Multi-Spectral Source-Segmentation using Semantically-Informed Max-Trees

In this paper, we propose an innovative approach to multi-band source segmentation that addresses the constraints of single-band max-tree-based methods and effectively manages component-graph complexity. Our method extends multiple max-trees by integrating semantically meaningful nodes, derived from statistical tests, into a structured graph. This integration enables the exploration of correlations among cross-band emissions, enhancing segmentation accuracy.

Evaluation with artificial multi-band astronomical images shows our method's superior accuracy in detecting and segmenting multi-spectral imagery. We achieve 98 % accuracy in identifying correlated cross-band sources. Compared to state-of-the-art methods, our approach improves detection precision from 0.92 to 0.95 without sacrificing recall.
Furthermore, quantitative analysis demonstrates significant speed enhancements, particularly on 3-channel images sized at 1,000 pixels squared, our method achieves up to an approximately 31x acceleration when compared to a component-graph-based approach. The versatility and effectiveness of the proposed method suggest applications in remote sensing and multi-spectral large-scale image data analysis.