Spatial dependence of predictions from image segmentation : a methods to determine appropriate scales for producing land-management information

A challenge in ecological studies is defining scales of observation that correspond to relevant ecological scales for organisms or processes. Image segmentation has been proposed as an alternative to pixel-based methods for scaling remotely-sensed data into ecologically-meaningful units. However, to date, selection of image object sets has been largely subjective. Changing scale of image segmentation affects the variance and spatial dependence (amount and range of spatial autocorrelation) of measured variables, and this information can be used to determine appropriate levels of image segmentation. Our objective was to examine how scaling via image segmentation changes spatial dependence of regression-based predictions of landscape features and to determine if these changes could identify appropriate segmentation levels for a given objective. We segmented an Ikonos image for southern Idaho (USA) into successively coarser scales and evaluated goodness-of-fit and spatial dependence of regression predictions of invasive western juniper (Juniperus occidentalis) density. Correlations between juniper density estimates and imagery increased with scale initially, but then decreased as scale became coarser. Scales with highest correlations generally exhibited the most spatial dependence in the regression predictions and residuals. Aggregating original juniper density estimates by image objects changed their spatial dependence, and the point at which spatial dependence began to diverge from the original observations coincided with the highest correlations. Looking at scale effects on spatial dependence of observations may be a simple method for selecting appropriate segmentation levels. The robustness of ecological analyses will increase as methods are devised that remove the subjectivity of selecting scales. * Corresponding author. Email: jkarl@nmsu.edu, phone: 575-646-7015. 1, INTRODUCTION A significant challenge in ecological studies has been defining scales of observation that correspond to relevant ecological scales for organisms or processes. Scale is a characteristic of a set of observations that controls what patterns can be detected from the observations (Wiens, 1989; Burnett and Blaschke, 2003). Operationally, scale is defined by the smallest observable unit (i.e., grain) and the maximum areal coverage (i.e., extent) of a set of observations (Turner et al., 1989). Because objects smaller than the grain size generally cannot be resolved, and patterns larger than the extent cannot be completely defined, selection of an appropriate scale is important to detecting and describing patterns that result from ecological processes. Traditionally, scaling of datasets has been accomplished through aggregating observations into successively coarser units of the same size (Wu and Li, 2006). Scaling effects of such methods that use a consistently-shaped support (i.e., analysis unit) include changes in mean and variance (Wu, 2004; Chen and Henebry, 2009) and the existence of scaling domains (Wiens, 1989). However, scaling via consistently-shaped supports can obscure boundaries and can magnify the effects of the modifiable areal unit problem (Dark and Bram, 2007). Image segmentation has been shown to be an effective method for scaling information such that relevant patterns are preserved and extraneous information (i.e., noise) is removed (Wu, 1999; Burnett and Blaschke, 2003; Hay and Marceau, 2004). This is because image segmentation is a "data-driven" approach to scaling where adjacent observations are merged based on similarity rules. The result of segmentation is a complete tessellation of an area into relatively homogeneous objects of differing shapes and sizes. By varying the degree of similarity needed to merge observations, finer or coarser scale segmentations can be achieved. Karl and Maurer (2010a) reported that higher and more consistent correlations between field and image data were achieved when scaling by image segmentation than by aggregating square pixels. However, to date, the selection of image object sets to represent landscape patterns has been largely subjective (Wang et al., 2004; Addink et al., 2007) and usually involves trial-and-error for selecting appropriate scales for analysis (Burnett and Blaschke, 2003; Feitosa et al., 2006; Navulur, 2007). However, the accuracy of image-derived products changes with segmentation level (Feitosa et al., 2006; Addink et al., 2007; Karl and Maurer, 2010a), and an optimal scale can be defined as the scale producing the most accurate results for a given objective. Changes in scale can also affect the magnitude and range of spatial autocorrelation (i.e., spatial dependence) between observations (Fortin and Dale, 2005), but this phenomenon has The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVIII-4/C7 not been studied extensively in object-based scaling. Karl and Maurer (2010b) showed that optimal object-based scales for predicting attributes of a semi-desert landscape minimized spatial autocorrelation of regression residuals. However, they also demonstrated that geostatistical techniques that could account for this residual spatial autocorrelation could yield predictions from scales finer than optimal that were about as accurate as predictions from the optimal scale. This suggests that spatial dependence of image objects relative to that of field observations may be an important trait in selecting objectsegmentation scales for analyses. Our objective was to examine how scaling via image segmentation changes the spatial dependence of regressionbased predictions of landscape features. Specifically we were interested in whether or not changes in observed spatial dependence of original measurements or regression predictions and residuals could be used to identify optimal (or nearoptimal) scales of segmentation.