Clear-Weighted Bit Allocation for Satellite Downlinks

Clear-Weighted Bit Allocation for Satellite Downlinks Alireza Furutanpey1,2,* , Qiyang Zhang3 , Yujie Huang4 , Philipp Raith1,2 , Schahram Dustdar2,5

arXiv:2608.01457v1 [cs.NI] 2 Aug 2026

1

Coovally, Barcelona, Spain 2 Technical University of Vienna, Austria 3 Peking University, China 4 Beijing University of Posts and Telecommunications, China 5 ICREA, Barcelona, Spain

Abstract—Earth-observation satellites capture more imagery than intermittent ground contacts can transmit. Onboard systems threshold a cloud detector, discard frames or tiles, and compress the survivors with a fixed codec. On expert-labeled imagery, these rules remove more than one-fifth of clear pixels, primarily through detector false positives. We train a neural codec with a clear-probability-weighted reconstruction loss, reallocating coded bytes from clouds to clear ground without requiring or transmitting a cloud map onboard. Each capture is encoded into a resumable base layer and a dependent refinement layer, while clear content is estimated from features produced by the encoder. At each contact, we causally rank arrived layers using estimated clear content, unfinished bytes, deadline slack, and aggregate deadline pressure. The scheduler serves base and computational deadlines, bounds stored residual bytes, and resumes interrupted packets. We evaluate the onboard-to-downlink pipeline using real entropy-coded bytes, orbit-derived interruptible contact capacities, and measured service time and energy on resource-constrained embedded accelerators. Clear-weighted codecs require up to 47.8% fewer bytes than learned-compression baselines at matched clear-region quality. The optimized encoder consumes less time and energy than one pass of the cloud detector used by the frame-discard rules. Relative to fixed two-stage service on the same streams, our scheduler more than doubles deadline-full clear-content delivery for the interrupted combined cohort, reaches 83.6% of a certified clairvoyant upper bound, and exceeds replayed reference orders in deadline-usable delivery. Index Terms—Earth observation, learned image compression, onboard computing, rate-distortion optimization, satellite downlinks

I. I NTRODUCTION Earth-observation satellites acquire imagery throughout an orbit but transmit only during intermittent ground contacts. With contact number and duration fixed, higher-resolution sensors produce more pixels per capture, and larger constellations produce more captures [1]–[3]. An onboard system must allocate source bits within each capture and downlink bytes among captures before their deadlines while respecting processing, storage, and downlink limits [4]–[6]. Prior onboard Earth-observation systems discard frames, select tiles, fill image regions, order payloads, or control receiver-side reconstruction from image content, cloud coverage, or task value [1]–[4], [6], [7]. Their authors compute these estimates outside the source codec and do not use them to allocate source bits between clear and cloudy regions. Studies of prior learned compression for satellite imagery optimize latent representations and entropy models while holding the * Corresponding author: a.furutanpey@dsg.tuwien.ac.at

acquired images and delivery schedule fixed, and therefore do not compare downlink byte allocations across captures [8]– [12]. Conventional onboard systems estimate cloud coverage, apply a threshold, and discard entire frames or tiles before source coding [3], [4], [13]. When an onboard system discards a frame or tile, it removes all clear-ground pixels from that frame or tile from encoding and the downlink queue. When the classification is wrong, the system removes clear ground that a correct decision would retain. With atomic payloads, no part of a capture can be reconstructed until every byte arrives. Even when payloads are layered, fixed two-stage service reserves contact capacity for base layers whose deadlines have arrived and admits refinement bytes only afterward. It does not jointly account for unfinished work at both deadlines. During training, we use pixelwise cloud probabilities from a frozen detector to assign larger reconstruction-error weights to likely clear ground. We confine that detector to training. During operation, we estimate each capture’s clear-pixel count using a convolutional readout of the already computed hypersynthesis features. We use this estimate as the codecnative semantic value for scheduling, replacing the separate operational detector, and produce a resumable base layer and a dependent refinement layer. At each contact, we compute a deadline-pressure max weight (DPMW) for every arrived layer from its capture’s codec-native value, unfinished item bytes, separate base and full deadline terms, and the current queue state. We serve arrived layer bytes in decreasing weight order. We assign earlier deadlines to base layers than to complete items, limit storage, and retain each partial packet’s byte offset so service can resume at a later contact. We isolate source allocation by holding the service rule and workload fixed while comparing clear-weighted and cloudagnostic encodings, and isolate service orders by comparing schedulers over the same clear-weighted byte streams and workload. At validation operating points matched for clearregion reconstruction quality, we measure 21% fewer transmitted bytes, including layer headers, with clear-weighted progressive encoding than with cloud-agnostic encoding from the same codec architecture. Clear-weighted encoding raises deadline-full delivery across orbit-derived contact capacities when both streams use fixed two-stage service. For the interrupted combined cohort across the trained codecs in the heldout contact workload, DPMW more than doubles deadlinefull delivery relative to fixed two-stage service. It reaches 83.6% of a certified clairvoyant upper bound and exceeds

the four replayed reference orders in deadline-usable delivery. For the primary Orin Nano Super configuration, we measure the source-coding service time from data preparation through entropy coding and find it below the computational deadline derived from reported capture intervals [4]. We provide the implementation and evaluation scripts in the accompanying repository1 . Our main contributions are: • We develop clear-weighted learned source coding by using pixelwise cloud probabilities during training to allocate source bits toward clear ground. • We design a progressive codec-network interface that represents each capture with a resumable base layer, a dependent refinement layer, and a codec-native semantic value estimating clear-pixel count. • We develop causal two-deadline scheduling that ranks arrived layers by codec-native value, unfinished work, and deadline slack. While we do not consider the engineering work to optimize entropy coding and execution paths a research contribution, we release these in a separate repository2 to reduce implementation confounds that otherwise favor systems-centric methods. To the best of our knowledge, no prior satellite-imagery system considers training-time clear-weighted source allocation with causal multi-contact service of resumable base and refinement bytes using a codec-native semantic value. II. BACKGROUND AND S YSTEM M ODEL Captured images arrive between intermittent ground contacts and are persisted in finite onboard storage. At each contact, only the current byte budget is available, and we allocate it causally among layers of arrived objects while unsent bytes remain queued. A usable base reconstruction has an earlier deadline than the full reconstruction, which requires both layers. We first assign bytes within each image through source coding, then choose which layer bytes to transmit so that bases and complete objects meet their deadlines. Each coded object contains a resumable base layer, a dependent refinement layer, and a codec-native semantic value that is fixed upon arrival. We measure frozen detector probabilities to weight reconstruction error during codec training and binary training-split cloud masks to supervise the native readout. Offline, we use validation reference masks for operating-point and resource selection and for validation-only comparisons, and test reference masks to construct held-out workloads and score test results. For detector-based baselines, we run the standalone detector. Let H and W be image height and width, xk ∈ [0, 1]H×W ×3 the RGB frame k, and x̂k its reconstruction. Detector output qki ∈ [0, 1] is the estimated cloud probability at pixel i. The evaluator-only indicators Cki ∈ {0, 1} and Gki ∈ {0, 1} mark reference-clear and reference-cloud pixels. For AllClear, we construct the reference regions from cloud-mask band 2 [14]. 1 https://github.com/rezafuru/Clear-Weighted-Bit-Allocation 2 https://github.com/rezafuru/RANSKit

detector false positives ref. 22%

70%

discard

78% of frame is clear and lost

est. 78%

frame-level granularity ref. 78%

whole frame discarded

discard

est. 84%

whole frame discarded 22% of frame is clear and lost

Fig. 1. Clear-content deletion caused by the CloudScout Rule.

