Classification of Remote Sensing Imagery Based on Density and Fuzzy c-Means Algorithm

Introduction

Remote sensing data clustering is an extremely important part of satellite image processing (Torahi & Chai, 2011; Mai, Trinh & Ngo, 2016; Mai & Ngo, 2015; Mai & Ngo, 2018). The results of satellite image classification can be used for a variety of purposes, such as natural resource research and environmental monitoring, urban planning and ensure national defense and security. Meanwhile, optical remote sensing data sources are often affected by weather conditions and the accuracy of the receiver, this make the image classification more complicate (Ngo, Mai & Pedrycz, 2015). In fact, uncertainty is inherently present in decision making. As such, it is increasingly imperative to research and develop new theories and methods based on fuzzy clustering (Li, 2017).

There are many satellite image classification methods (Han, Chi & Yeon, 2005; Gordo, Martinez, Gonzalo & Arquero, 2013), such as manual thresholds methods (Yang et al., 2016), unsupervised classification methods (Genitha & Vani, 2013), supervised classification methods (Jog and Dixit, 2016), fuzzy clustering method (Rauf, Valentin & Leonid, 2009) and method based on intuitionistic fuzzy sets (Li, 2004; Li & Cheng, 2002). These methods often use some common algorithms, such as K-means, c-Means, Iso-data, minimum distance and Fuzzy c-means. These clustering algorithms are essentially using the same strategy based on brightness to split into clusters (Jog & Dixit, 2016; Rauf et al., 2009) without regard to the density of the pixels, while high density pixels are usually located near the centroid of the cluster (Peherstorfer, Pflüger & Bungartz, 2012; Chen, Yan & Wang, 2014; Benmouiza & Cheknane, 2016).

Many scientists in the field of remote sensing data processing have proposed clustering methods based on density of pixels, in which density based spatial clustering of applications with noise (DBSCAN) is commonly used for satellite image classification (Khan, Rehman, Aziz, Fong & Saravady, 2014; Benmouiza & Cheknane, 2016). This algorithm requires only one input parameter and supports the users in determining an appropriate value for it. It discovers clusters of arbitrary shape and divides high density areas into cluster without depend on the size of data. In terms of implementation, this algorithm is also difficult to find the optimal radius of the density function around each pixel. In addition, the execution time of this algorithm is quite slow, especially when tested on large datasets, such as satellite imagery (Ngo, Mai & Nguyen, 2012). To overcome these limitations, many scientists are interested in improving this algorithm. Peherstorfer et al. (2012) presented a grid-based density estimation method to improve the speed of clustering. Chen et al. (2014) improved the DBSCAN algorithm by expanding the clusters which uses the margins of the objects, such as a pixel, to reduce the computation time. These improvements significantly reduce clustering time; however, affect the accuracy of clustering results.

To solve the above problem, this study proposed a method for approximating the centroid of cluster based on the density of pixels. Next step, the authors use approximation centroids for classification satellite imagery using the fuzzy c-means algorithm.