Geographic information system strategies for improving Landsat land use classification accuracy

Abstract

This study focuses on POWER Geographic Information System strategies for improving the land use classification and mapping accuracies of Landsat multi spectral scanner (MSS) data. The specific strategies are additional-band, class reduction, exclusion, and conditional.

Three classified images served as the basis upon which all accuracy improvements were determined. Generated from 4 bands of Landsat data, each of these 3 images possessed 13-classes and were composed of clusters greater than or equal to 5, 9, and 13 pixels, respectively. Image classification was accomplished through use of the General Image Processing System (GIPSY) at Virginia Tech.

The additional-band method consisted of augmenting each 4-band Landsat image with 5 ancillary bands producing a 9-band image. Only moderate improvements in select class accuracies were realized.

The reduction method involved decreasing the number of land use classes from 13 (Level 2) to 4 (Level 1) in both 4- and 9-band imagery. This resulted in significant increases in overall image and select class accuracies.

In the exclusion method, 7 data sets were removed sequentially from each 13-class image with 3 data sets removed from 4-class images. These data sets represented land uses with a low probability of consistent classification. Each exclusion produced moderate increases in overall image and select class accuracies.

A conditional classification method, utilizing a hierarchical decision-tree structure, was developed to determine its potential for increasing image accuracies. Slope angle, slope aspect, elevation, distance-from-water, and Landsat classified land use were included as environmental parameters. Decreased accuracies resulting from data registration and overlay problems warrant further research into and testing of this method.

Significant efficiencies in assessing subscene accuracies were realized using a distance-from-roads sampling strategy. Ground data, inclusive of 9 pixels from a primary or secondary highway, produced comparable accuracies to those derived from more extensive ground truth.