Spatial-spectral fuzzy k-means clustering for remote sensing image segmentation

Abstract

Spectral clustering is a clustering method based on algebraic graph theory. The clustering effect by using spectral method depends heavily on the description of similarity between instances of the datasets. Although, spechal clustering has been of significant interest in recent times, the raw spectral clustering is often based on Euclidean distance, and it is impossible to accurately reflect the complexity of the data. Despite having a well-defined mathematical framework, good performance and simplicity, it suffers from several drawbacks, such as it is unable to determine a reasonable cluster number and is sensitive to complex shape objects and noise elements. ln this paper, we present a new approach named spatial-spectral fuzzy clustering which combines spectral clustering and fuzzy clustering with spatial information of the data samples with their neighbors into a unified framework to solve these problems. The study consists of three main steps: Step 1, calculate the spatial information value of the pixels, step 2 applies the spectral clustering algorithm to change the data space from the color space to the new space, and step 3 clusters the data in new data space by fuzzy clustering algorithm. Experimental results on the remote sensing image were evaluated based on a number of indicators, such as IQI, MSE, DI and CSI, showing that it can improve the clustering accuracy and avoid falling into local optimum.