Band-Aware Fusion and Modulation

• Band-aware fusion and modulation are techniques that decompose signals into distinct frequency or spatial bands and process each with tailored methods.
• They employ adaptive gating, cross-band fusion strategies, and specialized modulation to optimize tasks like source separation, noise suppression, and data transmission.
• Recent studies demonstrate significant improvements in metrics such as accuracy, SNR, and complexity reduction across diverse applications including speech enhancement and hyperspectral imaging.

Band-aware fusion and modulation refer to a class of signal processing, machine learning, and communication methodologies that exploit the spectral or spatial band structure of signals to achieve improved fusion, representation, and processing. This approach explicitly leverages the differences and complementarities among frequency bands, sub-bands, or spectral groups for tasks such as source separation, noise suppression, multimodal fusion, and cross-band communications. Recent advances encompass domains as diverse as audio/speech enhancement, sound source localization, hyperspectral remote sensing, image fusion, and joint RF-optical communications.

1. Principles and Theoretical Foundations

Band-aware fusion involves the explicit decomposition of signals into frequency, spectral, or spatial bands or sub-bands, followed by selective or adaptive fusion and interaction among these components. The core principles are:

• Band Decomposition: Partitioning the input into bands (e.g., via STFT frequency bins, Haar wavelet sub-bands, or subcarrier groupings).
• Band-Specific Processing: Applying tailored feature extraction, enhancement, or modulation strategies to each band, reflecting the unique physical, perceptual, or statistical properties (e.g., low-band audio contains most perceptual information (Yu et al., 2022), low-frequency image bands encode structure (Zhang et al., 5 Sep 2025)).
• Cross-Band Fusion/Interaction: Techniques for information exchange or aggregation across bands, ranging from residual/additive skip connections and gating to deep attention-based fusion or learned affine modulations.
• Adaptive Modulation: In communications, band-aware modulation assigns resources (e.g., QAM order) to sub-bands based on SNR or channel conditions, maximizing efficiency and reliability (Song et al., 2023, Oikonomou et al., 21 Mar 2025).

Mathematically, band-aware operations are formalized through sequential or parallel mappings on band-decomposed tensors, often with band-dependent parameters and fusion rules.

2. Architectures and Implementation Strategies

A wide spectrum of architectures has been proposed for band-aware fusion and modulation, reflecting application-specific requirements:

• Alternating Full-Band and Narrow-Band Layers: In sound source localization, FN-SSL alternates BLSTM layers across the frequency and time axes, with additive skip connections to propagate context, achieving robust DP-IPD estimation (Wang et al., 2023).
• Dual-Path Transformers: In DPT-FSNet, intra-path transformers capture sub-band temporal dependencies, while inter-path transformers encode cross-band (full-band) correlations; their outputs are fused via gated convolutions (Dang et al., 2021).
• Sub-Band Cascades with Guided Modulation: For speech enhancement at high sampling rates, SF-Net splits the spectrum into contiguous bands, applies complex-domain enhancement to the perceptually critical low band, and uses knowledge-guided magnitude-only masking for higher bands, with learned gating for band interaction (Yu et al., 2022).
• Dual-Stream Band Order Networks: In HSLiNet, hyperspectral bands are reordered according to various strategies, and paired band orders are processed in separate streams, with features fused via adaptive FC-based modulation to leverage order-sensitive discriminative power (Yang et al., 27 Mar 2025).
• Frequency/Spatial-Domain Joint Fusion: The GD^2Fusion framework in infrared-visible image fusion integrates frequency-domain wavelet-based band selection and VLM-guided affine gating with spatial-domain attention and multi-scale aggregation, coordinated via end-to-end training (Zhang et al., 5 Sep 2025).
• Hybrid Multiband Modulation in Communications: FTN-NOFDM with NOM-p divides available bandwidth into L sub-bands, each assigned a modulation order matched to its channel SNR. At the receiver, sub-band decisions are fused (typically concatenated) to reconstruct the data stream (Song et al., 2023). For cross-band RF-optical systems, mappings are optimized linearly or by DNN to maximize MI and minimize SEP (Oikonomou et al., 21 Mar 2025).

