Hierarchical object-based classification of dense urban areas by integrating high spatial resolution satellite images and LiDAR elevation data

In Portugal, updating municipal plans (1:10 000) is required every ten years. High spatial resolution imagery has shown its potential for detailed urban land cover mapping at large scales. However, shadows are a major problem in those images and especially in the case of urban environments. The purpose of this study is to develop a less time consuming and less expensive alternative approach to the traditional geographic data extraction for municipal plans production. A hierarchical object-oriented classification method, combining a multitemporal data set of high resolution satellite imagery and Light Detection And Ranging (LiDAR) data, is presented for the Municipality of Lisbon. A histogram thresholding method and a Spectral Shape Index (SSI) are initially applied to discriminate shadowed from non-shadowed objects using a 2007 QuickBird image. These non-shadowed objects are then divided into vegetated and non-vegetated objects using a Normalized Difference Vegetation Index (NDVI). Through a rule-based classification using the height information from LiDAR data, vegetated objects are classified into grassland, shrubs and trees while non-vegetated objects are distinguished into low and high features. Low features are then separated into bare soil and roads, again using a NDVI, while high features are classified as buildings and high crossroads using the shape of the objects (density). The 2007 shadowed objects are classified based on the spectral and spatial information of a 2005 QuickBird image, where shadows are in different directions. The developed methodology produced results with an overall accuracy of 87%. Misclassifications among vegetated features are due to the fact that the nDSM did not express the height for permeable features, while among non-vegetated features are due to temporal discrepancies between the DTM and the DSM, to different satellite azimuths in the 2005 and 2007 images and to contextual rules.