We set Gki = 1 where the band equals one, Cki = 1 where it equals zero, and Cki = Gki = 0 elsewhere. We do not read the separate shadow bands. The unscored pixels remain in the HW denominator of the frame cloud fractions. For CloudSEN12+, we set Gki = 1 on the expert thick- and thincloud classes and Cki = 1 elsewhere. A. Cloud-Aware Source Decisions The conventional detect-then-drop pipeline thresholds clouds per frame and encodes only the survivors. For this pipeline, we compute the predicted and reference frame cloud fractions as 1 X 1 X 1{qki > 0.5}, ρk = Gki , (1) ρ̂k = HW i HW i where 1{·} is the indicator function. We discard frame k when ρ̂k > τ for a cloud-fraction threshold τ and compress the remaining frames. We instantiate τ at the published CloudScout 70% and Earth+ 50% frame thresholds, which we respectively call the CloudScout Rule and the Earth+ Rule [3], [13]. The pixel-removal adaptation of Wang et al. instead removes detector-predicted cloud pixels inside the codec and transmits the pixel mask as a side channel [15]. For clear-weighted encoding, we train on every frame with reconstruction error weighted by 1 − qki . P Let nC k = i Cki be the number of scored clear pixels in frame k. Capture-time deletion is P C nk 1{ρ̂k > τ } δ(τ ) = k P = δfp (τ ) + δsl (τ ), (2) C k nk where C k nk 1{ρ̂k > τ, ρk ≤ τ } P C , k nk P C n 1{ρ̂k > τ, ρk > τ } δsl (τ ) = k k P C . k nk

P

δfp (τ ) =

(3)

We count clear pixels from detector false positives in δfp and clear pixels inside correctly discarded cloudy frames, which we call clear slivers, in δsl (Fig. 1). The second term remains positive with perfect frame-level estimates because a frame above the cloud-fraction threshold can still contain clear ground. Clear-weighted encoding has zero capture-time deletion because we encode every frame. Reordering retained layer bytes after coding cannot recover a discarded frame.

B. Performance Measures For an emitted payload bk , we compute the coded rate as rk = 8|bk |/(HW ) bits per pixel (bpp). For learned codecs, we count every emitted image-latent and hyperlatent byte string, excluding dimensions, container metadata, and shared decoder state. For JPEG 2000 [16], we count the self-contained codestream after removing its comment marker, including 118 bytes of rate-independent marker segments per frame that the learned codec excludes. We use likelihood crossentropy as the training-rate proxy and the emitted bytes for every reported rate-distortion comparison. We compute the clear-region peak signal-to-noise ratio (PSNR) over the RGB reconstruction error on reference-clear pixels and omit frames without any scored clear pixels from per-frame statistics. Let K be a deadline-eligible cohort, AB (K) its items whose base layer meets its deadline, and AF (K) those whose base and refinement layers meet the full deadline. For the base and full deadlines θ ∈ {B, F }, reference-clear delivery is P C k∈A (K) nk (4) ηθ (K) = P θ C . k∈K nk We call ηB deadline-usable delivery and ηF deadline-full delivery. We measure the fused batch-1 service time S from tile preparation through host-to-device transfer, neural encoding, device-to-host transfer, and entropy coding, followed by one accelerator synchronization. We exclude queueing delay, and a batch size of 1 gives zero batch-fill delay. A satellite observes a new multi-tile image every 1–30 s [4]. Sustained singleprocessor encoding divides this interval among the tiles, our 256 × 256 frames, leaving milliseconds on wide sensors and seconds on narrow ones. We test the empirical 99th percentile of S against the 33 ms target inside this span. For a framediscard rule over K captured frames, we compute the survivor fraction as K 1 X s(τ ) = 1{ρ̂k ≤ τ }. (5) K k=1

We derive serial service demand from measured component medians, counting image tiling, host-to-device transfer, and detector time for every captured frame, and adding neural encoding, device-to-host transfer, and entropy encoding for the survivor fraction s(τ ). We measure energy from the total module rail over separate active and idle windows. Let P̄act and P̄idle denote mean active and idle power, Twin the active-window duration, and n the frames completed during that window. We compute the idlesubtracted energy per frame as

R

entropy model

Onboard

ẑi si hs

entropy coding native readout

 = R + λDw

weighted error Dw

x̂

queue

codec contract Bi biz biA

DPMW(t) WiB, WiE

biN

Ei

vi arrival ai

Ct

contact t sent xiℓt B

E

Q(t) ≤ Smax next byte offset

Fig. 2. Clear-weighted training and causal service of progressive layers.

A. Clear-Weighted Training Objective We omit frame index k in the per-frame formulation. We estimate cloud probability qi with a four-level U-Net of base width 64 from Sentinel-2 bands B8, B4, B3, and B2. We replace non-finite reflectance with zero, divide by 10,000, and clip to [0, 1]. For each clear-weighted codec seed, we freeze the detector checkpoint from that seed’s preceding AllClear detector-and-codec training run at distortion weight 300. Detector probabilities and RGB reconstructions use the same 256 × 256 crop without resampling. The clear weight is wi = 1 − qi .

(7)

We input RGB bands B4, B3, and B2 to the codec. For per-pixel RGB error ei = ∥xi − x̂i ∥22 /3, where xi ∈ [0, 1]3 is the color triple at pixel i, we use normalized clear-weighted distortion. Let R be the likelihood rate proxy, λ > 0 the distortion weight, and ϵ = 10−8 . P wi ei , L = R + λDw . (8) Dw = P i w i i+ϵ Let y and z denote the image latent and hyperlatent, with quantized forms ŷ and ẑ. Indices j and m enumerate their coordinates. We use pz,m for the modeled hyperlatent probability mass and py,j for the conditional image-latent probability mass. The rate proxy is X 1 R=− log2 pz,m (ẑm ) HW m  (9) X + log2 py,j (ŷj | ẑ, Cj ) , j

III. C LEAR -W EIGHTED P ROGRESSIVE C ODING

where Cj is the decoded checkerboard context. We train the entropy-bottleneck quantiles with a separate auxiliary loss. We use detector-derived wi only during codec optimization and encode each operational frame from RGB alone, transmitting no cloud map. For a generic binary clear indicator Zi , suppose the detector is calibrated so wi = Pr(Zi = 1 | V ) for observation V . If reconstruction error is conditionally independent of Zi given V , then E[wi ei ] = E[Zi ei ]. (10)

During training, we weigh reconstruction error with a frozen cloud detector. Onboard, we emit resumable base and refinement bytes, along with a codec-native semantic value, for each capture, and serve these layers across contacts in Section IV (Fig. 2).

