Local spatial autocorrelation characteristics of Landsat TM imagery of a managed forest area

In remote sensing, continuous landscapes are sampled into a grid of equally sized and regularly spaced pixels. One consequence of this surface regularization is that pixel values exhibit positive spatial autocorrelation. Accordingly, the extent and nature of spatial autocorrelation can be considered a characteristic of remotely-sensed data which may be exploited as an information source. However, existing global measures of spatial autocorrelation provide little insight into this characteristic since they summarize all spatial inter-relationships in a single measure. In contrast, local indicators of spatial association (LISA) measures assess for each pixel in the image both the degree of spatial dependence with neighbouring pixels and the magnitude of variate values in the neighbourhood of the pixel. In this study, one such LISA statistic, the Getis statistic (G(i)*), is applied to Landsat TM imagery of a managed forest region, Relationships are found between local spatial autocorrelation as measured by G(i)* and different Landsat TM image bands and differing cover types. Further, the spatial dependence information is presented in the context of forest inventory, polygons indicating the presence of spectral heterogeneity or homogeneity within forest polygon areas. This exploratory research confirms that spatial dependence information, as computed by G(i)*, constitutes a valuable new source of spatial information for the assessment of digital imagery, of forests.