⚡ electrical engineering

Clear-Weighted Bit Allocation for Satellite Downlinks

This paper proposes a system that combines a clear-probability-weighted neural codec with a causal scheduler to efficiently transmit Earth-observation imagery from satellites, significantly reducing bandwidth usage and improving the delivery of clear-ground content compared to existing frame-discard and fixed-compression methods.

Original authors: Alireza Furutanpey, Qiyang Zhang, Yujie Huang, Philipp Raith, Schahram Dustdar

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Alireza Furutanpey, Qiyang Zhang, Yujie Huang, Philipp Raith, Schahram Dustdar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Space Data Bottleneck

Imagine you are a photographer standing on a mountain, taking thousands of high-definition photos of a bustling city below. You have a super-fast camera, but you only have a tiny, shaky walkie-talkie to send the photos home. The walkie-talkie works for just a few minutes every hour, and the battery is dying. If you try to send every single photo, the most important ones—the ones showing the clear, sunny streets—might get lost in the middle of a pile of blurry, cloudy shots, or worse, the walkie-talkie might run out of battery before the good pictures get through.

This is the daily struggle of Earth-observation satellites. They are like those photographers, circling our planet and snapping images of everything from forests to cities. But they face a "data bottleneck": they capture far more pictures than they can ever send back to Earth because their connection to the ground is intermittent and slow. To cope, old systems used a blunt tool: a "cloud detector." If the computer thought a picture had too many clouds, it would delete the entire photo before even trying to send it. The problem? The computer often made mistakes, tossing out perfectly clear, sunny photos just because it thought a tiny patch was cloudy. This paper tackles the messy, high-stakes game of deciding which pixels to save, how to compress them, and when to shout them down to Earth, all while the satellite is hurtling through space with limited power and time.

The Paper's Story: A Smarter Way to Pack and Send

The authors of this paper, a team of researchers from Europe and China, propose a new way to handle this cosmic data traffic jam. Instead of blindly deleting cloudy photos, they built a "smart codec" (a fancy word for a compression system) that learns to care more about the clear ground than the clouds.

The "Clear-Weighted" Trick
Think of the satellite's memory as a backpack. In the past, the satellite would look at a photo, guess if it was cloudy, and if it guessed wrong, it would throw the whole backpack away. The new method is different. During its training phase (like studying for a test), the system learns to look at every single photo, even the cloudy ones. It assigns a "weight" to every pixel: clear ground gets a heavy weight (very important!), while clouds get a light weight (less important). When the system compresses the image, it uses this weight to decide where to spend its limited "bytes" (the digital space used to store the image). It effectively says, "I will spend most of my memory on the clear streets and parks, and I'll skimp on the fluffy clouds."

The result is a system that doesn't need to send a separate map telling the ground station which parts are cloudy. The compression itself is already biased toward the good stuff. When they tested this, they found that to get the same quality of clear-ground images, this new method needed 47.8% fewer bytes than standard learned compression methods. That's like fitting almost twice as many high-quality photos into the same backpack.

The "Resume-able" Layers
The paper also introduces a clever way to package these photos. Instead of sending one giant file that must arrive all at once, the satellite splits every image into two layers: a "base layer" (a rough sketch of the whole picture) and a "refinement layer" (the fine details).

The Base Layer: This is the most important part. It can be decoded on its own to give a usable, albeit slightly fuzzy, image.
The Refinement Layer: This adds the sharp details.

This setup is like sending a postcard with a sketch first, and the full-color photo later. If the satellite's connection to Earth gets cut off mid-sentence (which happens often in space), the ground station still has the sketch. The next time the satellite passes over, it can pick up exactly where it left off and send the rest of the details. This "resumable" feature ensures that even if the connection is choppy, no data is wasted.

The "Deadline-Pressure" Scheduler
The final piece of the puzzle is the "traffic cop" on the satellite. The satellite has many photos waiting to be sent, each with different deadlines. Some photos need to be sent immediately (the base layer), while others can wait a bit longer (the refinement layer). The authors created a smart scheduling algorithm called DPMW (Deadline-Pressure Max Weight).

