Introducing Shape Descriptors of Spectral Reflectance Curve (DSRC) for Improving Image Classification Accuracy

Abstract. Most algorithms used to extract information about objects in satellite images rely on data derived from groups of pixels (i.e. Training sites). The information space pertaining to individual pixels. However, remains largely underutilized. Here, we introduce a new approach to understanding pixel space by analyzing its shape descriptors derived from the spectral reflectance curve (SDSRC), and we illustrate an application of this concept. Six shape-based feature extraction and representation methods were explored. Among them, the Shape Center of Gravity (COG) or centroid, Area Under Curve (AUC), and the distance between the center of gravity to the Cartesian coordinate center are identified as parameters. We generated images based on these parameters, referred to as Pixel Shape Indices (PSI). PSI images were produced for a section of an OLI Landsat satellite image of the Los Angeles metropolitan area. Twenty-two datasets were created, combining original spectral bands with PSI images, and used for classification via the Maximum Likelihood classification scheme across four land use/land cover types. Classification Accuracy was assessed using the Kappa coefficient of agreement as well as the Hellden and Short coefficients. In most cases, results show improved values for all metrics, indicating enhanced classification performance across all land use / land cover categories due to the integration of PSI features. It is recommended that PSI images be incorporated into various image processing techniques- such as image fusion, filtering, enhancements, and texture analysis- for improved mapping and environmental monitoring.