Improving land cover classification using input variables derived from a geographically weighted principal components analysis

A - Papers appearing in refereed journals

Comber, A. J., Harris, P. and Tsutsumida, N. 2016. Improving land cover classification using input variables derived from a geographically weighted principal components analysis. ISPRS Journal of Photogrammetry and Remote Sensing. 119 (September), pp. 347-360. https://doi.org/10.1016/j.isprsjprs.2016.06.014

AuthorsComber, A. J., Harris, P. and Tsutsumida, N.
Abstract

This study demonstrates the use of a geographically weighted principal components analysis (GWPCA) of remote sensing imagery to improve land cover classification accuracy. A principal components analysis (PCA) is commonly applied in remote sensing but generates global, spatially-invariant results. GWPCA is a local adaptation of PCA that locally transforms the image data, and in doing so, can describe spatial change in the structure of the multi-band imagery, thus directly reflecting that many landscape processes are spatially heterogenic. In this research the GWPCA localised loadings of MODIS data are used as textural inputs, along with GWPCA localised ranked scores and the image bands themselves to three supervised classification algorithms. Using a reference data set for land cover to the west of Jakarta, Indonesia the classification procedure was assessed via training and validation data splits of 80/20, repeated 100 times. For each classification algorithm, the inclusion of the GWPCA loadings data was found to significantly improve classification accuracy. Further, but more moderate improvements in accuracy were found by additionally including GWPCA ranked scores as textural inputs, data that provide information on spatial anomalies in the imagery. The critical importance of considering both spatial structure and spatial anomalies of the imagery in the classification is discussed, together with the transferability of the new method to other studies. Research topics for method refinement are also suggested.

KeywordsGWmodel; GWPCA; Spatial heterogeneity; Accuracy
Year of Publication2016
JournalISPRS Journal of Photogrammetry and Remote Sensing
Journal citation119 (September), pp. 347-360
Digital Object Identifier (DOI)https://doi.org/10.1016/j.isprsjprs.2016.06.014
Open accessPublished as non-open access
FunderBiotechnology and Biological Sciences Research Council
Funder project or codeThe North Wyke Farm Platform [2012-2017]
Output statusPublished
Publication dates
Online19 Jul 2016
Publication process dates
Accepted17 Jun 2016
PublisherElsevier
Elsevier Science Bv
Copyright licensePublisher copyright
ISSN0924-2716

Permalink - https://repository.rothamsted.ac.uk/item/8v2x9/improving-land-cover-classification-using-input-variables-derived-from-a-geographically-weighted-principal-components-analysis

Restricted files

Publisher's version

Under embargo indefinitely

89 total views
1 total downloads
0 views this month
0 downloads this month