Efficacy of morphological approach in the classification of urban land covers.

Abstract

Understanding the often-heterogeneous land use land cover (LULC) in urban areas is critical for among others environmental monitoring, spatial planning and enforcement. Recently, several earth observation satellites have been developed with enhanced spatial resolution that provide for precise and detailed representation of image objects. This has generated new demand for enhanced processing capabilities. Thus, the need for techniques that incorporate spatial and spectral information in the analysis of urban LULC has drawn increasing attention. Enhanced spatial resolution comes with challenges for most pixel based classifiers. This include salt and pepper effects that arise from incapability of pixel based techniques in considering spatial or contextual information related to the pixel of interest during image analysis. These challenges have often contributed to the inaccuracy of heterogeneous LULC classification. Object based techniques on the other hand have been proposed to provide effective framework for incorporating spatial information in their analysis. However, challenges such as over/under segmentation and difficulty or non-robust statistical estimation hamper most object techniques in achieving optimum performance. Thus, to achieve optimum LULC classification, the full exploitation of both spectral-spatial information is essential. Hence, this study investigated the efficacy of Mathematical Morphological (MM) techniques referred to as morphological profiles (MP) in LULC classification of a heterogeneous urban landscape. The first objective of the study evaluated two MP techniques i.e. concatenation of morphological profiles (CMP) and multi-morphological profiles (MMP) in the classification of a heterogeneous urban LULC. Findings from this study indicated that both CMP and MMP provided higher accuracies in classifying a heterogeneous urban landscape. However, in evaluating their capability in preserving geometrical characteristics such as shape, theme, edge and positional similarity of image structures, CMP provided higher accuracies than MMP. This was attributed to the use of Principal Component Analysis (PCA) in the construction of MMP that resulted in the distorted edges of some of the image objects. However, in comparing the techniques in terms of the capability to discriminate image objects, MMP provided higher classification accuracies compared to CMP. This can be attributed to the former’s capability to exploit both spectral and spatial information from very high spatial resolution imagery. Hence in the second objective, MMP was adopted due to its ability to deal with dimensionality problem associated with CMP and its superior object discrimination capability. The findings indicated that MMP significantly enhanced ML and SVM classification accuracies. Specifically, the use of MMP as a feature vector for SVM and ML classification increased LULC distinction of objects with similar spectral signatures in a heterogeneous urban landscape. This is due to its capability to provide an effective framework for synthesis of spectral and spatial information. Overall the study demonstrated that morphological techniques provides robust novel image analysis techniques which can effectively be used for operational classification of a heterogeneous urban LULC.