Technical Summary: Generative Spectrum Cartography (GSC)

1. Problem Formulation

The paper addresses spectrum cartography, the task of reconstructing the spatial distribution of Power Spectral Density (PSD) over a target region from spatially incomplete measurements. This problem is particularly critical for satellite-based wide-area sensing, where data is often sparse, noisy, and low-bit quantized.

The authors identify two coupled challenges that existing methods typically address separately:

1. High-fidelity Reconstruction: Recovering the underlying PSD field from severely incomplete measurements. This is a severely ill-posed inverse problem, especially when measurements are subject to non-ideal factors like additive Gaussian noise and low-bit analog-to-digital conversion (quantization).
2. Efficient Sensing Resource Allocation: Determining which unobserved locations to measure next to maximize map quality under a limited sensing budget. Existing approaches often fail to use the current reconstruction's uncertainty to guide subsequent measurements, leading to redundant data collection.

The paper argues that classical priors (e.g., smoothness, sparsity) are insufficient for complex, multi-source spectrum patterns, and that existing diffusion-based inverse solvers (like DPS) are computationally expensive and unstable under low-bit quantized measurements due to their reliance on likelihood-gradient guidance.

2. Methodology

The authors propose Generative Spectrum Cartography (GSC), a unified Bayesian inference framework that integrates spectrum map reconstruction and active sensing using a Denoising Diffusion Probabilistic Model (DDPM) as a learned prior.

A. Measurement-Conditioned Posterior Inference

Instead of using iterative likelihood-gradient guidance (which requires costly gradient evaluations and is unstable for quantized models), GSC derives closed-form analytical posterior mean updates that are embedded directly into the reverse diffusion process.

• Linear Measurement Model: For linear measurements ($y = Hx_0 + n$), the posterior mean is updated using a correction term based on the residual between the measurement and the current estimate, scaled by the noise variance and diffusion parameters.
• Quantized Measurement Model: For low-bit quantized measurements ($y = Q(Hx_0 + n)$), the authors derive a closed-form update using the probability mass of the quantization intervals. This involves calculating the difference between the PDF and CDF of a standard normal distribution over the specific quantization interval defined by the measurement.
• Gradient-Free Sampling: By embedding these analytical updates into the reverse diffusion step ($x_{t-1} = \tilde{\mu}_t(x_t, \hat{x}_{0|t,y}) + \tilde{\sigma}_t z$), GSC achieves measurement-consistent posterior sampling without computing likelihood gradients.

B. Uncertainty-Aware Active Sensing

The framework utilizes the stochastic nature of the diffusion model to estimate spatial uncertainty and guide active sensing:

1. Uncertainty Estimation: The reconstruction process is run $M$ times with different random seeds to generate an ensemble of posterior samples. An empirical uncertainty map (variance) is computed from this ensemble.
2. Diversity-Aware Selection: To avoid selecting redundant measurements from clustered high-uncertainty regions, the authors employ a K-means clustering strategy. Unmeasured grid locations are represented as feature vectors combining normalized spatial coordinates and normalized uncertainty. The algorithm clusters these candidates and selects the most uncertain location from each cluster, ensuring spatial diversity in the additional sensing budget.

3. Key Contributions

The paper outlines four primary contributions:

1. Unified GSC Framework: A posterior inference framework that unifies spectrum map reconstruction and active sensing under a learned diffusion prior, establishing a closed-loop approach for spectrum cartography.
2. Closed-Form Posterior Updates: The derivation of analytical, measurement-conditioned posterior mean updates for both linear and quantized measurement models. This enables gradient-free, measurement-consistent sampling, avoiding the computational costs and instability of likelihood-gradient guidance.
3. Uncertainty-Aware Active Sensing: A strategy that estimates spatial uncertainty from posterior samples and uses diversity-aware selection (via K-means) to guide the allocation of additional sensing resources to informative, non-redundant locations.
4. Validation: Comprehensive evaluation on simulated electromagnetic maps and a high-fidelity satellite monitoring scenario, demonstrating superior performance in both reconstruction fidelity and sensing efficiency.

4. Experimental Results

The authors evaluated GSC against representative baselines (IDW, Nasdac, Dowjons, DPS, and UNN) using Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS).

• Reconstruction Performance:

  • Sparse/Noisy Data: Under sparse (down to 5% measurement ratio) and noisy conditions, GSC consistently outperformed both classical interpolation and other learning-based methods. For instance, at $\rho=0.05$ (noise-free), GSC achieved 27.51 dB PSNR compared to 25.26 dB for DPS.
  • Quantized Data: In low-bit regimes (1-bit to 3-bit), GSC showed significant gains over the UNN baseline. Under 1-bit quantization with a 15% measurement ratio, GSC improved PSNR from 10.95 dB (UNN) to 20.98 dB and reduced LPIPS from 0.491 to 0.183.
  • Satellite Scenario: In a high-fidelity simulated satellite monitoring scenario (MEO satellite monitoring LEO uplinks), GSC maintained high fidelity (37.35 dB PSNR at $\rho=0.10$) where the IDW baseline degraded significantly (25.30 dB).

• Active Sensing Efficiency:

  • The uncertainty-aware strategy consistently outperformed random sensing across all additional sensing budgets ($\Delta\rho$).
  • The improvement was observed in both linear and quantized settings, with the advantage becoming more pronounced as the bit-depth increased. The strategy successfully directed measurements toward weakly constrained regions, improving reconstruction quality more effectively than random selection.

5. Significance and Claims

The paper claims that GSC represents a significant advancement by addressing the coupled challenges of reconstruction and sensing within a single, tractable Bayesian framework.

• Robustness to Quantization: A key claim is that GSC's closed-form updates make it uniquely suitable for low-bit quantized measurements, a regime where gradient-based diffusion solvers often fail or become unstable.
• Closed-Loop Efficiency: By integrating uncertainty estimation directly into the sensing loop, the framework maximizes the value of limited sensing resources, which is critical for satellite-based monitoring where energy and bandwidth are constrained.
• Generative Priors: The work demonstrates that learned diffusion priors, when combined with analytical measurement conditioning, can effectively model complex, multi-source electromagnetic fields that traditional statistical priors cannot capture.

The authors conclude that while the current work focuses on fixed-frequency 2D maps, the framework provides a foundation for future extensions to spatiotemporal cartography and distributed multi-agent sensing.