Streamlining Land Cover Classification

by | Jul 8, 2011

Combining multispectral aerial imagery with automated processing shows promise while revealing challenges.

By Anna Gibson, University College of London Department of Civil, Environmental and Geomatic Engineering (www.cege.ucl.ac.uk), London.

Table 1. Imagery acquired by Ordnance Survey's airborne Intergraph Z/I DMC sensor.

Ordnance Survey, the U.K.'s national mapping agency, currently verifies land cover types manually. To make the process more efficient, a semi-automated approach was evaluated using readily available aerial imagery (Table 1). The goal was to accurately discriminate between different intertidal land cover types.

Project Specifics
A team from University College London's Department of Civil, Environmental and Geomatic Engineering compared a range of unsupervised, supervised and rule-based classification algorithms for a 1- by 1-kilometer area of Chichester Harbour on England's South Coast (Figure 1).

Ground truthing verified actual land cover types by acquiring sample points that were used as the reference dataset for accuracy assessment. The project focused on comparing the performance of classification algorithms available within ENVI EX image processing software, developed by ITT Visual Information Solutions (www.ittvis.com).

Figure 1. Based on spectral properties, the study compared the performance of classification algorithms for discriminating between different intertidal land cover types through the use of multispectral aerial imagery. The study area encompassed West Wittering Beach and East Head Spit on England's South Coast.

 

Several U.K. land cover mapping projects and classifications already have been completed, leading to a variety of land cover classification schemes. For this study, definitions were adopted from a 2008 report by the U.K. Biodiversity Action Plan Steering Group (www.ukbap.org), which defined 65 priority habitats as a result of the 1992 European Union Habitats Directive.

Table 2. Comparison of overall accuracies for best-performing classifications.

 

The following classifications were performed with ENVI EX, testing different parameters within each algorithm:

Unsupervised (ISODATA)

” 4 Classes

” 5 Classes

” 6 Classes

Figure 2. A comparison shows the best-performing classifications.

” 7 Classes

Supervised

” Maximum Likelihood

” Minimum Distance

” Mahalanobis Distance

Spectral Angle Mapper

Supervised Rule Based

The researcher's used the following methodology:

1. Preliminary classifications were based on assumed land cover from reference sources and prior studies.

2. Ground truthing was performed through a stratified random sampling strategy, with location determined through real-time kinematic Global Positioning System technology.

3. Following ground truthing, classes were reduced to five in total (Sublittoral Sands and Gravels, Mudflats, Coastal Sand Dunes, Coastal Saltmarsh and Water), and spectral profiles were created (Figure 2).

4. Coincidence of actual land cover against mapped land cover was assessed with ArcGIS 9.3 geographic information system software from Esri (www.esri.com), and accuracy was quantitatively assessed through contingency tables: Producer's Accuracy, User's Accuracy, Omission and Commission Errors.

Results
Overall the best classification accuracy (72.4 percent), assessed through Producer's Accuracy and User's Accuracy, was achieved by the Mahalanobis Distance classifier when five classes were specified (Table 2). This implies a restriction to which four-band aerial imagery can be used to discriminate between classes when basing pixel allocation to classes on spectral properties.

Coastal Sand Dunes generally was the most accurately defined class, with the Mahalanobis Distance classifier producing Producer's Accuracy and User's Accuracy of 95.8 percent. Mudflats was the most difficult class to discriminate. Figure 2 shows final visual classifications. Supervised Rule-Based classification didn't perform as well as expected, perhaps because of the similar spectral profiles of the land cover types involved.

Future work will involve revisiting the site at the same time of year as image acquisition for best comparisons, assessing the use of aerial imagery for monitoring intertidal land cover change over time, further investigating rule creation and refinement in ENVI EX, and comparing airborne multispectral imagery performance against hyper-spectral data.              

 

NEWEST V1 MEDIA PUBLICATION

October Issue 2023