Band-Selection Stability and Semantic Segmentation Performance: A Study on Hyperspectral City

Jiarong Li; Imad Ali Shah; Enda Ward; Martin Glavin; Edward Jones; Brian Deegan

cs.CVSep 25, 2026

Band-Selection Stability and Semantic Segmentation Performance: A Study on Hyperspectral City

Authors: Jiarong Li, Imad Ali Shah, Enda Ward, Martin Glavin, Edward Jones, Brian Deegan

Organizations: School of Engineering and Ryan Institute, University of Galway, Ireland · Valeo Vision Systems, Tuam, Ireland

Abstract

Resource constraints make high-dimensional hyperspectral imaging challenging in autonomous perception, motivating the use of band selection methods. However, the sensitivity of band-selection methods to sampled data and their relationship to semantic segmentation models (SSMs) remain underexplored. This study evaluates six band selection methods on ten independently sampled, class-balanced region-of-interest (ROI) sets, yielding 60 top-25 band subsets from the Hyperspectral City V2 (128 bands: 450-950nm) dataset. Top- $K$ bands ( $K\in\{3,5, ... 13\}$ ) from the first three ROI sets are evaluated with three SSMs against the corresponding 128-band baseline. Experiments show that intra-method stability is method-dependent: Sim-LP shows the highest stability (pairwise Jaccard similarity) and, together with JMIM+CSNR, yields the best segmentation results. Top- $K$ based SSMs remain competitive with baselines, with gains of up to 2.01 mIoU and 1.72 mF1 points, and 18-22x faster CPU inference for $K=9$ . However, performance does not improve monotonically with $K$ , and stability shows no consistent association with SSM performance. These findings suggest that intra-method stability is informative but an unreliable indicator of downstream segmentation performance, highlighting the need to evaluate band-selection methods across repeated samples, subset sizes, and SSMs.

Figures & tables

Fig. 1 : Method-specific top-25 band recurrence across ten class-balanced ROI sets. Color indicates selection count.

Fig. 2 : Intra-method pairwise Jaccard similarities of top-25 subsets across ten ROI sets. The dashed line marks the random-subset expectation ( $0.11$ ).

Model	Set Size	CMIM		JMIM		JMIM+CSNR		JMIM+CSNR+Corr		mRMR		Sim-LP		Freq.		128-band
		mIoU	mF1	mIoU	mF1	mIoU	mF1	mIoU	mF1	mIoU	mF1	mIoU	mF1	mIoU	mF1	mIoU	mF1
U-Net	3	70.73 $\pm$ 1.23	79.68 $\pm$ 1.32	68.43 $\pm$ 0.32	77.95 $\pm$ 0.25	68.79 $\pm$ 0.39	78.01 $\pm$ 0.31	67.76 $\pm$ 1.75	77.72 $\pm$ 1.23	68.62 $\pm$ 0.48	77.63 $\pm$ 0.69	69.88 $\pm$ 1.01	79.31 $\pm$ 0.73	69.66 $\pm$ 1.27	78.94 $\pm$ 0.88	70.67 $\pm$ 0.61	79.56 $\pm$ 0.47
	5	70.06 $\pm$ 0.61	79.34 $\pm$ 0.43	69.94 $\pm$ 0.61	78.94 $\pm$ 0.73	68.51 $\pm$ 0.49	77.98 $\pm$ 0.23	68.03 $\pm$ 1.83	77.74 $\pm$ 1.61	70.79 $\pm$ 0.76	79.85 $\pm$ 0.25	70.01 $\pm$ 0.77	79.67 $\pm$ 0.60	68.97 $\pm$ 0.85	78.65 $\pm$ 0.77
	7	70.61 $\pm$ 0.31	79.90 $\pm$ 0.20	70.02 $\pm$ 0.23	79.25 $\pm$ 0.95	70.53 $\pm$ 1.19	79.52 $\pm$ 0.88	66.64 $\pm$ 0.85	76.56 $\pm$ 0.67	70.01 $\pm$ 0.29	79.50 $\pm$ 0.12	70.39 $\pm$ 1.36	79.92 $\pm$ 1.20	69.97 $\pm$ 1.32	79.11 $\pm$ 1.42
	9	70.77 $\pm$ 0.37	79.66 $\pm$ 0.50	70.34 $\pm$ 1.28	79.63 $\pm$ 1.07	71.10 $\pm$ 1.09	80.10 $\pm$ 0.94	67.98 $\pm$ 0.98	77.84 $\pm$ 0.79	70.91 $\pm$ 1.04	79.58 $\pm$ 1.09	71.05 $\pm$ 0.32	80.26 $\pm$ 0.50	69.96 $\pm$ 0.17	79.01 $\pm$ 0.15
	11	70.02 $\pm$ 1.75	79.34 $\pm$ 1.10	68.67 $\pm$ 1.23	78.33 $\pm$ 1.09	70.34 $\pm$ 0.43	79.81 $\pm$ 0.55	68.15 $\pm$ 0.38	77.96 $\pm$ 0.72	69.89 $\pm$ 0.27	79.13 $\pm$ 0.52	69.09 $\pm$ 1.23	78.88 $\pm$ 1.05	68.82 $\pm$ 1.10	78.11 $\pm$ 0.92

