Segmentation and classification of multitemporal data: methodology and results of a modified Gaussian Markov random field model classification system

Constructive and destructive interference have long been a significant problem for the classification of radar imagery. Several authors have noted the ability of segmentation to work around the problem by incorporation of spatially adaptive filters and fuzzy logic. We present the methodology and results of such a study using multitemporal datasets in the place of multispectral data. For the study 27 standard and ScanSAR images were collected over the region from 1996 to 2001. The majority of the images were S4 ascending, with ScanSAR narrow collected simultaneously. The ScanSAR images allowed us to assess the methodology over a wider swath, yielding results on a regional rather than local scale. Northern Australia exhibits an extremely seasonal climate. The monsoonal nature of this climate means a huge variation in the amount of available water to plant life both on and surrounding the flood plains of the major river systems. Within this area is Kakadu National Park, a heritage listed and Ramsar identified wetland of international significance. Kakadu has a long history of scientific study, and is an ideal site on which to conduct studies such as this. The seasonal nature of this area presents a unique opportunity, that of a worst case scenario for classification schemes, whereby each location on the flood plain undergoes significant change over the seasonal cycle, similar in many respects to the change between bands of a multi-spectral dataset. A modified Gaussian Markov random field model segmentation routine was used to cluster areas exhibiting similar radar response (both numerical and textural) at each successive date in the time series. For each date cluster statistics were then generated. This allowed the construction of temporal curves due to the fact that as the target material dries the material´s dielectric constant decreases. As dielectric constant is significantly dependent on water content, these temporal curves act as a proxy measure for water availability, and hence aid in the discrimination of wetland from non-wetland areas and any change in size and location of these areas throughout the season. The output of the segmentation routine is a series of three bands. The first and second bands are the statistics (mean and standard deviation) of t- he original image and the third a vector file of the edge locations. By outputting statistical information instead of arbitrary segment number standard classification routines may be used successfully on radar imagery. Classification of the results of the segmentation may be either supervised or unsupervised, however analysis indicated that the Isodata algorithm adequately addressed classification needs. As a combined system the segmentation and classification system used in this paper has shown some excellent results when applied to multi-date radar imagery of Kakadu National Park in Northern Australia. Results for a segmentation of multitemporal radar dataset are presented alongside the results of the subsequent classification. By creating a system that requires minimal user input a wide variety of applications may be addressed. Indeed, the registration of the imagery becomes the most user intensive portion of the process. Automated routines such as this allow for analysis of large areas on a regular basis, an exceptional result for monitoring change and establishing baselines for later comparison.