If calibration and conditional independence hold, we estimate expected squared error on clear pixels from the numerator of Dw and clear area from its denominator. We implement the objective with the continuous score wi = 1 − qi , including pixels outside the scored clear and cloud regions.

E = max{0, P̄act − P̄idle }

Twin . n

(6)

B. Rate-Allocation Interpretation

D. Codec-Native Semantic Value

We model latent coordinate j as a separable Gaussian with variance νj , distortion dj , and ideal rate Rj (dj ) = [ 12 log2 (νj /dj )]+ , where [u]+ = max{u, 0}. Let w̃j denote its effective clear weight, averaged over the decoder receptive field and normalized to unit mean. The local allocation is X min [Rj (dj ) + λw̃j dj ] . (11)

We estimate each item’s clear-pixel count from the encoderside parameters si = hs (ẑi ) already computed while coding item i. With the codec fixed, we train a convolutional readout gψ against 16 × 16 average-pooled binary cloud masks from AllClear training images. The predicted clear-pixel value is ! 1 X σ(gψ (si )u ) , (14) vi = HW 1 − |Ωs |

{dj }j 0<dj ≤νj ∀j

j

Let rj⋆ = Rj (d⋆j ) denote the minimizing rate. The solution is   1 d⋆j = min νj , , 2 ln 2 λw̃j (12) +  1 log2 (2 ln 2 λνj w̃j ) . rj⋆ = 2 For w̃j = 0, we set d⋆j = νj and rj⋆ = 0. When both solutions have positive rate, changing the same coordinate from unit weight to w̃j changes its ideal rate by 12 log2 w̃j . Receptive fields and shared parameters couple coordinates after retraining, so we use the derivation only to predict that larger effective clear weight lowers distortion and raises rate where the positive-rate solution applies. C. Progressive Codec Interface We implement the analysis and synthesis transforms with Cheng-style residual blocks, strided convolutions, and subpixel upsampling [17]. The analysis transform ga maps x to y, and the hyper-analysis transform ha maps y to z. After quantizing z, the hyper-synthesis transform hs (ẑ) produces side parameters for the conditional distribution of ŷ. The synthesis transform gs reconstructs the image from ŷ. We code image latents with a two-pass checkerboard context [18]. Anchors use only the hyperprior side parameters. Non-anchors additionally use a masked convolution over decoded anchors and separate pointwise networks for Gaussian means and scales, adapting the improved checkerboard context of Fu et al. [19]. The encoder uses a 5 × 5 context over 12 opposite-parity anchors. We emit the hyperlatent, anchor, and non-anchor streams in dependency order. The base layer contains the hyperlatent and anchors. To decode it, we reconstruct the anchors, fill missing non-anchors with their conditional means, and apply gs . We decode the refinement layer’s non-anchor stream to recover the unchanged full reconstruction. We encode each image once, with no codec retraining and no second low-rate payload. In the network formulation, i indexes coded items rather N than pixels. Let bzi , bA i , and bi denote the three coded strings for item i, and let h = 16 bytes be the framing cost per layer. The network-facing sizes are Bi = |bzi | + |bA i | + h,

Ei = |bN i | + h.

(13)

We retain each queued layer’s next byte offset so interrupted service can resume in a later contact, and we admit refinement only after base completion.

u∈Ωs

where Ωs indexes the readout grid and σ is the logistic function. We implement the readout with a 1 × 1 convolution from 384 to 128 channels, a 3 × 3 convolution from 128 to 64 channels, and a 1 × 1 scalar output, with rectified linear units after the first two layers. The readout has 123k parameters and requires 31.5 M multiply-accumulate operations per tile. We infer vi from si once, without a separate onboard detector pass, fix it at arrival, and pass the contract (ai , Bi , Ei , vi ) to the network queue, where ai is item i’s arrival contact. E. Onboard Realization We optimize the onboard coded path. For each frame, we prepare one 256 × 256 tile, copy it to the accelerator, evaluate the analysis transform, hyperprior, checkerboard context, and entropy-parameter networks, copy symbols and probability indexes to the host, and entropy-code all streams. We exclude ground-side decoding from onboard service time. We replace per-symbol Gaussian-scale lookup with vectorized table indexing. Because the input size is fixed, we cache hyperlatent indexes and expanded medians, then replay the neural computation as a CUDA Graph after an explicit input copy [20]. We preserve the codec architecture, training loss, entropy coder, and bitstream syntax across the optimized and reference paths, yielding identical symbols, probability indices, and coded bytes in both paths. IV. D EADLINE -P RESSURE C ONTACT S CHEDULING With variable contact capacity, we must choose between new bases and refinements that complete older items. A fixed base allowance cannot respond to residual item work or aggregate deadline load. We rank arrived layers by semantic value, unfinished work, deadline slack, and the current queue’s deadline pressure, and call this rule DPMW. A. Contact and Queue Model Item i arrives at contact ai through the codec contract (ai , Bi , Ei , vi ). Its base and full deadlines are dB i = ai + 1,

dF i = ai + 3.

(15)

At contact t, we observe the current integer capacity Ct , and stored residual bytes cannot exceed Smax . Let xiℓt be bytes sent from layer ℓ ∈ {B, E} during contact t, where B denotes the base layer and E the refinement layer, let riℓ (t) be its

contact-start residual, and let Q(t) be stored residual bytes. Feasible service satisfies X xiℓt ≤ Ct , xiℓt ∈ {0, . . . , riℓ (t)}, (16) i,ℓ Q(t) ≤ Smax . Let b̄CA,s be the expected transmitted bytes per item for the full cloud-agnostic (CA) validation stream after weighting the data to 90% mean cloud cover, including both 16-byte layer headers. For codec seed s and A = 300 arrivals per contact, one contact’s capture volume is Vs = ⌊Ab̄CA,s ⌋, and we set Smax = 3Vs . We retain the byte offset after partial service across contacts and admit refinement only after prior or samecontact base completion. Let TiF denote the full-completion contact. The full-delivery utility is X UF = vi 1{TiF ≤ dF (17) i }. i

Meeting dB i is a separate usable-delivery objective, and we use one causal index policy for both. We compute operational weights from the native value vi and evaluate ηB and ηF with evaluator-only reference-clear pixels. B. Deadline Pressure At contact t, item i has base residual bi (t), refinement residual ei (t), and full-delivery feasibility fi (t) ∈ {0, 1}, which becomes zero after either layer is evicted. The remaining feasible full-delivery work is
RiF (t) = fi (t) bi (t) + ei (t) . (18) We compute bi (t), ei (t), fi (t), and RiF (t) only from arrived items that remain stored at contact t. At the start of each positive-capacity contact, we freeze the aggregate pressures PB (t) =

bi (t) , B d −t+1 i:b (t)>0 i X

(19)

i

t≤dB i

PF (t) =

X i:RiF (t)>0 t≤dF i

RiF (t) . F di − t + 1

(20)

We interpret PB as the service rate required by incomplete bases and PF as that required by items whose full delivery remains feasible.