The table below summarizes representative architectural motifs:

Domain	Band-Aware Decomposition	Fusion/Modulation Mechanism
Audio/Speech	Frequency bands (STFT, sub-bands)	Alternating/recurrent layers, residual/gating, cross-band attention (Wang et al., 2023, Dang et al., 2021, Yu et al., 2022)
Remote Sensing	Spectral permutations (HSI)	Dual-stream by band order, adaptive fusion (Yang et al., 27 Mar 2025)
Image Fusion	Wavelet sub-bands (DWT)	VLM-guided affine gating, multi-scale conv/attention (Zhang et al., 5 Sep 2025)
Communications	Subcarrier groups, cross-bands	Adaptive-QAM per band, O(1) detection, DNN-optimized mapping (Song et al., 2023, Oikonomou et al., 21 Mar 2025)

3. Modulation and Gating Mechanisms

Band-aware modulation—distinct from conventional signal modulation—refers to the dynamic adjustment or gating of information flow according to band-specific cues or contextual knowledge:

• Simple Additive/Residual Gating: As in FN-SSL, additive skip connections modulate band representations without parameterized gates (Wang et al., 2023).
• Learned Sigmoid/FC Gating: Sub-band interaction modules in SF-Net employ learned sigmoidal gates to dynamically weigh low-band contributions to higher-band processing (Yu et al., 2022); HSLiNet uses post-summation FC layers to adaptively reweight fused order streams (Yang et al., 27 Mar 2025).
• VLM-Guided Affine Modulation: In GD^2Fusion, CLIP prompt embeddings generate affine parameters for band-wise gating, aligning band selection/refinement with semantic degradations (Zhang et al., 5 Sep 2025).
• Nonlinear Gating in Transformers: DPT-FSNet applies elementwise tanh/sigmoid gated convolutions after dual-path aggregation, blending frequency and temporal features (Dang et al., 2021).
• Modulation Order Adaptation: In multi-band communications, QAM levels are allocated to sub-bands as a function of per-band SNR (i.e., $$Q_l = \max\{q: \mathrm{SNR}(f_{\rm center},l) \ge \gamma_q\}$$) (Song et al., 2023).
• Cross-Band Symbol Mapping: In hybrid RF-optical communications, linear combinations or DNN-generated mappings project RF constellations into optical intensity values, subject to tractable detection and MI/SEP optimization (Oikonomou et al., 21 Mar 2025).

These gating paradigms share the implicit aim of maximizing downstream discriminative power, fidelity, or efficiency via band-aware decision making.

4. Quantitative Impact and Empirical Results

Across domains, band-aware fusion and modulation confer demonstrable empirical advantages:

• Sound Source Localization: FN-SSL achieves 86.7% accuracy within ±5° (mean error 3.2°) at SNR=0 dB, surpassing full-band baselines by 6.3% and reducing MAE by 0.6° (Wang et al., 2023).
• Speech Enhancement: SF-Net improves PESQ to 3.02 vs. 2.78 (dual-stream DS-Net) and achieves 19.90 dB SDR, while ablation shows the sub-band interaction module yields up to 0.89 dB SSNR gain (Yu et al., 2022). DPT-FSNet, using full-band/sub-band transformers, outperforms previous SOTA in PESQ/SI-SDR (Dang et al., 2021).
• Hyperspectral Classification: HSLiNet's dual-order fusion achieves +1.0–1.5% OA improvement on Houston 2013, and +3–5% OA in few-shot scenarios, setting new SOTA (OA 0.9989) (Yang et al., 27 Mar 2025).
• Image Fusion: GD^2Fusion reports qualitative and quantitative superiority under dual-source degradations, directly attributing gains to frequency-domain sub-band gating and spatial-domain multi-modal fusion (Zhang et al., 5 Sep 2025).
• Communications: Adaptive-multiband FTN-NOFDM reduces BER by over 2× (from $5.04\times10^{-4}$ to $2.54\times10^{-4}$ for $L=3$ bands) and delivers 97% complexity reduction vs. single-band designs (Song et al., 2023). Cross-band modulation with DNN-Gen offers ≥2 dB SNR gains in MI and SEP over linear mapping or conventional cross-band PAM (Oikonomou et al., 21 Mar 2025).

