Satellite edge computing for real-time and very-high resolution Earth observation

I. INTRODUCTION

• High-data-rate inter-satellite links distribute images to neighboring satellites, forming a cluster that processes and compresses data before ground download.Source: §I · p. 3
• Processing and communication at each satellite link must finish within a time slot, whose maximum capture interval is set by the ground track frame period.The capture frequency depends on the satellite's field of view and altitude.Source: §I · p. 3
• Early-discard approaches, can discard partially obstructed images that remain valuable for mission-critical applications such as emergencies.Source: §I · p. 4
• Prior satellite video-coding work used a single satellite and fixed downlink rate without optimizing energy efficiency, limiting its fit for high-frame-rate applications.Source: §I · p. 4
• The paper extends prior work to real-time, high-resolution Earth observation, with timing requirements defined by system dynamics.It frames this as a major extension of its earlier work and a departure from the SMEC literature discussed in the introduction.Source: §I · p. 5
• Its general SMEC framework segments imaging data, allocates processing tasks, scatters segments, processes them, and gathers results at the ground station.Source: §I · p. 5

• Distributing data across four satellites reduces per-link transmission and per-CPU processing loads, fulfilling the GTFP real-time constraint and stabilizing communication and processing subsystems.Source: §I · p. 5
• The paper defines a general model for high-resolution Earth-observation imagery with real-time constraints imposed by LEO orbital parameters and imaging instruments.Source: §I · p. 5
• It formulates a global optimization of segmentation, allocation, and processing to minimize energy while respecting satellite CPU-frequency and ISL and satellite-to-ground link-rate limits.Source: §I · pp. 5–6
• The framework optimizes which satellites participate, their CPU frequencies, and compression ratios, generalizing the previous work.Source: §I · p. 6
• Up to 12× more images reach the GS with SMEC than with direct download, and up to 2× more than with source-satellite processing.Source: §I · p. 6
• The SMEC approach uses up to 10× less energy than direct download; jointly optimizing task parameters for a whole task with K frames saves up to 90% energy versus optimizing each frame independently.Source: §I · p. 6
• In the La Palma scenario, optimizing parameters for K = 82 frames reduces energy consumption by up to 11% versus frame-by-frame optimization, and only distributed SMEC avoids buffer overflows.Direct download and local processing would overflow communication and/or processing buffers.Source: §I · p. 7

A. Framing

• The system comprises a ring of N uniformly spaced LEO satellites, with cameras and communication modules for ground links and neighboring-satellite ISLs.The satellites serve pre-planned tasks that scan an area of interest for delivery to the ground station.Source: A. Framing · p. 7
• The ground sample distance depends on image width in pixels, camera field of view, and satellite altitude, and determines the ground area covered by each image.The frame size is D_k = W_kD_img bits.Source: A. Framing · pp. 7, 9
• Camera pointing and rotation let the source satellite capture side-by-side images as it moves over the area of interest.Each frame may contain a different number of images; the stated example uses three images in the first two frames and two in the third.Source: A. Framing · pp. 7–8
• Frames are captured at the ground-track frame period so successive frames have no coverage gaps or overlapping pixels.The period is derived from the satellite orbital period at its deployment altitude.Source: A. Framing · p. 9
• Frames may be processed locally or distributively before transmission, reducing their size by a compression factor greater than one.Processing may include optical-anomaly correction, compression, and encoding.Source: A. Framing · p. 9

B. SMEC framework

• The framework accounts for LEO communication constraints, including limited bandwidth and links with heterogeneous capacity.Processed segments are sent to the ground station through a destination satellite.Source: B. SMEC framework · pp. 9–10
• Each frame is divided into segments, allocated among cluster satellites, distributed over ISLs, processed to reduce data size, and gathered at the ground station.A fraction of the frame may instead be designated for direct download without satellite processing.Source: B. SMEC framework · pp. 9–10

C. Compression

• Compression complexity is modeled as increasing exponentially with compression ratio, enabling a broader range of algorithms than the constant-complexity JPEG model.This model creates a trade-off between processing energy and communication energy.Source: C. Compression · p. 10
• The satellites adapt CPU frequency using DVFS, while FIFO processing queues are evaluated for stability rather than their effect on overall latency.The model also accounts for waiting until the CPU is available to process received data.Source: C. Compression · pp. 10–11

D. Communication

• The network is represented by a time-slotted graph whose links are updated from snapshots at the beginning of each slot.Intra-plane ISLs are fixed, full-duplex links with fixed data rate and power, while the satellite-to-ground connection is established with the closest available satellite.Source: D. Communication · p. 11
• The satellite-to-ground link uses RF, with its data rate selected from DVB-S2X rates whose SNR thresholds are met throughout the slot.The downlink rate adapts continuously as satellite movement changes the link.Source: D. Communication · p. 12

E. Energy consumption

• Processing energy is modeled from CPU frequency and the number of cores, with energy per clock cycle proportional to the square of clock frequency.The effective capacitance coefficient is denoted by ν.Source: E. Energy consumption · p. 13
• The communication energy model includes transmission energy, power-amplifier inefficiency, and static module power for RF and FSO links.The destination satellite's static RF power is excluded from the framework's energy evaluation.Source: E. Energy consumption · p. 13