Table 1: Mean

\pm

SD mIoU and mF1 (%) by SSM, method, and subset size. Individual methods use nine evaluations, while frequency-consensus and baseline entries use three. Cell color encodes the signed difference from SSM-specific 128-band baseline (green: positive; red: negative; white: parity), and bold marks each SSM-metric maximum.

Fig. 3 : CPU input-layer inference time and averaged mIoU/mF1 across methods in Table 1 : Annotations indicate input and 128-band points. Inference used batch size 1, 10 warm-up iterations, excluded data loading, and was benchmarked on Intel Xeon Gold 6252, Debian 12.

Fig. 4 : Mean pairwise Jaccard similarity versus baseline-relative mIoU/mF1 differences, averaged across SSMs and subset sizes. Differences are reported in percentage points. Dashed lines mark the random-subset expectation ( $0.11$ ) and baseline parity. Shading indicates positive performance differences and stability below the random-subset reference.

Explore similar work

Adaptive Band Selection for Hyperspectral Classification with Spatially Disjoint Evaluation
Jun 4, 2026Ikram El-Hajri, Ouassim Karrakchou, Alejandro MousistHyperspectral Imaging Feature Selection
International University of Rabat, Rabat, Morocco · Thales Alenia Space, Spain
A PyTorch Library for Hyperspectral Image Models: Technical Report
Sep 30, 2026Tanishq Rachamalla, Aryan Das, Srishti Kaushik +1Remote Sensing Benchmarks Hyperspectral Image Classification
Department of Information Technology Siddhartha Academy of Higher Education Vijayawada, Andhra Pradesh 521108, India · Department of Computer Science and Engineering Vellore Institute of Technology Bhopal, Madhya Pradesh 466114, India · Department of Computer and Information Sciences Indira Gandhi National Open University New Delhi 110068, India +1
HyperImageNet: A Large-Scale High-Spatial Resolution Hyperspectral Imagery Classification Benchmark
Jul 23, 2026Chuguang Zeng, Jingtao Li, Yinhe Liu +1Remote Sensing Image Segmentation Land Cover Classification
Wuhan University 430072 Wuhan, China

Jun 4, 2026J/K move · Enter open · S save

Adaptive Band Selection for Hyperspectral Classification with Spatially Disjoint Evaluation

Ikram El-Hajri, Ouassim Karrakchou, Alejandro Mousist

International University of Rabat, Rabat, Morocco · Thales Alenia Space, Spain

Hyperspectral Imaging Feature Selection Hyperspectral Image Classification

Open paper AskPDF arXiv

Hyperspectral band selection methods based on differentiable selectors can be sensitive to initialization and to extracting a final discrete subset, while prescribed band counts limit flexibility. We propose SGBR-HC (Spectral-Group Band Ranking with Hard-Concrete initialization), a two-stage method that uses a supervised spectral ranking to initialize trainable sparse gates rather than treating ranking as a fixed selection rule, letting the number of selected bands be determined by training. Stage-1 scores candidate bands from training pixels by class discriminability and spectral diversity; this ranking seeds the gate logits for Stage-2, which trains the sparse gates jointly with a spatial classifier. Under spatially disjoint evaluation on Pavia University and Houston 2013, verified by retraining a fresh classifier on the selected bands, SGBR-HC achieves the highest mean overall accuracy and Cohen's kappa with approximately twenty bands. Bypassing Stage-1 degrades OA by 8.84 pp on Pavia University and 22.15 pp on Houston 2013, confirming the ranking prior's role. Random pixel splits inflate OA on Pavia University by 30.56 pp, underscoring spatial leakage as a critical evaluation confound.

Abstract

Explore similar work

Adaptive Band Selection for Hyperspectral Classification with Spatially Disjoint Evaluation

A PyTorch Library for Hyperspectral Image Models: Technical Report

HyperImageNet: A Large-Scale High-Spatial Resolution Hyperspectral Imagery Classification Benchmark