Automated Land Cover Classification

This is the corresponding unclassified mask

This is the corresponding unclassified mask

See Know Issues for help in interpreting the results.

Quickstart Tutorial

In a Python terminal:

import gbdxtools

gbdx = gbdxtools.Interface()

lulc = gbdx.Task('lulc')
lulc.inputs.image = 's3://gbd-customer-data/32cbab7a-4307-40c8-bb31-e2de32f940c2/platform-stories/coastal-change/images/pre'

# Run workflow and save results
wf = gbdx.Workflow([lulc])
wf.savedata(lulc.outputs.image, 'platform-stories/trial-runs/lulc')
wf.execute()

Task Runtime

There is no runtime data available for this algorithm.

Input Options

Name Type Description Required
image Directory Contains input image. The input image must be a WV02/WV03 multispectral image which is atmospherically compensated. If more than one images are contained in this directory, one is picked arbitrarily. True
vegetation String If True, the output is a vegetation mask. Default is False. False
water String If True, the output is a water mask. Default is False. False
soil String If True, the output is a bare soil mask. Default is False. False
clouds String If True, the output is a cloud mask. Default is False. False
shadows String If True, the output is a shadow mask. Default is False. False
unclassified String If True, the output is an unclassified material mask. Default is False. False
tiles String Number of tiles to tile input image into if it is too big. In that case, the recommended number is 2. Only use this if the default option fails. Default is 1. False
verbose String If True, save algorithm config files in output directory. To be used for debugging purposes. Default is False. False

Note that if more than one class is set to True, the corresponding mask includes all the classes set to True.

Outputs

Name Type Description
image Directory Contains output image.

Known Issues

  • Shadows may be misinterpreted as water.
  • Thin water bodies may be discarded.
  • Small vegetation patches may be lost.
  • Cloud holes are to be expected.
  • Regions than appear as bare soil in the original image and interpreted as vegetation in the LULC are due to spatial aggregation of small grass patches which are not necessary evident.
  • The shadows class contains a limited subset of the true set of all shadow regions.
  • The unclassified class can be used as a rough approximation of built-up.