5. Algorithmic Design Considerations

Optimal design of band-aware fusion and modulation systems depends on several factors:

• Band Partitioning Strategy: Must balance per-band SNR, energy concentration, or task-specific feature localization. For speech, band cutoffs are chosen to separate harmonic/perceptual structure from noise (Yu et al., 2022); for HSI, band order permutations can reveal discriminative patterns (Yang et al., 27 Mar 2025).
• Fusion Symmetry and Regularization: Dual-stream or multi-order aggregation imposes symmetry/invariance, which can regularize and improve generalization (e.g., treating ascending/descending orders identically).
• Complexity vs. Performance Trade-off: Coordinated band-wise processing enables smaller, specialized modules and reduced decoding complexity in communications (down to 2.7% of single-band cost), while maintaining or improving aggregate performance (Song et al., 2023).
• Adaptation to Degradation: In multimodal/image fusion, VLM-guided modulation and explicit degradation perception allow selective enhancement or suppression of band features compromised by specific modalities' degradations (Zhang et al., 5 Sep 2025).
• Loss Functions and Supervision: Hybrid objectives may integrate time-domain, spectral-domain, or band-consistency terms to reinforce fidelity and band-aware inheritance.

6. Broader Contexts and Future Prospects

Band-aware fusion and modulation continue to expand across domains:

• Remote Sensing: Band-order insights and adaptive multi-order fusion may generalize to other multispectral/multimodal sensing scenarios (Yang et al., 27 Mar 2025).
• Audio/Acoustics: Alternating and cascaded band-aware architectures demonstrate greater robustness to non-stationary noise and reverberation (Wang et al., 2023, Yu et al., 2022).
• Communications: Cross-band and multi-band modulation strategies are becoming increasingly critical for managing the heterogeneity of hybrid RF-optical, VLC, or THz systems, where channel conditions and SNR vary widely by frequency (Oikonomou et al., 21 Mar 2025).
• Vision-Language Integration: The use of VLMs for frequency-domain guidance and band-aware gating, as in GD^2Fusion, points to stronger integration of semantic/linguistic priors for adaptive fusion (Zhang et al., 5 Sep 2025).

A plausible implication is that as sensor, communication, and computation paradigms become more heterogeneous, explicit modeling of band structure—via band-aware fusion and modulation—will play an increasingly central role in bridging performance, robustness, and efficiency.

7. Summary Table of Core Band-Aware Fusion/Modulation Strategies

Application Domain	Core Band-Aware Method	Empirical Benefit	Reference
Speech Enhancement	Sub-band cascades w/ gating	+0.24 PESQ, +1 dB SDR	(Yu et al., 2022)
HSI-LiDAR Fusion	Dual-order band-wise feature fusion	+1–1.5% OA (overall), +3–5% OA (few-shot)	(Yang et al., 27 Mar 2025)
Sound Source Localization	Alternating full/narrow-band BLSTM	−0.6° MAE, +6% acc vs. baseline	(Wang et al., 2023)
Infrared/Visible Image Fusion	Wavelet band gating + spatial fusion	Robust under degradation, SOTA fusion	(Zhang et al., 5 Sep 2025)
FTN-NOFDM Optical	Sub-band adaptive QAM mapping	97% complexity reduction, 2× BER gain	(Song et al., 2023)
RF-Optical Comms	Cross-band linear/DNN symbol mapping	≥2 dB MI/SEP gain at SOTA complexity	(Oikonomou et al., 21 Mar 2025)