Mapping urban landscape heterogeneity: agreement between visual interpretation and digital classification approaches

Visual interpretation of remotely sensed imagery has long been used for landscape pattern analysis. Few studies, however, have investigated human variation in estimates of within-patch composition for classification of those patches, particularly in urban settings. This paper compares the agreement of two approaches—visual interpretation and object-based—to estimate the proportion cover of landscape features within delineated patches, and investigates the spatial patterns of patches with large disagreement between the two approaches. The two approaches were compared for the Gwynns Falls watershed, Maryland, USA. Three methods were used to assess agreement: a traditional error matrix based procedure and two fuzzy methods, a plus-one modification of the traditional procedure, and a fuzzy set theory method. We found that while visual interpretation does not work effectively when patches contain a mix of different types of features, accuracy increases with patches that are either dominated by a specific feature, or do not contain a specific feature. The overall accuracies of estimates by visual interpretation also vary by features, ranging from 63.3% for pavement to 93.8% for bare soil. Patches with large disagreement between the two approaches cluster spatially at locations where the urban landscape is more structurally complex, suggesting the accuracy of visual interpretation may be affected by patch shape complexity, and the spatial configuration of the landscape features within the patches. These results provide important insights into the accuracy of thematic maps based on visual interpretation, not only for ecologists and managers who are using the maps, but also for those who produce the maps.