Supporting Quality-Based Image Retrieval Through User Preference Learning

It is common for modern geospatial libraries to contain multiple datasets that cover the same area but differ only in some specific quality attributes (e.g., resolution and precision). This is affecting the concept of content-based geospatial queries, as simple coverage-based query mechanisms (e.g., declaring a specific area of interest) as well as theme-based query mechanisms (e.g., requesting a black and white aerial photo or multispectral satellite imagery) are rendered inadequate to identify and access specific datasets in such collections. In this paper we introduce a novel approach to handle data quality attributes in geospatial queries. Our approach is characterized by the ability to model and learn user preferences, thus establishing user profiles that allow us to customize image queries for improving their functionality in a constantly diversifying geospatial user community. Introduction GIS data in digital format are collected and stored at constantly increasing rates. Furthermore, the GIS user community is also expanding, with users of various levels of expertise aiming to access and use this information. Considering geospatial imagery in particular, it is now common for geospatial libraries to contain multiple datasets that cover the same area but differ only in some specific quality attributes (e.g., resolution and precision). This evolution is affecting the concept of content-based geospatial queries, as simple coveragebased (e.g., declaring a specific area of interest) and themebased (e.g., requesting a black and white aerial photo or multispectral satellite imagery) query mechanisms are rendered inadequate to identify and access specific datasets in such collections. Thus, we find ourselves in need of query solutions to support complex content-based queries where the term content refers (beyond spatial and temporal coordinates) to quality attributes (e.g., resolution and precision) of geospatial datasets. In this paper we address this issue by introducing a novel approach to geospatial queries that takes into account quality attributes. Data quality in this case is used as an umbrella term. Concepts that play into our definition of data quality are, among others: accuracy, error, and precision (Buttenfield, 1993). In typical image database queries users provide an ideal set of quality attributes that best satisfies their needs, and the database is searched to identify the dataset that best resembles the user request in terms of these attributes. This involves a comparison of attribute values of each dataset to the userdefined ideal, and the ranking of the datasets. The degree to Supporting Quality-Based Image Retrieval Through User Preference Learning Giorgos Mountrakis, Anthony Stefanidis, Isolde Schlaisich, and Peggy Agouris which a dataset meets the specifications requested by a user to address a specific task at hand is termed fitness of use. We keep this well-established approach, but modify the mechanism that compares quality attributes among datasets, to accomodate the fact that user preferences are much more complex than traditional nearest neighbor approaches. Indeed, common solutions make use of standard distance metrics (for example, nearest neighbor) that are rather simplistic (e.g., linear) and ignore the fact that attribute variations may have different importance to different types of users. For example, scale variations may affect dataset fitness differently for a geologist as compared to a cartographer. Current solutions ignore these particularities of different user communities, using the same distance function to evaluate datasets regardless of a user’s preferences. To overcome this shortcoming we introduce a novel approach that allows users to express their own preferences, models these preferences in complex distance metric functions, and uses these functions to evaluate dataset similarities for each request. Our approach could be used to model various attributes of a geospatial dataset; it is particularly suitable for quality attributes like the above mentioned ones, as they are the ones that convey inherently application-dependent preferences that cannot be satisfied by existing dataset similarity metrics. Essentially, our approach leads to the establishment of user profiles in geospatial queries, expressing the manner in which attribute variations affect a user’s preferences. These profiles may be user dependent, application dependent, or both. A user profile is a set of parameters, describing the relationship between attribute variations and user preferences in a mathematical manner. We proceed by first defining, then training, and finally using these profiles in the query process. To accomplish our goal we introduce a novel user-adaptive learning algorithm that iteratively learns user preference regarding data quality. Within the process we assume that users have previously performed a filtering in non-quality attributes like space and time. Thus, our approach complements existing spatio-temporal queries, supporting the additional evaluation of quality attributes during the query process. It should also be mentioned that even though our motivation stems from geospatial image queries, the approach presented here could be applicable to other types of geospatial datasets as well. The paper is organized as follows: • Related Work: An overview of literature relevant to our work, • Data Quality Attributes: A discussion of the data quality attributes that are used to convey image data quality, P H OTO G R A M M E T R I C E N G I N E E R I N G & R E M OT E S E N S I N G Augus t 2004 9 7 3 Department of Spatial Information Science & Engineering, and National Center for Geographic Information and Analysis (NCGIA), University of Maine, 348 Boardman Hall, Orono, ME 04469-5711 ([email protected]; [email protected]; [email protected]; [email protected]). Photogrammetric Engineering & Remote Sensing Vol. 70, No. 8, August 2004, pp. 973–981. 0099-1112/04/7008–0973/$3.00/0 © 2004 American Society for Photogrammetry and Remote Sensing SHI-25.qxd 7/9/04 15:10 Page 973