For an eligible refinement, the weight is WiE (t) = 1{RiF (t) > 0, t ≤ dF i }

PF (t) vi . F F Ri (t) di − t + 1

(22)

Through vi /bi (t) and vi /RiF (t), we prioritize valuable items that require fewer residual bytes. The pressure-to-slack factors increase with queued work and approaching deadlines. D. Causal Service Rule Arrivals enter before service. If they push stored residual bytes above Smax , we evict refinement layers in increasing vi /ei (t), followed by base layers in increasing vi /bi (t). When we evict a base, we also remove its refinement. For positive capacity, we freeze PB (t) and PF (t) and insert every positive-weight incomplete base and eligible refinement into a max heap. We serve the highest-weight layer until that layer or the contact capacity is exhausted. When a base completes, we immediately insert its refinement with the frozen pressures and the updated residual work RiF (t) = ei (t). After exhausting all positive-weight work, we serve expired work by decreasing native value per residual byte. Eviction ties use earlier arrival and lower item identifier. Positive-weight ties use the earlier active deadline, then earlier arrival, base before refinement, and lower item identifier. Expired work ties omit the deadline key. For Lt active layers, pressure construction takes O(Lt ) time, and heap service takes O(Lt log Lt ) time and O(Lt ) memory. At each contact, we use current capacity, arrived item metadata, derived deadlines, residual bytes, and completion or eviction state. We use no future arrivals, interruption labels, evaluator-only clear pixels, fitted coefficients, capacity forecasts, or an offline solver. E. Clairvoyant Full-Delivery Bound We bound any scheduler on a fixed workload with a clairvoyant mixed-integer relaxation. Let K be the deadlineC eligible cohort, Si = Bi + Ei , Ti = {ai , . . . , dF i }, and ni be evaluator-only reference-clear pixels. For a P schedule π, define F its evaluator full-delivery utility UFC (π) = i∈K nC i 1{Ti ≤ F di }. Binary ζi marks full completion, continuous ϕit denotes aggregate item bytes assigned to contact t, and ξit = Si ζi − Pt−1 ′ ϕ t′ =ai it is selected residual work at contact start. We solve X (UFC )UB = max nC i ζi ζ,ϕ

s.t.

i∈K

X

ϕit = Si ζi

∀i,

t∈Ti

X

ϕit ≤ Ct

∀t,

ξit ≤ Smax

∀t,

(23)

i:t∈Ti

C. Base and Refinement Weights For an incomplete base, the service weight is WiB (t) = 1{t ≤ dB i }

PB (t) vi B bi (t) di − t + 1

vi PF (t) + 1{RiF (t) > 0, t ≤ dF . i } F Ri (t) dF i −t+1 (21)

X i:t∈Ti

ζi ∈ {0, 1},

ϕit ≥ 0.

We give the model future arrivals and capacities, allow immediate eviction of unselected arrivals, omit unscored traffic, and drop the earlier base deadline. For any feasible system schedule π, retain only bytes of cohort items completed by

V. E VALUATION A. Experimental Design We fit models on AllClear training data. We use validation data to select operating points and network resources and to report the progressive-byte and 501-frame deletion comparisons. We use the region-disjoint test split for rate-distortion, per-image, capacity-response, and scheduler results. Each split contains 256 × 256 Sentinel-2 crops and machine-generated cloud masks [14]. We test unchanged same-sensor transfer on expert-labeled CloudSEN12+ [21]. We train four codec seeds at five distortion weights and use three shared rate points for transfer and scoring sensitivity. We count emitted range asymmetric numeral system (rANS) bytes from 32-bit floatingpoint encoding of every image and hyperlatent stream and compute Bjøntegaard delta rate (BD-rate) at matched clearregion PSNR with 95% Student-t intervals over codec seeds. We retrain ELIC [22] and Minnen 2018 [23] from scratch on the same training data with four seeds and five distortion weights per codec spanning a comparable rate range with one shared early-stopping schedule. We measure encoder and standalone-detector execution on a Jetson Orin Nano Super and Jetson AGX Orin. For the primary Nano configuration, we use 16-bit floating point, batch size one, 15 W mode, and maximum clocks without TensorRT. We compute linearly interpolated p99 from 100 synchronized calls after 30 warm-ups and subtract separately measured idle module-rail power. Before test evaluation, we use validation data to select one clear-weighted (CW) point and one qualitymatched CA point. We encode each capture into base and refinement layers with 16-byte headers while retaining crosscontact state. For the orbit-profile experiment, we draw 300 arrivals per contact from 4,052 test frames resampled to 90% mean cloud cover. We use one-contact base and three-contact full deadlines, packet resumption, five backlog orders per codec seed, and seven capacity fractions. We propagate a 24-hour Sentinel-2B orbit over a Svalbard ground station with a 5◦ elevation mask and set the 15 relative contact capacities to the contacts’ usable byte volumes. Each backlog order is one seeded arrival draw. At fraction f , we distribute ⌊15f Vs ⌋ bytes across contacts according to the orbit profile. We hold arrivals, integer capacities, storage limits, and backlog order fixed between paired CA and CW cells. We repeat the sweep with native, shared-detector, and uniform values. With fixed twostage service, we maximize complete due-base native value before admitting refinement while retaining at least 99% of

