|Title||Human Geography—A Link between Image Analysis and the Social Sciences?|
|Abstract||For human geographers, map analysis has traditionally been an important tool of research. Also, in the 1960s and 1970s, geographers were among the first to embrace computer technology for analyzing spatial distributions and satellite imagery. Since then, computer-based image analysis has developed into a major research field whereas human geography has become increasingly dominated by non-quantitative research methods.
In recent years, however, there has been a renewed interest in quantitative approaches. One reason is the increased accessibility of longitudinal, individual-level, population-based databases that allows detailed research on a broad range of topics. Another reason is an increased accessibility of fast computers able to process large satellite image files (a typical Landsat image can be around 700 MB).
It could, therefore, be of interest to discuss to what extent research problem that face human geographers might be approached with image analysis methods. We will present three research areas that might contain image analysis related problems: the PLACE database, SRPC-based analysis of farming systems, and window independent context based segmentation (WICS).
The PLACE database contains one record per year for each individual residing in Sweden between 1990 and 2006. Each record holds information about income, education, employment, and family. In addition PLACE contains geo-coordinates (100 meter square) based on the home address and geo-coordinates based on the address of their employer. Because of this link to 100 meter square “pixels” we would argue that PLACE could be seen as an image file with large amounts of attribute information connected to each pixel. The question is to what extent image-analysis methods could be applied to this data, for example, in order to construct standardized statistical areas that are large enough to reduce sampling errors and still fairly homogenous with respect to attribute data.
SRPC is a method developed at the Department of Human Geography for the analyses composite land-cover complexes in satellite imagery. The idea is to use on distance-relations between different spectral classes in order to identify the presence of certain farming system in a satellite image scene. One question is if similar approaches have been used in image analysis.
Window independent context based segmentation (WICS, patent pending) can be seen as a generalization of SRPC and is carried out in three steps. The first step is an unsupervised classification of a satellite image based on spectral values. The second step is to apply a distance transform to each of the classes obtained in the first step. If the unsupervised classification has resulted in 20 classes then the second step will generate 20 distance values for each pixel, on for each spectral class. The third step is to run a cluster analysis on the values obtained in the second step that assigns pixels to the same segment if they have similar distance relationships to the classes used in the analysis to the same segment. Since pixels located close to each other will have similar distance relationships they will be assigned to the same segment. However, if there is a drastic shift in the composition of classes the distance relationships will change and pixels will be assigned to different classes. We have not been able to find any references to this type of approach in the remote sensing literature and it would, also in this case, be of interest to know if a similar approach has been described in the image analysis literature.