III. PROBLEM FORMULATION

• The objective sums energy for scatter, processing, and gather, with scatter accounting for data distribution and gather for transmitting compressed segments toward the ground station.The scatter energy includes both reaching the ground station and distributing data among satellites.Source: §III · pp. 15–16
• The ground station jointly selects each frame's allocation, compression ratio, and satellite CPU frequencies to minimize energy while satisfying processing and communication constraints.Resource slicing requires average processing and transmission times to fit within the ground-track frame period.Source: §III · pp. 14, 16
• The routing algorithm affects ISL constraints and energy, but the formulation treats routing variables as parameters using hop-count shortest paths.Finding optimal routing for each phase is outside the paper's scope.Source: §III · p. 16
• The optimization is non-convex jointly in allocation, compression ratio, and CPU frequency, so the authors decompose it into three subproblems and solve them iteratively with block coordinate descent.The procedure first obtains a closed-form optimal CPU frequency, then uses it to optimize allocation and compression ratio iteratively.Source: §III · pp. 16–17

A. Optimizing the CPU frequency

• The frequency optimization considers CPU frequency f(n)_k for each frame k and satellite n, bounded by 0 ≤ f(n)_k ≤ f_CPU.It formulates the problem for a given ρ and X in P1.Source: A. Optimizing the CPU frequency · p. 17
• For feasible X and ρ, the optimal CPU frequency is constant and equals the minimum required to satisfy the processing constraint.The paper gives the optimal frequency for each satellite n given X and ρ.Source: A. Optimizing the CPU frequency · p. 17
• When the processing constraint cannot be fulfilled, CPU frequency is temporarily set to f_CPU during iterative optimization.Source: A. Optimizing the CPU frequency · p. 17

B. Optimizing the task allocation

• For fixed F* and ρ, task allocation X is optimized through an iterative convex approach because PX is linear in X with affine constraints.The processing constraint is moved into an augmented-Lagrangian objective as a penalty.Source: B. Optimizing the task allocation · p. 18
• The IPDD procedure solves for X at a given F, then updates F* after each iteration.Source: B. Optimizing the task allocation · p. 18

C. Optimizing the compression ratio

• For fixed X and F*, the compression ratio ρ is optimized with projected gradient descent.The formulation removes the constant scatter-energy term and includes processing constraints for each satellite.Source: C. Optimizing the compression ratio · p. 19
• The compression-ratio problem uses slack variables for downlink, ISL, and ρ constraints, while communication limits reflect average rates over task execution.Source: C. Optimizing the compression ratio · p. 19

D. Global optimization algorithm and practical considerations

• The iterative global optimization algorithm includes optimizing ρ given F* and X with projected gradient descent.Source: D. Global optimization algorithm and… · pp. 20–21
• IPDD and projected gradient descent convergence depend on the initial penalty terms and termination condition.Source: D. Global optimization algorithm and… · p. 20
• The methods can converge with finite penalties and iterations, but the iterations required vary widely with objective and constraint characteristics.Source: D. Global optimization algorithm and… · p. 21
• A real-system deadline can be met by limiting iterations, but this produces a sub-optimal solution; execution time can be characterized for the platform.Source: D. Global optimization algorithm and… · p. 21

IV. RESULTS

• Multi-frame optimization is optimal across the K frames; its gains are illustrated in synthetic and realistic scenarios representing upper and lower energy-savings bounds.Source: §IV · p. 22

A. Per-frame optimization

• Direct download has much higher energy and is feasible only for Wk ≤3 images; local processing has similar energy to the global solution within its Wk ≤18 feasible region.Source: A. Per-frame optimization · p. 22
• The global solution has minimum energy in all cases and supports Wk ≤37 for vd = v0 and Wk ≤36 for vd = v5.Source: A. Per-frame optimization · pp. 22–23, Figure 3 · p. 24
• 12× more images per frame are supported than with direct download, and 2× more than with local processing.Source: A. Per-frame optimization · p. 23, Figure 3 · p. 24
• Both local processing and global optimization select the same compression ratio; it decreases through Wk = 9 for vd = v0 and Wk = 14 for vd = v5.Source: A. Per-frame optimization · pp. 23–24, Figure 4 · p. 24

B. Multi-frame optimization

• With K = 5 frames and four empty frames, energy consumption is 90% lower than at K = 1 for vd = v0 and 56% lower for vd = v5.The task is modeled as an initial frame of W0 images followed by empty frames.Source: B. Multi-frame optimization · p. 26
• For La Palma’s K = 82 frames, global optimization saves 9% of energy at η = 0.1 and 11% at η = 1 versus per-frame optimization.The scenario is only feasible with distributed SMEC; direct download and local processing are infeasible.Source: B. Multi-frame optimization · p. 26
• Across the La Palma task, data is more evenly distributed across four satellites at η = 0.1 than at η = 1, when source satellite n = 5 processes more than 78% of the data.Optimal compression ratios vary with η and generally increase with the number of images in a frame.Source: B. Multi-frame optimization · p. 27

V. CONCLUSIONS

• Distributed SMEC captures, processes, and downloads up to 6× more images than direct download and 2× more than local SMEC.The conclusions also report energy savings up to 90% through selection of data allocation, compression ratio, and processing frequency.Source: §V · p. 27

APPENDIX

• For a satellite with data to process, the optimal CPU frequency is equal across all frames and remains below fCPU.The appendix derives this result through complementary-slackness conditions.Source: APPENDIX · p. 29
• When a satellite has no data to process, any CPU frequency can be selected.Source: APPENDIX · p. 29
• The complementary-slackness conditions can hold jointly only when the satellite has no data to process.The alternative would imply infinite processing time and is excluded.Source: APPENDIX · p. 29