that optimum. We average capacity curves within codec seed and apply one-sided Student-t inference to the normalized capacity-range integral. For the held-out scheduler experiment, we reuse the four selected CW streams over 30 contacts. We evaluate five backlog orders, two capacity fractions, and three interruption events in 240 paired interrupted and matchedcontrol cells. For each event, we set one to three consecutive contact capacities to zero and either redistribute the removed bytes over the next six contacts or remove them permanently. We give the matched controls the same total bytes on the uninterrupted profile. Only byte-service order changes across the three schedulers. We score the combined cohort, the items arriving from three contacts before the first interrupted contact through six after the last one. Earlier arrivals reach their full deadlines before the interruption, and the six-contact tail spans the redistribution window. For deadline-full delivery ηF , we credit reference-clear pixels in items whose base and refinement meet the three-contact deadline. For deadlineusable delivery ηB , we use the same denominator and credit a completed base at one contact. We average balanced cells within codec seed and compute one-sided 95% lower bounds and interruption-specific difference-in-differences (DiD) conditional on the four seeds. In post hoc replays, we apply four online orders with identical state-transition rules: earliest deadline first (EDF) with native value per residual byte as tiebreak, native value, native value per residual full-delivery work, and that ratio divided by deadline slack. We solve (23) with HiGHS at relative gap 10−6 . B. Source Representation and Allocation 1) Capture-Time Deletion: We apply one fixed standalone detector to the CloudScout 70% and Earth+ 50% frame rules [3], [13] on a 501-frame trace and sweep cloud cover through weighted resampling with five 4,000-frame backlogs. We separate deletion caused by frame-level false positives from deletion of clear slivers, the clear regions inside correctly discarded cloudy frames, in Fig. 3. We measure 5.36% clear-pixel deletion with the CloudScout Rule on the fixed trace. Frame-level false positives account for 3.08 points and clear slivers for 2.28 points. With the Earth+ Rule, we measure 9.76% deletion. Our pixel-removal adaptation of Wang et al. deletes 7.36% [15]. At 90% mean cover, deletion reaches 21.4 ± 1.7% for the CloudScout Rule and 30.4 ± 1.8% for the Earth+ Rule. Captured clear pixels deleted (%)

dF i and set their ζi = 1. Removing all other traffic cannot violate capacity or storage, so the retained service is feasible in (23) with the same evaluator utility. Thus UFC (π) ≤ UFC,∗ ≤ (UFC )UB , and UFC (π)/(UFC )UB is a certified lower bound on its instance-specific ratio to the unknown clairvoyant optimum UFC,∗ . Aggregate service for a selected item has a basefirst realization by labeling its first Bi cumulative bytes as base. This certificate gives no worst-case approximation or stochastic stability guarantee.

False positive Clear sliver

(a) 12 9

(b) 30

CloudScout Rule Earth+ Rule

20

6

10

3 0

0

24 ted ule ule igh out R ng 20 th+ R we a r c a r S W a E d Cle Clou

Clear-weighted

0 0.1

0.3

0.5

0.7

0.9

Global cloud cover

Fig. 3. Clear-pixel deletion from frame- and pixel-level cloud removal.

Clear-weighted

45

Cloud-agnostic

ELIC

Clear-region PSNR (dB)

Minnen 2018

JPEG 2000

Scene

Cloud-agnostic Clear-weighted

Scene

Cloud-agnostic Clear-weighted

Cloud-region PSNR (dB)

40 35 30 0.01

0.02

0.05

0.1

0.01

0.02

0.05

0.1

Real coded bits per pixel

Low codelength share

High codelength share

Fig. 4. Clear- and cloud-region rate-distortion on AllClear.

At distortion weight 300, we allocate 5.09 ± 0.51 times as many image-latent bits per clear pixel as per cloudy pixel with clear-weighted coding. For cloud-agnostic coding, the ratio is 0.52 ± 0.02. We compute conditional image-latent codelength from the exact rANS cumulative-distribution tables, exclude hyperlatents and stream flush, and normalize each map to within-image codelength shares. TABLE I C APTURE - TIME DELETION AND FROZEN TRANSFER TO C LOUD SEN12+. Metric Deletion, total % (false-positive points) CloudScout Rule Earth+ Rule Clear-weighted vs. CA BD-rate (%)

AllClear

CloudSEN12+

5.36 (3.08) 9.76 (4.73) −37.6 ± 4.6

21.19 (16.95) 28.74 (17.80) −26.4 ± 5.0

TABLE II R EAL - BYTE BD- RATE AT MATCHED CLEAR - REGION PSNR OVER FIVE RATE POINTS . Method Clear-weighted Cloud-agnostic

Reference Cloud-agnostic ELIC Minnen 2018 JPEG 2000 ELIC

−39.0 ± −47.8 ± −42.0 ± −62.2 ± −13.7 ±

BD-rate (%) 7.4 7.4 6.4 4.7 4.1

Byte change (%)

We transfer the fixed detector and both codecs to CloudSEN12+ without retraining (Table I). Both frame rules delete more than one-fifth of the captured clear pixels on CloudSEN12+, primarily due to false positives. 2) Matched Clear-Region Quality: Clear-weighted coding reduces real bytes by 39.0% at matched clear-region quality against the same architecture trained without clear weighting (Fig. 4). Relative to Efficient Learned Image Compression (ELIC), we measure 47.8% fewer bytes, with fewer bytes for every trained seed than for either reference. Cloud-agnostic training with our internal architecture reduces bytes by 13.7% relative to ELIC. Minnen 2018 is the stronger learned reference at the evaluated rate points (Table II). At 0.0164 bpp, clear-region PSNR is 1.5 dB higher and cloud-region PSNR is 6.7 dB lower with clear weighting. Against the same cloud-agnostic codec over the three shared rate points, we measure a 37.6% byte reduction on AllClear and 26.4% after unchanged transfer of both codecs to CloudSEN12+ (Table I). 3) Conditional Codelength: We compare normalized image-latent codelength for four partly cloudy scenes in Fig. 5.

Clear-region PSNR change (dB)

Fig. 5. Normalized image-latent codelength for four partly cloudy scenes. Mostly clear 2 Distortion weight 300

Mixed Mostly cloudy Distortion weight 3000

0 −2 −4 −100

100 50 0 −50 −100

0

−50

0

50

100 −100

−50

Byte change (%)

0.33

0.66

Ground-truth cloud fraction

0

50

100

1

Fig. 6. Per-image real-byte and clear-region PSNR changes across cloud cover at distortion weights 300 and 3000.

