Multiple feature fusion in nonlinear canonical correlation analysis for semi-supervised land cover segmentation

Wei, H. ORCID: https://orcid.org/0000-0002-9664-5748, Xiao, J. and Hong, X. ORCID: https://orcid.org/0000-0002-6832-2298 (2026) Multiple feature fusion in nonlinear canonical correlation analysis for semi-supervised land cover segmentation. Applied Soft Computing, 201 (Part C). 115595. ISSN 1568-4946 doi: 10.1016/j.asoc.2026.115595

Abstract/Summary

Segmenting complex multi-spectral remote sensing images, such as urban areas, with limited training data is a challenging task in computer vision research. To address this challenge, this study introduces a novel semi-supervised clustering framework using multiple feature fusion and dimensionality reduction designed for effective segmentation. We propose a pipeline that first constructs a high-dimensional feature space by concatenating textural and spatial features at the pixel level, followed by dimensionality reduction using t-distributed Stochastic Neighbour Embedding (t-SNE). A core contribution of this study is the introduction of RBF-CCA, a modified Canonical Correlation Analysis (CCA) algorithm that leverages Radial Basis Functions (RBF) centred on limited labelled data to learn an optimal projection matrix. By projecting the RBF of the full image data using this learned canonical space, a robust partitioning via k-means clustering of canonical variables with a desired number of clusters is employed to segment the given image. Experimental results on several multi-spectral remote sensing benchmark images demonstrate that the proposed approach effectively captures spectral and textural similarities from highly sparse labelled data and significantly improves land cover segmentation accuracy compared to traditional schemes. It offers a novel solution for semi-supervised segmentation of high-dimensional imagery, specifically optimised for land cover data with limited labelling efforts.

Item Type	Article
URI	https://centaur.reading.ac.uk/id/eprint/130807
Identification Number/DOI	10.1016/j.asoc.2026.115595
Refereed	Yes
Divisions	Science > School of Mathematical, Physical and Computational Sciences > Department of Computer Science
Publisher	Elsevier
Download/View statistics	View download statistics for this item