Using Texture to Tackle the Problem of Scale in Land-cover Classification

This paper discusses the problem of scale in current approaches to Object Based Image Analysis (OBIA), and proposes how it may be overcome using theories of texture. It is obvious that aerial images contain land-cover that is textured and thus any features used to derive a land-cover classification must model texture information and as well as intensity. Previous research in the area of OBIA has attempted to derive land-cover classification using intensity features only, ignoring the presence of texture. This has led to a number of issues with the current theory of OBIA. Using only intensity it is impossible to perform segmentation of textured landcover. In an attempt to tackle this problem it has become practice in OBIA to run segmentation at a number of different scales in the hope that each textured region will appear correctly segmented at some scale. This process of performing segmentation at multiple scales is not in line with current theories of visual perception. Julesz (Julesz 1983) states that when we view an object our aperture is adjusted to view that object in its true form. Also in theories of visual object recognition each object or feature is represented only once in its true form. The result of integrating segmentation at multiple scales is the generation of a land-cover hierarchy in a bottom-up manner but this is not how our visual system generates such hierarchies. This process in the visual system is conversely very top-down, with the aggregation of objects not only being driven by their relative intensity or texture features but also our knowledge, desires and expectations. Quantitative evaluation is also made increasingly difficult due to the lack of ground truth for each scale; it is impossible to predict the appropriate appearance of ground truth at each scale. Given the fact that each land-cover is represented at a number of scales, the number of context relationships between objects which must be managed is exponentially large. This makes the task of deriving land-use from landcover increasingly difficult. If a robust set of intensity and texture features can be extracted and integrated correctly it would be possible to represent each land-cover in its true form within the one segmentation. Using a non-linear diffusion process and a geostatistical feature extraction algorithm we extract a set of intensity and texture feature respectively. Theses features are then integrated in such a manner to perform discriminate land-cover based on intensity where possible and texture where not. The motivation being that intensity features do not suffer from the uncertainty principle unlike texture thus giving more accurate boundary localization. * Corresponding author.