4) Per-Image Effects: We report byte and clear-region PSNR changes across reference cloud fraction in Fig. 6. We define mostly clear and mostly cloudy groups using the lower and upper thirds of reference cloud fraction. At distortion weight 300, mostly clear scenes use 18.2% more bytes at the median and gain 0.25 dB median clear-region PSNR. Mostly cloudy scenes use 77.7% fewer bytes, with a 0.86 dB median reduction among scenes containing scored clear pixels. At weight 3000, mostly clear scenes use 11.5% more bytes while median clear-region PSNR falls by 0.17 dB. 5) Scoring Sensitivity: We compare BD-rate with ELIC using four clear-region definitions in Table III. For every definition, we measure at least 43% fewer bytes with clear-weighted coding than with ELIC at the three shared rate points, while the cloud-agnostic reduction remains near 16%. When we exclude clear pixels within one and two pixels of a cloud boundary, the clear-weighted reduction increases. With the population scored by OmniCloudMask, we still measure fewer bytes with clear-weighted coding than with cloud-agnostic coding [24]. C. Deadline-Constrained Delivery 1) Progressive Capacity Response: At the selected operating points, CW transmits 21% fewer bytes than CA, including layer headers, on the full 501-frame validation split without cloud-cover resampling. With the same fixed two-stage service and codec-native values, we measure a 20.25-point higher normalized capacity-

Model Clear-weighted Cloud-agnostic

AllClear −50.2 ± 6.4 −16.0 ± 4.5

+1 px −51.5 ± 6.7 −15.9 ± 4.3

+2 px −52.7 ± 6.6 −16.0 ± 4.3

Omni −43.7 ± 9.6 −16.2 ± 5.0

DPMW

Full delivery (%)

TABLE III R EAL - BYTE BD- RATE (%) AGAINST ELIC AT THE THREE SHARED RATE POINTS .

Fixed two-stage

100 75 50 25 0

0.25

(b) Interrupted

0.35

0.25

CW

CA

Native

Shared

0.35

Capacity fraction

Uniform

50 25 0 .20

.30

.40

.50

.75

1.00

1.25

Capacity / CA capture Fig. 7. Deadline-full delivery across contact capacities for fixed two-stage service.

range integral for CW than for quality-matched CA, with a 17.41-point lower bound (Fig. 7). With the shared-detector score, we measure a 20.93-point increase. When we replace uniform with native values, the CW integral increases by 3.16 points. We reach 95% mean deadline-full delivery at about 47% lower capacity with CW than with CA for both native and shared scores. The validation-selected CW delivery point lies above the plotted rate range of Fig. 4. Across the four CW test streams, scored without cloud-cover resampling, we measure a base-layer fraction of 66.0% of mean full wire bytes and base clear-region PSNR of 44.63 dB, 1.88 dB below full reconstruction. 2) Deadline-Pressure Scheduling: For deadline interleaving, we serve due bases, due refinements, future bases, and future refinements in fixed phases and order each phase by deadline and native value per residual byte. DPMW raises mean deadline-full delivery from 38.1% to 83.5% versus fixed two-stage service and by 74.77 points versus deadline interleaving (Fig. 8, Table IV). The DPMW–fixed gain is 45.41 points, with a 35.95-point lower bound. We compute a post hoc paired cluster bootstrap over codec seeds and backlog orders, stratified by event and capacity cell, and obtain a two-sided 95% interval of 37.31– 50.21 points. All three schedulers transmit 99.72% of available capacity on average across interrupted profiles, so service volume does not explain the delivery difference. Deadline interleaving keeps a base in the due phase after its deadline passes and serves due bases oldest first, so expired bases precede just-due ones and due refinements wait behind the whole due-base backlog. With DPMW across the six post-interruption contacts, we TABLE IV C OMBINED - COHORT DPMW CONTRASTS .

Contrast Mean (points) 95% lower bound (points) Interrupted full, vs. fixed +45.41 +35.95 Interrupted full, vs. interleave +74.77 +65.72 Interruption-specific full DiD +9.66 +8.43 Interrupted usable, vs. fixed +5.22 +2.88

(a) DPMW service mix 100

Refinement

50 0

Base 1

2

3

4

5

6

Clear delivery (%)

Fig. 8. Combined-cohort deadline-full delivery for DPMW and both online baselines.

75

Mean cell share (%)

Full delivery (%)

Capacity fraction

100

Deadline interleave

(a) Matched control

Contact after interruption

(b) Separate deadline metrics 100 DPMW Fixed two-stage Deadline interleave

50 0

Usable 1 contact

Full 3 contacts

Fig. 9. DPMW service mix and combined-cohort delivery at two deadlines.

allocate 25–33% of transmitted bytes to refinement while continuing base service (Fig. 9). Mean full and usable delivery remain closer with DPMW than with either online baseline. DPMW raises mean full delivery by 11.30 points and mean usable delivery by 6.28 points relative to EDF on interrupted profiles (Table V). The two density policies exceed DPMW’s full delivery by 2.13–2.21 points but fall 2.44–3.21 points below its usable delivery. We found no improvement from the aggregate-pressure multipliers in post hoc ablations. DPMW reaches 83.60% of the certified bound and closes 73.49% of the fixed-to-bound full-delivery gap. The certified bound completes about one-third of cohort items at the lower capacity fraction, and the uncompleted items hold under 1% of cohort clear pixels. HiGHS reports optimal status for all 240 solves. During interruptions, deadline-usable delivery is 5.22 points higher with DPMW than with fixed two-stage service, with a 2.88-point lower bound. The full-delivery improvement is 9.66 points larger for interrupted profiles than for matched controls, with an 8.43-point lower bound. In matched controls, full delivery is higher with DPMW, while one-contact usable delivery is 1.99 points lower. D. Onboard Cost Optimizing the coding path reduces fused 99th-percentile service time from 47.03 to 19.89 ms and idle-subtracted energy from 148.25 to 103.17 mJ per frame on the primary Nano configuration (Table VI). Optimized component medians remain below 33 ms on both devices. On 4,052 held-out tiles per codec seed, the native readout has a mean Spearman correlation of 0.852 and a mean absolute error of 6,820 pixels, or 10.41% of a tile. The separate detector reaches 0.887 and 6,131 pixels, or 9.35%. In separate active windows, detector-to-encoder energy ratios for the full encoder are 2.21–2.62 across device and power configurations. On the Nano, detection takes 36.9 ms versus 19.6 ms for full-frame encoding, and detect-then-drop remains slower than

Per-frame median demand (ms)

TABLE V P OST HOC REFERENCES ON INTERRUPTED PROFILES . Policy or bound ηB (%) ηF (%) Clairvoyant upper bound – 99.89 Density / deadline slack 84.04 85.72 Value / residual full work 83.27 85.64 Native value 80.74 83.73 DPMW 86.47 83.51 EDF 80.20 72.21

