Spending a Leased Band on the Right Half of a Hyperspectral Image

Drones mapping a disaster zone often carry two cameras that see the scene differently. One camera splits light into many narrow colors (a hyperspectral camera) and therefore produces blurry, coarse pixels. The other records only a few broad colors and produces sharp pixels. Back on the ground, a fusion algorithm can combine the two into an image that is both sharp and spectrally rich. The drones need radio spectrum to send this data home, and one way to get it is cooperative spectrum leasing: a drone relays a struggling licensed user's traffic, and in return the licensed user lets the drone fleet use the band for part of each frame. Existing leasing schemes treat the drone's imagery as an undifferentiated pile of bits and split the frame at a fixed ratio. We propose to treat two choices as one decision under a single airtime budget. The first is how much of the frame to spend relaying versus transmitting. The second is how much spatial and spectral resolution each camera stream should carry. Reconstruction quality at the ground fusion stage is the objective.

Put plainly, the trade-off between spatial and spectral resolution that limits the camera hardware reappears as a trade-off over transmitted bits. Fusion methods already know how to recover the full image cube from a coarse-but-colorful image and a sharp-but-few-colored image. So the cheapest way to deliver a hyperspectral scene over a leased band may be to send deliberately degraded versions of both streams and let fusion undo the degradation. How much to degrade should depend on how much airtime the relaying deal leaves over.

Core Claim

For a fixed primary-user protection requirement, choosing the relay time fraction together with the spatial and spectral downsampling factors of the two transmitted streams gives better fused-image fidelity per leased second than two alternatives. One is a fixed lease split followed by conventional compression of the full cube. The other is a throughput-optimal lease split followed by a separately chosen compression scheme. Fidelity is measured by reconstruction SNR and by spectral angle, which is the angle between true and recovered per-pixel spectra and therefore measures color distortion.

Let $\alpha$ be the fraction of a frame of length $T$ spent relaying primary traffic. Let $s$ be the spatial downsampling factor applied to the hyperspectral stream and $b$ the number of bands kept in the sharp stream. Let $D(s,b,R)$ be the fused-image distortion when both streams are coded into $R$ bits. The problem is then

$$ \min_{\alpha,s,b}\; D(s,b,R) \quad \text{s.t.}\quad R \le (1-\alpha)\,T\,W\log_2(1+\gamma_{S}),\qquad R_P(\alpha) \ge R_P^{\min}, $$

where $W$ is the leased bandwidth, $\gamma_S$ is the secondary link's signal-to-noise ratio, and $R_P(\alpha)$ is the primary user's rate with relaying. The primary rate grows with $\alpha$ and stops growing once relaying has done its work, which matches the queue-clearing results in Protocol Design and Stability Analysis of Cooperative Cognitive Radio Users. In this formulation, the image side enters the leasing decision only through the curve $D(\cdot)$. We claim that curve is far from flat and far from what a bit-count objective assumes.

Prior Work and What It Leaves Open

The motivating gap is clearest in the UAV leasing work. An Autonomous Spectrum Management Scheme for Unmanned Aerial Vehicle Networks in Disaster Relief Operations has one drone relay for ground users while the rest of the fleet downlinks mission imagery. The frame is fixed at 0.5 relaying and 0.5 transmitting, and the imagery is modeled as raw data volume. Distributed Cooperative Spectrum Sharing in UAV Networks Using Multi-Agent Reinforcement Learning learns how many drones relay versus sense, using a reward of primary throughput, secondary throughput and a fairness index. Neither paper asks what the secondary bits are for. We keep their leasing model and simulation settings but replace secondary throughput with fused-image distortion. We do not tackle their open mobility problem.

Other leasing papers supply alternative ways to pay for spectrum that our formulation can absorb in place of $\alpha$. Cooperative Spectrum Sharing Between D2D Users and Edge-Users: A Matching Theory Perspective uses a three-phase broadcast, relay and private-transmission frame. Reciprocally Benefited Spectrum Access Scheme With Joint Power and Subcarrier Allocation in a Software-Defined Network allocates subcarriers and power. Cognitive Non-Orthogonal Multiple Access with Cooperative Relaying superimposes secondary data on relayed primary data. All three optimize rate, not what the rate delivers.

On the imaging side, the fusion methods supply the decoder. Hyperspectral Super-Resolution: A Coupled Nonnegative Block-term Tensor Decomposition Approach (CNN-BTD) recovers the full cube by factoring it into a few low-rank abundance maps and nonnegative spectral signatures, with recovery guarantees. Hyperspectral Super-Resolution: A Coupled Tensor Factorization Approach shows that recovery remains identifiable even when the blur is unknown. That matters here because we choose the degradation ourselves and therefore know it exactly. Multispectral and Hyperspectral Image Fusion by MS/HS Fusion Net (MHF-net) builds the known blur and band-averaging operators into an unrolled optimization network. SSIF learns a continuous function of position and wavelength that generalizes to unseen scale factors. These methods were designed to undo hardware limits. We reuse them to undo choices made deliberately for transmission.

SSIF's reported ability to stay accurate at 14x scale after training only up to 8x is what makes a variable $s$ practical: a single decoder can serve every downsampling factor the optimizer picks. MHF-net's authors also note that fusion assumes the two images are perfectly registered. That constrains the design: we transmit both streams from a single drone carrying co-mounted cameras, not from two drones.

