A Zone-Based Approach To Identifying Urban Land Uses Using Nationally-Available Data

Abstract

Accurate identification of urban land use is essential for many applications in environmental study, ecological assessment, and urban planning, among other fields. However, because physical surfaces of land cover types are not necessarily related to their use and economic function, differentiating among thematically-detailed urban land uses (single-family residential, multi-family residential, commercial, industrial, etc.) using remotely-sensed imagery is a challenging task, particularly over large areas. Because the process requires an interpretation of tone/color, size, shape, pattern, and neighborhood association elements within a scene, it has traditionally been accomplished via manual interpretation of aerial photography or high-resolution satellite imagery. Although success has been achieved for localized areas using various automated techniques based on high-spatial or high-spectral resolution data, few detailed (Anderson Level II equivalent or greater) urban land use mapping products have successfully been created via automated means for broad (multi-county or larger) areas, and no such product exists today for the United States. In this study I argue that by employing a zone-based approach it is feasible to map thematically-detailed urban land use classes over large areas using appropriate combinations of non-image based predictor data which are nationally and publicly available. The approach presented here uses U.S. Census block groups as the basic unit of geography, and predicts the percent of each of ten land use types - nine of them urban - for each block group based on a number of data sources, to include census data, nationally-available point locations of features from the USGS Geographic Names Information System, historical land cover, and metrics which characterize spatial pattern, context (e.g. distance to city centers or other features), and measures of spatial autocorrelation. The method was demonstrated over a four-county area surrounding the city of Boston. A generalized version of the method (six land use classes) was also developed and cross-validated among additional geographic settings: Atlanta, Los Angeles, and Providence. The results suggest that even with the thematically-detailed ten-class structure, it is feasible to map most urban land uses with reasonable accuracy at the block group scale, and results improve with class aggregation. When classified by predicted majority land use, 79% of block groups correctly matched the actual majority land use with the ten-class models. Six-class models typically performed well for the geographic area they were developed from, however models had mixed performance when transported to other geographic settings. Contextual variables, which characterized a block group’s spatial relationship to city centers, transportation routes, and other amenities, were consistently strong predictors of most land uses, a result which corresponds to classic urban land use theory. The method and metrics derived here provide a prototype for mapping urban land uses from readily-available data over broader geographic areas than is generally practiced today using current image-based solutions.