50 40 30 20 10 0.0

0.2

Nano 15 W

Energy (mJ) 148.25 103.17 Energy (mJ) 148 → 103 36 → 18

encoding every frame across both device curves (Fig. 10). For the energy-only size comparison, the full, half, and quarter encoders contain 15.9, 7.7, and 4.2 million onboard parameters. Across twelve maximum-clock device, power, and encoder-size configurations, detector-to-encoder energy ratios span 2.21–6.78 (Fig. 11). VI. R ELATED W ORK Onboard Earth-observation systems reduce downlink demand through frame or tile rejection, cloud filling, adaptive coding and reconstruction, and re-imaging schedules [3], [4], [6], [7], [13], [15]. Zhang et al. select and deduplicate tiles for object counting in TargetFuse, while Liu et al. remove spatial and temporal redundancy from space-situational-awareness imagery through distributed on-orbit sparse coding [2], [25]. These systems either make content decisions outside the codec or exploit generic image redundancy. We use dense cloud probabilities only during training to allocate rate within every RGB capture. During operation, we encode without a separate onboard detector pass and transmit no cloud map. We expose resumable base and refinement bytes to the scheduler. Prior studies optimize satellite codecs for global, metadataconditioned, temporal, or task-specific reconstruction [8]–[12]. Other authors allocate rate through importance maps, training masks, segmentation priors, or global distortion [22], [23], [26]–[28]. SPIFF, AdaStreamer, AdaSem, and OMNIS adapt imagery or features to semantic relevance, bandwidth, channel feedback, or radio resources [29]–[32]. These systems infer relevance during operation or bind communication to a fixed task. We retain general image reconstruction and export exact layer sizes with a codec-native value. Liu et al. assign tasks to local, sunlit, or ground resources in PHOENIX to meet completion deadlines while minimizing maximum battery depth of discharge [33]. Park et al. coordinate satellite handoff and video rate from throughput predictions to maximize streaming quality [34]. Neither group allocates contact bytes among incomplete base and refinement

Detector energy / encode energy

Full

Margin (ms) 13.4 27.0

0.6

0.8

1.0

AGX max

Clear-weighted

Detect-then-drop

Fig. 10. Per-frame service demand across survivor fractions on two devices.

TABLE VI O NBOARD ENCODER SERVICE TIME AND ENERGY. Fused full encoder, Orin Nano 15 W, max clocks Execution p99 (ms) 33-ms margin Reference 47.03 −14.03 Optimized 19.89 +13.11 Component medians, reference→optimized Device Mode Encoder Service time (ms) Nano 15 W full 46.0 → 19.6 AGX max full 23.6 → 6.0

0.4

Survivor fraction

Half

Quarter

Detector = encode

6 4 2 0 Nano 15 W

Nano max

AGX 50 W

AGX max

Fig. 11. Detector energy divided by optimized encoder energy at maximum clocks.

layers from multiple captures. We rank residual base and refinement bytes with DPMW while limiting storage, enforcing distinct deadlines, resuming packets, and using no futurecapacity forecast. Our earlier work, FOOL [5], pairs feature compression via shallow variational bottleneck injection [35] with accounting for target-device throughput, transfer cost, energy, and mission capacity. There, we deliberately set clouds aside as a separate challenge and preserved downstream task performance without prior knowledge of the tasks. In contrast, this work addresses the more general objective of preserving clear-region reconstruction quality, saving the bits that cloud regions would otherwise consume as high-entropy content. We additionally consider capture-time deletion, progressive coded bytes, and deadline delivery. VII. C ONCLUSION Clear-weighted training reallocates source bits toward clear content and emits resumable base and refinement layers with codec-native values. DPMW ranks arrived layers by value, residual work, and two deadlines while limiting storage, resuming packets, and using no future-capacity forecast. The selected progressive stream uses 21% fewer bytes than the cloud-agnostic stream from the same codec architecture on the validation split. In held-out interrupted contacts, DPMW raises deadline-full delivery from 38.1% to 83.5%, with a 35.95point lower bound. It reaches 83.6% of the certified clairvoyant upper bound and exceeds the four replayed reference orders in deadline-usable delivery. The optimized encoder stays below the latency target on the primary Nano configuration. Future work will bridge this system with FOOL’s task-agnostic feature compression and extend the clear-weighted objective and the scheduler to preserve performance on downstream tasks without prior knowledge of their labels.

R EFERENCES [1] B. Denby and B. Lucia, “Orbital edge computing: Nanosatellite constellations as a new class of computer system,” in Proc. 25th ACM Int. Conf. Archit. Support Program. Lang. Oper. Syst. (ASPLOS), 2020, pp. 939–954. [2] Q. Zhang, X. Yuan, R. Xing, Y. Zhang, Z. Zheng, X. Ma, M. Xu, S. Dustdar, and S. Wang, “Resource-efficient in-orbit detection of earth objects,” in IEEE INFOCOM 2024 - IEEE Conference on Computer Communications, Vancouver, BC, Canada, May 20-23, 2024. IEEE, 2024, pp. 551–560. [3] K. Du, Y. Cheng, P. Olsen, S. Noghabi, and J. Jiang, “Earth+: On-board satellite imagery compression leveraging historical earth observations,” in Proc. 30th ACM Int. Conf. Archit. Support Program. Lang. Oper. Syst. (ASPLOS), 2025, pp. 361–376. [4] B. Denby, K. Chintalapudi, R. Chandra, B. Lucia, and S. Noghabi, “Kodan: Addressing the computational bottleneck in space,” in Proc. 28th ACM Int. Conf. Archit. Support Program. Lang. Oper. Syst. (ASPLOS), 2023, pp. 392–403. [5] A. Furutanpey, Q. Zhang, P. Raith, T. Pfandzelter, S. Wang, and S. Dustdar, “FOOL: Addressing the downlink bottleneck in satellite computing with neural feature compression,” IEEE Trans. Mobile Comput., vol. 24, no. 8, pp. 6747–6764, 2025. [6] C. Sun, Y. Zhang, B. Tao, D. Vasisht, and M. Marina, “DeepSpace: Super resolution powered efficient and reliable satellite image data acquisition,” in Proc. ACM SIGCOMM, 2025, pp. 311–328. [7] J. Liu, Q. Li, Y. Li, C. Ji, and S. Wang, “SCo-Cloud: Satellite constellation collaboration for cloud-aware onboard-computed imaging and transmission,” in Proc. AAAI Conf. Artif. Intell., vol. 40, no. 9, 2026, pp. 7159–7167. [8] Z. Zhang, H. Qiu, M. Zhang, J. Liu, B. Chen, T. Zhang, and H. Li, “COSMIC: Compress satellite image efficiently via diffusion compensation,” in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 37, 2024, pp. 91 951–91 982. [9] G. Guerrisi, F. Del Frate, and G. Schiavon, “Artificial intelligence based on-board image compression for the Phi-Sat-2 mission,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 16, pp. 8063–8075, 2023. [10] J. Costa-Watanabe, I. Wittmann, B. Blumenstiel, and K. Schindler, “TerraCodec: Compressing optical earth observation data,” arXiv preprint arXiv:2510.12670v3, 2026. [11] D. Liu, J. Zhang, X. Liang, Y. Qi, Y. Song, Y. Xi, and J. Jin, “RSLLIC: A lightweight learned image compression model with knowledge distillation for onboard remote sensing,” IEEE Trans. Geosci. Remote Sens., vol. 64, pp. 1–13, 2026. [12] C. Mollière, I. Cumplido, M. Zeulner, L. Liesenhoff, M. Schubert, and J. Gottfriedsen, “Learned image compression for earth observation: Implications for downstream segmentation tasks,” arXiv preprint arXiv:2512.01788, 2025. [13] G. Giuffrida, L. Diana, F. de Gioia, G. Benelli, G. Meoni, M. Donati, and L. Fanucci, “CloudScout: A deep neural network for on-board cloud detection on hyperspectral images,” Remote Sens., vol. 12, no. 14, p. 2205, 2020. [14] H. Zhou, C.-H. Kao, C. P. Phoo, U. Mall, B. Hariharan, and K. Bala, “AllClear: A comprehensive dataset and benchmark for cloud removal in satellite imagery,” in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 37, 2024, pp. 53 571–53 597. [15] K. Wang, J. Jia, P. Zhou, H. Ma, L. Yang, K. Liu, and Y. Li, “Efficient on-board compression for arbitrary-shaped cloud-covered remote sensing images via adaptive filling and controllable quantization,” Remote Sens., vol. 16, no. 18, p. 3431, 2024. [16] A. Skodras, C. Christopoulos, and T. Ebrahimi, “The JPEG 2000 still image compression standard,” IEEE Signal Process. Mag., vol. 18, no. 5, pp. 36–58, 2001. [17] Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, “Learned image compression with discretized gaussian mixture likelihoods and attention modules,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 7936–7945. [18] D. He, Y. Zheng, B. Sun, Y. Wang, and H. Qin, “Checkerboard context model for efficient learned image compression,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 14 766–14 775.

