Spectral Haar Binarizer: Compact Boolean Contrast Representations for Hyperspectral Classification

Michał Romaszewski; Przemysław Głomb

cs.LGOct 30, 2026

Spectral Haar Binarizer: Compact Boolean Contrast Representations for Hyperspectral Classification

Authors: Michał Romaszewski, Przemysław Głomb

Organizations: Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, Bałtycka 5, 44-100, Gliwice, Poland

Abstract

Tsetlin Machines offer an interpretable, hardware-friendly alternative to conventional machine learning models, but applying them to continuous hyperspectral measurements requires converting spectral signals into Boolean representations. Standard binarisation methods typically discretise each spectral band independently, potentially discarding important relationships between neighbouring wavelengths. In this paper, we introduce the Spectral Haar Binarizer (SHB), a contrast-based spectral representation method that constructs Boolean predicates from differences between selected spectral intervals. Inspired by Haar-like contrast features, the method uses an AdaBoost-style selection procedure to obtain a compact set of discriminative spectral predicates. We evaluate SHB on five heterogeneous hyperspectral datasets covering forensic blood detection, agricultural and urban remote sensing, and rainforest tree-species classification. Across binary and multiclass tasks, SHB consistently outperforms standard per-band binarisation strategies and provides the strongest or near-strongest results across classifiers. In binary tasks, TM-SHB achieves the best average performance, while RF-SHB is strongest in multiclass settings. Bayesian and rank-based analyses confirm that SHB's advantage is robust across datasets and not driven by isolated cases. The results show that representing continuous hyperspectral measurements is a central processing step for efficient classification. That contrast-based Boolean features provide an effective bridge between information-rich sensor measurements and compact, logic-based processing.

Figures & tables

Figure 1: Hyperspectral imaging of blood and blood-like substances, based on the dataset introduced in [ 31 ] . (a) Laboratory scene containing blood and visually similar red substances. (b) Absorbance spectra of blood, red paint, and artificial blood. Although these materials may appear similar in RGB imagery, their hyperspectral signatures reveal distinct spectral characteristics. In particular, the blood spectrum exhibits characteristic haemoglobin absorption features, including the $\alpha$ and $\beta$ bands around 576 nm and 542 nm, respectively, as well as a further feature near 950 nm. This illustrates the potential of hyperspectral imaging for discriminating substances and analysing their material properties beyond what is accessible from conventional RGB data.

Task	Method	Avg Acc	Avg BAcc	Avg F1
B	TM-SHB	$97.08\pm 0.29$	$94.89\pm 0.58$	$95.58\pm 0.43$
	RF-SHB	$96.81\pm 0.41$	$94.50\pm 0.73$	$95.31\pm 0.61$
	RF-Raw	$94.28\pm 0.58$	$90.50\pm 0.94$	$91.65\pm 0.82$
	RF-Std	$92.95\pm 0.72$	$88.11\pm 1.17$	$89.26\pm 1.01$
	RF-MO-T	$92.92\pm 0.65$	$87.77\pm 1.18$	$89.02\pm 0.93$
	RF-MO-OH	$92.58\pm 0.55$	$87.22\pm 1.11$	$88.55\pm 0.81$

Table 1: Averaged performance of binarizer-classifier pairs across all datasets. Shortcuts: B: Binary, M: Multiclass, MO: MultiOtsu, T: Thermometer, OH: OneHot, Quant: Quantile. Methods ordered by F1.

