Object-Based Urban Land Cover Extraction Using the Synergy of Lidar Data and Very High Resolution Multispectral Imagery

Abstract

Up-to-date urban land cover information plays a critical role for urban planning and management. In this study, an approach was presented for classifying urban land cover types using the integration of very high resolution (VHR) multispectral aerial imagery and airborne discrete return LiDAR (Light Detection and Ranging) data. The integration of aerial imagery and LiDAR data was conducted during object-oriented (OO) classification. Image segmentation prior to OO classification was performed using the Simple Non-Iterative Clustering (SNIC) algorithm, which is a state-of-the-art image segmentation algorithm that exhibits the advantages of efficiency and high accuracy. The features used in the classification consist of the optical bands of aerial imagery, an NDVI index and seven grey-level co-occurrence matrix (GLCM) texture metrics (contrast, dissimilarity, homogeneity, second moment, entropy, variance, correlation) calculated from the optical bands, one normalized digital surface model (nDSM) and one intensity band derived from LiDAR data. Adaptive Boosting (AdaBoost), a machine learning algorithm, was selected as the classifier. The sensitivitiy of AdaBoost to feature selection (FS), by applying recursive feature elimination (RFE) method, was also investigated. The methods were applied to fused VHR aerial imagery and LiDAR data of the city of Hradec Kralove, Czech Republic. Three sub-areas were chosen as the study areas. The results demonstrated that the fusion of aerial imagery and the LiDAR derived nDSM and intensity image features significatly improved the results (overall accuracy-OA) up to 24.7%. The highest classification accuracy achieved (OA = 85.5%) was based on the selected best features (21 features) from 56 input features. The second highest classification accuracy (OA = 84.8%) obtained was based on the fused dataset of aerial imagery and the LiDAR derived nDSM and intensity image features. The integrated dataset of aerial imagery and the LiDAR derived features proved to be effective in urban land cover classification. However, combining object-based GLCM texture measures in the AdaBoost classifier reduced the classification accuracy.