Imagine a busy airport control tower. The tower doesn't just let planes take off in the order they arrived. It looks at how much fuel they have left, how urgent their destination is, and how much space is left on the runway. Similarly, DPMW looks at:

How much "clear ground" is in the photo (the value).
How much of the photo is already sent (the unfinished work).
How close the photo is to its deadline.

If the connection to Earth is interrupted (like a sudden storm grounding flights), this smart scheduler rearranges the queue to make sure the most valuable clear-ground pixels get sent first. In their tests, when connections were interrupted, this smart scheduler managed to deliver 83.6% of the possible "perfect" clear-ground images, compared to only 38.1% with the old, rigid methods. It essentially doubled the amount of useful data that made it through the chaos.

The Cost of Being Smart
One might worry that being this smart requires a super-computer on the satellite, which uses too much power. However, the authors tested their system on real, resource-constrained satellite hardware (specifically the Jetson Orin Nano). They found that their optimized system was actually faster and used less energy than the old method of running a separate cloud detector and then deleting frames. The new system saved time and battery, proving that you don't need a massive computer to be efficient; you just need a smarter way of thinking.

In summary, this paper shows that by teaching satellites to prioritize clear ground during the compression process and by using a flexible, resume-able delivery system, we can send back significantly more useful information about our planet, even when the connection is spotty and the power is low. It turns a clumsy, all-or-nothing approach into a nimble, strategic game of data survival.

Technical Summary: Clear-Weighted Bit Allocation for Satellite Downlinks

Problem Statement

Earth-observation satellites capture significantly more imagery than their intermittent ground contacts can transmit. Current onboard systems typically employ a "detect-then-drop" pipeline: a cloud detector thresholds a frame, and if the cloud coverage exceeds a set limit, the entire frame or specific tiles are discarded before source coding. This approach suffers from two critical inefficiencies:

False Positives: Expert-labeled imagery analysis shows that these rules discard more than one-fifth of clear pixels, primarily due to detector false positives.
Suboptimal Bit Allocation: Even when frames are retained, conventional learned compression methods treat all pixels equally or optimize latent representations without considering the semantic value of clear versus cloudy regions. They do not reallocate source bits from cloudy areas to clear ground, nor do they dynamically manage downlink bytes across multiple captures with varying deadlines.

Furthermore, existing schedulers often use fixed two-stage service (serving base layers first, then refinements) that fails to jointly account for unfinished work at both base and full deadlines, leading to missed opportunities for delivering usable clear content during interrupted contacts.

Methodology

The proposed solution integrates a clear-weighted learned source codec with a causal deadline-pressure scheduler.

1. Clear-Weighted Progressive Coding

The authors propose a neural codec that reallocates bits toward clear ground without requiring a separate cloud detector pass during operation or transmitting a cloud map.

Training: The codec is trained using pixel-wise cloud probabilities ( $q_i$ ) derived from a frozen cloud detector. The reconstruction loss is weighted by $w_i = 1 - q_i$ , penalizing errors on likely clear ground more heavily than on cloudy regions.
Inference & Semantic Value: During operation, the codec estimates the clear-pixel count ( $v_i$ ) for each capture using a lightweight convolutional readout on the encoder's internal hyper-synthesis features. This provides a "codec-native semantic value" without a separate detection step.
Progressive Structure: Each capture is encoded into a resumable base layer (hyperlatent and anchor anchors) and a dependent refinement layer (non-anchor anchors). This allows transmission to be interrupted and resumed later.
Rate Allocation: Theoretically, this weighting shifts the ideal rate allocation, lowering distortion and increasing rate for coordinates with higher effective clear weights.