Task Type	Control ( $C$ )	Baseline ( $B$ )	$P(C>B)$	$P(C\approx B)$	$P(B>C)$
Binary	TM-SHB	RF-SHB	8.34%	91.08%	0.57%
		TM-Std	99.55%	0.00%	0.45%
		RF-Std	99.84%	0.00%	0.16%
		RF-Raw	99.46%	0.00%	0.54%
		TM-MultiOtsu-Thermometer	99.72%	0.00%	0.28%
		TM-KMeans-Thermometer	99.77%	0.00%	0.23%

Table 2: Comparison between TM-SHB pipeline and reference models in binary and multiclass tasks, using Bayesian methodology from [ 4 ] . Probabilities denote whether the control method is strictly better (

P(C>B)

), worse (

P(B>C)

), practically equivalent (

P(C\approx B)

). The value of ROPE is set to 1%

Panel of Figure 2 — Figure 2: Visualization of probability that TM-SHB is better, worse or practically equivalent, indicated by p(rope), than selected reference, based on Bayesian methodology from [ 4 ] . Note that compared classifiers use either the proposed binarisation scheme or continuous data, without binarisation.

Method	Macro-F1			Bal-Acc
	Bin.	Multi.	Total	Bin.	Multi.	Total
RF-SHB	26	15	41	24	15	39
TM-SHB	26	10	36	26	12	38
RF-Raw	17	12	29	15	11	26
RF-Std	12	4	16	9	4	13
RF-MultiOtsu-Thermometer	10	5	15	7	5	12

Table 3: Ranking table for tested binariser-classifier combination, in binary and multiclass experiments. Models with the highest number of points performed better across datasets.

Panel of Figure 3 — Figure 3: Ablation study: visualization of the impact of the number of bits parameter on binary (a) and multiclass (b) balanced accuracy. Panel (c) presents the impact of the number of features (filters) on the SHB method.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Task / Pair	Method	Test Acc	Test BAcc	Test F1
Multiclass	RF-SHB	$98.73\pm 0.16$	$97.66\pm 0.24$	$97.83\pm 0.22$
	TM-SHB	$98.31\pm 0.33$	$97.04\pm 0.48$	$94.86\pm 5.28$
	RF-Raw	$97.25\pm 0.15$	$96.13\pm 0.18$	$96.23\pm 0.18$
	RF-Quantile-OneHot	$92.44\pm 0.15$	$90.26\pm 0.28$	$90.36\pm 0.21$
	RF-Quantile-Thermometer	$92.41\pm 0.10$	$90.25\pm 0.18$	$90.37\pm 0.15$
	RF-Std	$92.36\pm 0.16$	$90.13\pm 0.24$	$90.25\pm 0.21$

Appendix

Table 4: Comprehensive experimental results for the HBE1 dataset (Mean

\pm

Standard Deviation across 5 CV folds). Ordered by BAcc.

Task / Pair	Method	Test Acc	Test BAcc	Test F1
Multiclass	RF-Raw	$99.46\pm 0.11$	$95.60\pm 0.54$	$96.04\pm 0.20$
	TM-SHB	$99.61\pm 0.04$	$95.60\pm 0.44$	$95.93\pm 0.35$
	RF-SHB	$99.61\pm 0.03$	$95.30\pm 0.38$	$95.85\pm 0.23$
	TM-KMeans-OneHot	$99.15\pm 0.14$	$95.13\pm 0.46$	$95.48\pm 0.21$
	TM-KMeans-Thermometer	$99.01\pm 0.38$	$95.10\pm 0.94$	$95.34\pm 0.82$
	TM-MultiOtsu-OneHot	$98.98\pm 0.09$	$94.86\pm 0.55$	$95.19\pm 0.28$

Appendix

Table 5: Comprehensive experimental results for the HBF1 dataset (Mean

\pm

Standard Deviation across 5 CV folds).

