Skip to main content
ARS Home » Plains Area » Las Cruces, New Mexico » Range Management Research » Research » Publications at this Location » Publication #225742
Title: Correlation of object-based texture measures at multiple scales in sub-decimeter resolution aerial photography

Author
item LALIBERTE, ANDREA - NEW MEXICO STATE UNIV
item Rango, Albert

Submitted to: Meeting Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 3/25/2008
Publication Date: 8/7/2008
Citation: Laliberte, A., Rango, A. 2008. Correlation of object-based texture measures at multiple scales in sub-decimeter resolution aerial photography. Proceedings of the International Archives of the Photogrammetery, Remote Sensing, and Spatial Information Sciences: GEOBIA 2008. Pixels, Objects, Intelligence. August 6-7, 2008 Calgary, AB. ISPRS, Vol. No. XXXVII-4/C1.

Interpretive Summary:

Technical Abstract: Texture measures are commonly used to increase the number of input bands in order to improve classification accuracy, especially for panchromatic or true color imagery. While the use of texture measures in pixel-based analysis has been well documented, this is not the case for texture measures calculated in an object-based environment. In addition, the latter application is computer intensive, and may occur at multiple segmentation scales. Therefore fewer variables are preferred and a knowledge of correlation and how it changes across segmentation scales is required. The objectives of this study were to assess correlations between texture measures as a function of segmentation scale while mapping rangeland vegetation structure groups. A 5-cm resolution true-color aerial photo mosaic was segmented at 15 consecutively coarser scales using the object-based image analysis Definiens Professional. We investigated 10 gray-level co-occurrence matrix (GLCM) texture measures, determined the optimal texture measures for each scale with a decision tree, and calculated correlation coefficients for all texture pairs. Entropy, mean and correlation were least correlated with other texture measures at all scales. The highest correlation that remained stable across all segmentation scales was found for contrast and dissimilarity. We observed both increasing and decreasing correlation coefficients for texture pairs as segmentation scale increased, and there was larger variability from one scale to the next at finer segmentations and more consistency in correlation at medium to coarse scales. This was attributed to the fact that at finer segmentation scales, smaller objects were more numerous, and the ratio of edge to interior pixels for an image object was higher than at coarser scales. At medium to coarse scales, vegetation patches were delineated more precisely, and as a result, classification accuracy and class separability were highest. Due to the sensitivity of texture to scale, an optimal segmentation scale is important in object-based analysis. This approach allowed for determining the most suitable and uncorrelated texture measures at the optimal image analysis scale, was less computer intensive than a series of test classifications, and shows promise for incorporation into rangeland monitoring protocols with very high resolution imagery.