2. Deadline-Pressure Max Weight (DPMW) Scheduling

The system employs a causal scheduler that ranks arrived layers based on:

Codec-Native Value ( $v_i$ ): The estimated clear-pixel count.
Unfinished Work: The remaining bytes for base ( $b_i$ ) and refinement ( $e_i$ ) layers.
Deadline Slack: The time remaining until the base ( $d^B_i$ ) and full ( $d^F_i$ ) deadlines.
Aggregate Deadline Pressure: The system calculates aggregate pressure ( $P_B$ and $P_F$ ) representing the service rate required by incomplete bases and feasible full deliveries.

The scheduler computes a weight $W_i$ for each layer, prioritizing valuable items that require fewer residual bytes and are under higher deadline pressure. It supports packet resumption, storage bounds, and eviction policies that prioritize high-value items.

3. Onboard Optimization

The implementation is optimized for resource-constrained embedded accelerators (e.g., NVIDIA Jetson Orin Nano). Techniques include vectorized table indexing, CUDA Graph replay for neural computation, and caching of hyperlatent indexes to ensure the encoding service time remains within the capture interval deadlines.

Key Contributions

Clear-Weighted Learned Source Coding: A method using pixel-wise cloud probabilities during training to allocate source bits toward clear ground, reducing transmitted bytes while maintaining clear-region quality.
Progressive Codec-Native Interface: A network design that outputs a resumable base layer, a refinement layer, and a codec-native semantic value (clear-pixel count) derived from internal features, eliminating the need for a separate onboard detector.
Causal Two-Deadline Scheduling (DPMW): A scheduling policy that ranks layers by semantic value, unfinished work, and deadline slack, jointly optimizing for base and full deadlines without future capacity forecasts.

Results

Source Coding Efficiency

Bit Reduction: At matched clear-region reconstruction quality, the clear-weighted codec requires 47.8% fewer bytes than learned-compression baselines (ELIC) and 39.0% fewer bytes than a cloud-agnostic version of the same architecture.
Deletion Avoidance: Unlike frame-discard rules (CloudScout, Earth+) which delete ~21–30% of clear pixels due to false positives, the clear-weighted approach encodes every frame, achieving zero capture-time deletion of clear pixels.
Per-Image Effects: For mostly cloudy scenes, the method reduces byte usage by ~77% while maintaining quality for the small clear regions present.

Scheduling Performance

Delivery Improvement: In interrupted contact scenarios, the DPMW scheduler increases deadline-full delivery (complete reconstruction of clear content) from 38.1% (fixed two-stage service) to 83.5%.
Optimality: The scheduler achieves 83.6% of a certified clairvoyant upper bound (which has access to future arrivals and capacities).
Usable Delivery: It also exceeds reference orders (EDF, native value) in deadline-usable delivery (base layer completion), delivering 5.22 percentage points more clear content than fixed two-stage service during interruptions.

Onboard Performance

Latency & Energy: The optimized encoder service time on the Jetson Orin Nano is 19.89 ms (p99), well below the 33 ms target derived from capture intervals.
Efficiency: The optimized encoder consumes less time and energy than a single pass of the standalone cloud detector used in traditional frame-discard rules. The detector-to-encoder energy ratio ranges from 2.21 to 6.78, indicating the proposed method is computationally more efficient than the "detect-then-drop" alternative.

Significance and Claims

The paper claims to be the first satellite-imagery system to combine training-time clear-weighted source allocation with causal multi-contact service of resumable base and refinement bytes using a codec-native semantic value.

The significance lies in shifting the paradigm from discarding potentially useful data (frames/tiles) to intelligently reallocating limited downlink bandwidth toward the most valuable content (clear ground). By integrating the semantic value directly into the codec and the scheduler, the system avoids the overhead of separate detection passes and the inefficiencies of fixed scheduling policies. The authors emphasize that while they do not claim engineering optimizations (like entropy coding path tuning) as novel research contributions, they release these implementations to ensure fair comparisons. The system demonstrates that preserving clear-region quality is achievable with significantly lower bandwidth and computational cost than current state-of-the-art onboard processing methods.