[19] H. Fu, F. Liang, J. Liang, Y. Wang, Z. Fang, G. Zhang, and J. Han, “Fast and high-performance learned image compression with improved checkerboard context model, deformable residual module, and knowledge distillation,” IEEE Trans. Image Process., vol. 33, pp. 4702–4715, 2024. [20] A. Furutanpey, C. Walser, P. Raith, P. A. Frangoudis, and S. Dustdar, “Leveraging neural graph compilers in machine learning research for edge-cloud systems,” IEEE Trans. Parallel Distrib. Syst., vol. 37, no. 9, pp. 2065–2077, 2026. [21] C. Aybar et al., “CloudSEN12+: The largest dataset of expert-labeled pixels for cloud and cloud shadow detection in Sentinel-2,” Data Brief, vol. 56, p. 110852, 2024. [22] D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y. Wang, “ELIC: Efficient learned image compression with unevenly grouped spacechannel contextual adaptive coding,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 5708–5717. [23] D. Minnen, J. Ballé, and G. D. Toderici, “Joint autoregressive and hierarchical priors for learned image compression,” in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 31, 2018, pp. 10 771–10 780. [24] N. Wright, J. M. A. Duncan, J. N. Callow, S. E. Thompson, and R. J. George, “Training sensor-agnostic deep learning models for remote sensing: Achieving state-of-the-art cloud and cloud shadow identification with OmniCloudMask,” Remote Sens. Environ., vol. 322, p. 114694, 2025. [25] Y. Liu, H. Jin, W. Yao, Y. Chen, Y. Zhao, L. Kong, R. Li, X. Liu, and G. Chen, “Distributed on-orbit sparse coding for efficient space situational awareness image transmission,” in IEEE INFOCOM 2025 - IEEE Conference on Computer Communications, London, United Kingdom, May 19-22, 2025. IEEE, 2025, pp. 1–10. [26] M. Li, W. Zuo, S. Gu, J. You, and D. Zhang, “Learning content-weighted deep image compression,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 43, no. 10, pp. 3446–3461, 2021. [27] M. Li, S. Gao, Y. Feng, Y. Shi, and J. Wang, “Content-oriented learned image compression,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 632–647. [28] Y. Liu, W. Yang, H. Bai, Y. Wei, and Y. Zhao, “Region-adaptive transform with segmentation prior for image compression,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2024, pp. 181–197. [29] M. Palena, J. A. Ayala-Romero, A. Garcia-Saavedra, and C. Chiasserini, “SPIFF: selective preservation of image fidelity for bandwidthconstrained heterogeneous networks,” in IEEE INFOCOM 2026 - IEEE Conference on Computer Communications, Tokyo, Japan, May 18-21, 2026. IEEE, 2026, pp. 1–10. [30] A. Zhu, S. Zhang, K. Cheng, X. Shi, Z. Qian, and S. Lu, “AdaStreamer: Machine-centric high-accuracy multi-video analytics with adaptive neural codecs,” in IEEE INFOCOM 2024 - IEEE Conference on Computer Communications, Vancouver, BC, Canada, May 20-23, 2024. IEEE, 2024, pp. 1161–1170. [31] Q. Liao and T. Tung, “AdaSem: Adaptive goal-oriented semantic communications for end-to-end camera relocalization,” in IEEE INFOCOM 2024 - IEEE Conference on Computer Communications, Vancouver, BC, Canada, May 20-23, 2024. IEEE, 2024, pp. 1111–1120. [32] L. Qin, I. Harshbarger, L. Nasraoui, C. Chiasserini, and M. Levorato, “OMNIS: semantic RAN slicing via dynamic split neural networks,” in IEEE INFOCOM 2026 - IEEE Conference on Computer Communications, Tokyo, Japan, May 18-21, 2026. IEEE, 2026, pp. 1–10. [33] W. Liu, Z. Lai, Q. Wu, H. Li, Q. Zhang, Z. Li, Y. Li, and J. Liu, “In-orbit processing or not? sunlight-aware task scheduling for energy-efficient space edge computing networks,” in IEEE INFOCOM 2024 - IEEE Conference on Computer Communications, Vancouver, BC, Canada, May 20-23, 2024. IEEE, 2024, pp. 881–890. [34] K. Park, Z. He, C. Luo, Y. Xu, L. Qiu, C. Ge, and M. Muaz, “Joint optimization of handoff and video rate in LEO satellite networks,” in IEEE INFOCOM 2026 - IEEE Conference on Computer Communications, Tokyo, Japan, May 18-21, 2026. IEEE, 2026, pp. 1–10. [35] A. Furutanpey, P. Raith, and S. Dustdar, “FrankenSplit: Efficient neural feature compression with shallow variational bottleneck injection for mobile edge computing,” IEEE Trans. Mobile Comput., vol. 23, no. 12, pp. 10 770–10 786, 2024.

Related documents

• Invitation to the Second Symposium on Problems in the Optimization of Data Communications Systems#131380
• Taxonomy of Coding Techniques for Efficient Network Communications#139258
• Information-Theoretic Foundations of Mismatched Decoding#1051929
• North-Bound Distribution of Link-State and Traffic Engineering (TE) Information Using BGP#209013
• Transparent Interconnection of Lots of Links (TRILL): Appointed Forwarders#262438