Evaluating the attributes of remote sensing image pixels for fast k-means clustering

Abstract

Clustering process is an important stage for many data mining applications. In this process, data elements are grouped according to their similarities. One of the most known clustering algorithms is the k-means algorithm. The algorithm initially requires the number of clusters as a parameter and runs iteratively. Many remote sensing image processing applications usually need the clustering stage like many image processing applications. Remote sensing images provide more information about the environments with the development of the multispectral sensor and laser technologies. In the dataset used in this paper, the infrared (IR) and the digital surface maps (DSM) are also supplied besides the red (R), the green (G), and the blue (B) color values of the pixels. However, remote sensing images come with very large sizes (6000 × 6000 pixels for each image in the dataset used). Clustering these large-size images using their multiattributes consumes too much time if it is used directly. In the literature, some studies are available to accelerate the k-means algorithm. One of them is the normalized distance value (NDV)-based fast k-means algorithm that benefits from the speed of the histogram-based approach and uses the multiattributes of the pixels. In this paper, we evaluated the effects of these attributes on the correctness of the clustering process with different color space transformations and distance measurements. We give the success results as peak signal-to-noise ratio and structural similarity index values using two different types of reference data (the source images and the ground-truth images) separately. Finally, we give the results based on accuracy measurement for evaluating both the success of the clustering outputs and the reliability of the NDV-based measurement methods presented in this paper.