The nearest structural cousin is Spectrum Cartography via Coupled Block-Term Tensor Decomposition. It applies the same block-term model to radio maps but operates in a passive sensing setting with no leasing. We do not pursue that symmetry. The file suggests it is an analogy rather than a mechanism.

Research Program

The first step is to measure $D(s,b,R)$ empirically. On Indian Pines, Chikusei and Pavia Centre, we generate pairs following Wald's protocol, the standard way of producing matched coarse and sharp images by synthetically blurring and downsampling a reference cube. We use a grid of spatial factors and kept-band counts, then code each stream at a range of bit rates with a standard per-band image codec. The pair is fused with CNN-BTD, MHF-net and SSIF, and we record reconstruction SNR and spectral angle against the original cube. The result is a family of rate-distortion curves for "send a degraded pair and fuse." We compare them with the same codec applied to the full cube.

The second step embeds these curves in the two UAV leasing simulators from the file, the 81×81 disaster grid and the 100 m × 100 m Rayleigh-fading swarm. For each channel realization we solve the problem above. The search over $(s,b)$ is a small discrete grid, and $\alpha$ is a one-dimensional search, so the joint solve is cheap. We also train a lightweight lookup policy that maps measured link quality to $(\alpha,s,b)$, so the decision can run on the drone without new feedback signaling. The third step adds the deployment problem the file flags as open, channel bit errors on the two streams. Here we test whether losing part of the hyperspectral stream hurts fusion less than losing the same fraction of a compressed full cube.

Baselines

The naive reference is the scheme in An Autonomous Spectrum Management Scheme for Unmanned Aerial Vehicle Networks in Disaster Relief Operations as published: a 0.5/0.5 frame with the full hyperspectral cube compressed by a per-band codec into whatever bits fit. We run it in the same simulator with identical channel draws and primary protection threshold. Beating it means higher fused reconstruction SNR and lower spectral angle at equal primary rate. No reported image-quality result exists on Lacuna for this baseline, because the paper reports only training time and exploration counts.

The strongest competitor on the communication side is a separable design. We first choose $\alpha$ to maximize secondary throughput subject to the primary constraint, as the MARL reward in Distributed Cooperative Spectrum Sharing in UAV Networks Using Multi-Agent Reinforcement Learning effectively does. We then pick the best compression for the resulting bit budget, including the degraded-pair option. If this matches our joint design, the coupling claim fails. It is the decisive comparison.

The strongest competitor on the imaging side is fusion itself at full resolution, used as an upper anchor. CNN-BTD on Indian Pines reaches a reconstruction SNR of 27.17 dB and a spectral angle of 0.0352 against 25.38 dB for STEREO (CNN-BTD). MHF-net reaches 37.23 dB PSNR on CAVE against about 33.59 dB for coupled nonnegative matrix factorization (MHF-net). Our pipeline should approach these numbers as the lease budget grows. Measuring how fast it degrades as the budget shrinks is the result of interest.

Three ablations isolate the ingredients: optimized $(s,b)$ with fixed $\alpha=0.5$; optimized $\alpha$ with fixed $(s,b)$; and the joint optimization using a black-box ResNet decoder instead of a fusion decoder that knows the degradation. All decoders receive identical training data and tuning budgets.

Why not just…?

Why not just compress the full hyperspectral cube with a good codec and send it? This is the real threat. Choosing degradation factors is itself a crude, structured form of compression, and a strong codec may exploit spectral redundancy just as well. We have found no reported result on Lacuna for hyperspectral codecs in this setting. The first experiment therefore compares the two rate-distortion curves directly at matched bits. If the full-cube codec dominates at every rate, the proposal reduces to choosing $\alpha$ against a codec's distortion curve. That is still distortion-aware leasing, but without the resolution-trade-off mechanism, and we would report it that way. Our expectation that fusion wins at low rates rests on the low-rank structure that CNN-BTD exploits: a few signatures and abundance maps describe the cube, and a coarse cube plus a sharp image pins them down cheaply. This is an expectation, not an established result.

Why not just maximize throughput in the lease and let the imaging pipeline cope? If distortion falls steadily as bits increase, maximizing bits is close to optimal. It stops being optimal when the primary constraint and fusion quality saturate at different points. In that case, extra relay time that buys the primary user little could have bought much better spectral fidelity, or the reverse. The separable baseline above tests exactly this. We will report the regimes where the gap disappears.

Why not just send the sharp image only, or a hand-picked subset of bands? For mapping or detection, a sharp few-band image might be enough. The downstream classification gains SSIF reports on Pavia Centre, 1.7% to 7% over LIIF, suggest that spectral detail matters for land-use labels. To check this directly, we include sharp-only and band-subset points in the rate-distortion grid and evaluate land-use classification accuracy alongside reconstruction metrics.

Proposed Experiments

The decisive check is a single plot per dataset of fused reconstruction SNR against leased airtime at fixed primary rate. It shows the joint design, the separable design, the fixed-split naive scheme and the full-cube codec. The proposal succeeds if the joint curve sits measurably above the separable curve over a meaningful range of lease budgets, with lower spectral angle as well. Showing better spectral angle rules out the possibility that the gain comes from spatial sharpness alone.

A second check perturbs channel quality across the Rayleigh-fading simulator and tests whether the learned lookup policy tracks the per-realization optimum. A third injects bit errors and measures which transmission scheme degrades more gracefully.

A negative result on the first plot would itself be useful. It would show that the leasing literature's bit-count abstraction is adequate for imaging payloads, a claim no paper in the file has tested.