Classifying Patterns of Land Cover Using Mutual Information and Clustering

Comparison between two maps of land use/land cover (LULC) is a fundamental task in remote sensing and geospatial data analysis with application to change detection, validation of models, and accuracy assessment [1, 2]. We present a methodology of calculating a similarity measure between every pair of LULC maps in a collection in order to classify them into characteristic LULC patterns. A pattern is a specific composition of LULC categories and their particular spatial arrangement; it represents a higher level abstraction of landscape than a single LULC category. For example, a LULC map of a city provides visual assessment of spatial relations between its constituent categories, but it also defines a pattern characteristic fingerprint of this particular city in terms of LULC. A collection of different cities can be grouped into classes on the basis of similarities between their patterns. Traditional approaches to quantification of LULC patterns focus on either composition (histogram of LULC categories) or spatial configuration (various pattern indicators). However, for the purpose of classification, it is sufficient to quantify only a similarity between two patterns without quantifying patterns themselves. We use a similarity measure [2] that simultaneously takes into account composition and configuration information. First, this similarity measure is applied to every pair of maps in a collection; the output of this step is a similarity matrix that encapsulates pairwise similarities between the maps. Second, we use a combination of clustering and visualization methods to translate the information contained in the similarity matrix into classification of maps. The output of this step is a similarity map that depicts the overall structure of similarities between various maps in the collection, and a classification scheme that groups maps into characteristic LULC-patterns.