Spatial accuracy measures of soft classification in land cover

Abstract

GIScience 2016 Short Paper Proceedings Spatial accuracy measures of soft classification in land cover N. Tsutsumida 1 and A. Comber 2 Graduate School of Global Environmental Studies, Kyoto University Email: naru@kais.kyoto-u.ac.jp Department of Geography, University of Leeds Email: a.comber@leeds.ac.uk Abstract Accuracy of land cover maps is important for map users. The soft classification of land cover has been developed for avoiding mixed pixel problem, however the proportional map is traditionally assessed only by a global measure, such as R-squared and root mean square error (RMSE), lacking local information of accuracy. We developed a way of local measures of accuracy employed by a geographically weighted (GW) model. GW-Rsquared and GW- RMSE are locally assessed a soft classification map of urban agglomeration as a case study. Lower accuracies are found at the edge of urban boundary surrounding the core of the urban area and such local information is valuable for a deeper understanding of spatial accuracy. 1. Introduction Land cover maps are important for those who are interested in climate change, biodiversity and anthropogenic impacts on terrestrial environments and the accuracy of the map is an important consideration. Traditionally land cover maps classified as categorized classes (hard classification) are assessed by building a confusion matrix that compares predicted and observed classes, with predicted classes derived from the classification and observed classes from independent validation data. Measures of user, producer and overall accuracy and kappa index are calculated from the matrix. The reliability of land cover data classified using continuous measures such as fuzzy set memberships (soft classification) is frequently assessed using measures such as R-squared and root mean square error (RMSE) (Chen et al. 2010; Tsutsumida et al. 2016; Yuan et al. 2008). However these measures only provide global measures of reliability and accuracy, and they do not take spatial configuration into account. Local assessments would be valuable for a deeper understanding of spatial accuracy. 2. Background Spatial accuracy assessments for hard classification of land use and land cover have been considered by some previous studies (Foody 2005; Pontius et al. 2011). In particular, geographically weighted (GW) logistic regression model have recently been developed (Comber et al. 2012; Comber 2013). These generate local confusion matrices at discrete location in the study area and generate spatially distributed estimates (surfaces) of user, producer and overall accuracy. However, little work has focused on spatially distributed accuracy measures for soft classifications. In this study we develop the spatial accuracy measures of soft classification by determining R-squared and RMSE locally. A GW model is applied to develop such measures spatially. 3. Materials A map of fractional impervious surface area (ISA) in Jakarta metropolitan areas, the biggest urban agglomeration in Indonesia, in 2012 was used in this study (Figure 1). This map was