Task / Pair	Method	Test Acc	Test BAcc	Test F1
Multiclass	RF-SHB	$89.53\pm 0.70$	$86.50\pm 3.74$	$88.05\pm 2.39$
	TM-SHB	$86.51\pm 1.81$	$78.96\pm 3.42$	$81.92\pm 3.33$
	RF-Raw	$84.71\pm 0.68$	$77.96\pm 1.47$	$80.92\pm 1.58$
	RF-MultiOtsu-OneHot	$76.67\pm 0.63$	$71.34\pm 3.33$	$72.82\pm 3.35$
	RF-MultiOtsu-Thermometer	$76.44\pm 0.68$	$70.74\pm 3.46$	$72.59\pm 3.06$
	RF-KMeans-OneHot	$76.03\pm 1.06$	$69.82\pm 2.33$	$72.48\pm 2.22$

Appendix

Table 6: Comprehensive experimental results for the INPI dataset (Mean

\pm

Standard Deviation across 5 CV folds). Ordered by BAcc.

Task / Pair	Method	Test Acc	Test BAcc	Test F1
Multiclass	RF-SHB	$83.90\pm 0.69$	$77.05\pm 0.91$	$80.13\pm 0.90$
	TM-SHB	$77.28\pm 0.92$	$65.27\pm 1.40$	$66.90\pm 5.06$
	RF-Raw	$75.09\pm 0.45$	$63.81\pm 0.51$	$67.46\pm 0.62$
	RF-MultiOtsu-Thermometer	$68.80\pm 0.55$	$55.51\pm 0.54$	$59.39\pm 0.57$
	RF-MultiOtsu-OneHot	$68.77\pm 0.56$	$55.48\pm 0.57$	$59.37\pm 0.66$
	RF-Std	$68.38\pm 0.54$	$55.48\pm 0.85$	$58.76\pm 1.00$

Appendix

Table 7: Comprehensive experimental results for the MOSA dataset (Mean

\pm

Standard Deviation across 5 CV folds). Ordered by BAcc.

Task / Pair	Method	Test Acc	Test BAcc	Test F1
Multiclass	RF-SHB	$93.77\pm 0.30$	$90.88\pm 0.42$	$91.90\pm 0.41$
	RF-Raw	$92.71\pm 0.26$	$90.32\pm 0.33$	$91.39\pm 0.27$
	TM-SHB	$92.92\pm 0.30$	$88.66\pm 0.39$	$90.49\pm 0.43$
	RF-Quantile-OneHot	$85.25\pm 0.25$	$80.25\pm 0.30$	$81.23\pm 0.25$
	RF-Quantile-Thermometer	$85.23\pm 0.15$	$80.11\pm 0.20$	$81.11\pm 0.16$
	RF-Std	$83.56\pm 0.35$	$79.53\pm 0.49$	$80.61\pm 0.55$

Appendix

Table 8: Comprehensive experimental results for the PAVI dataset (Mean

\pm

Standard Deviation across 5 CV folds). Ordered by BAcc.

Hyperspectral Image Classification

Hyperspectral image classification assigns land-cover, material, or biological labels using the spectral and spatial information in image pixels.

See all 58 papers

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
SSFT: A Lightweight Spectral-Spatial Fusion Transformer for Generic Hyperspectral Classification
Apr 17, 2026Alexander Musiat, Nikolas Ebert, Oliver WasenmüllerRepresentation Learning Hyperspectral Imaging
Mannheim University of Applied Sciences, Germany
MBTI: A Multi-Branch Efficient Fine-Tuning Framework for Hyperspectral Image Classification with Foundation Models
Jul 14, 2026Mingzhen Xu, Haonan Guo, Di Wang +9Hyperspectral Image Classification Parameter-Efficient Fine-Tuning
School of Computer Science, Wuhan University, Wuhan, China. · State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, China. · China Petroleum Pipeline Engineering Corporation, Langfang, China, and also with Hebei Key Laboratory of Underground Energy Storage Technology, Langfang, China. +4

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

Hyperspectral Image Classification

Explore similar work

Adaptive Band Selection for Hyperspectral Classification with Spatially Disjoint Evaluation

SSFT: A Lightweight Spectral-Spatial Fusion Transformer for Generic Hyperspectral Classification

MBTI: A Multi-Branch Efficient Fine-Tuning Framework for Hyperspectral Image Classification with Foundation Models