arXiv:2604.13361v1 [cs.NI] 14 Apr 2026
Joint Semantic Coding and Routing for Multi-Hop Semantic Transmission in LEO Satellite Networks Hong Zeng
Jiangtao Luo*
Yongyi Ran*
School of Communications and Information Engineering Chongqing University of Posts and Telecommunications Chongqing, China [email protected]
School of Communications and Information Engineering Chongqing University of Posts and Telecommunications Chongqing, China [email protected]
School of Communications and Information Engineering Chongqing University of Posts and Telecommunications Chongqing, China [email protected]
Abstract—Low Earth Orbit satellite networks pose significant challenges to multi-hop semantic transmission because rapidly changing topology, link variability, and queue dynamics make end-to-end performance jointly depend on routing, relay processing, and semantic payload adaptation. Existing studies usually optimize routing or semantic transmission separately and are therefore not well suited to dynamic satellite scenarios under local observations. To address this issue, this paper proposes GraphJSCR, a graph-based joint routing and semantic coding method for multi-hop semantic transmission in dynamic Low Earth Orbit satellite networks. The satellite constellation is modeled as a time-varying directed graph, and the forwarding process is formulated as a partially observable sequential decision problem. A graph representation learning module is designed to encode local topology, link status, queue conditions, packet context, and semantic transmission states. Based on the learned representation, the proposed decision network jointly determines next-hop selection, relay processing level, and semantic transmission budget to balance end-to-end semantic quality and transmission delay. The semantic encoder-decoder is developed with reference to the SwinJSCC framework. Simulation results demonstrate that GraphJSCR achieves faster convergence and a better tradeoff between semantic fidelity and transmission efficiency than benchmark methods. Index Terms—LEO satellite networks, multi-hop transmission, joint routing and semantic coding, graph attention network, reinforcement learning
I. I NTRODUCTION Low Earth Orbit (LEO) satellite networks are expected to play a key role in future integrated space-air-ground communication systems because of their wide-area coverage, flexible deployment, and low dependence on terrestrial infrastructure [1]. However, their high orbital mobility, rapidly time-varying topology, and frequent inter-satellite link switching make reliable multi-hop transmission highly challenging. Meanwhile, emerging satellite services increasingly involve image- and video-oriented traffic, for which conventional bit-level reliable delivery is often inefficient in terms of transmission overhead, This work is supported by the National Natural Science Foundation of China(No. U25B2033, No. U23A20275 and No. 62171072 ). * Jiangtao Luo and Yongyi Ran are the Corresponding authors (Email: [email protected], [email protected]).
latency, and task effectiveness [2]. This motivates semanticoriented transmission, where the objective is to preserve taskrelevant information rather than exact bit recovery. Although recent advances in intelligent routing for LEO networks [3] and semantic communication based on Deep JSCC [2] and its Swin Transformer variant [4] provide useful foundations, they do not directly resolve the core challenge of multi-hop semantic transmission in dynamic LEO networks. In a multi-hop semantic system, end-to-end performance is jointly affected by path selection, relay behavior, and semantic transmission budget adaptation. Forward-only relaying can cause representation mismatch and accumulation of distortions across heterogeneous links, whereas decode-and-reencode relaying can improve robustness at the cost of additional delay and resource consumption onboard [5]. Therefore, the essential problem is not routing alone or coding alone, but how to coordinate forwarding and semantic processing over multiple hops under dynamic topology and local observations. Existing studies have started to uncover the coupling in multi-hop semantic transmission. Early efforts mainly focused on characterizing semantic degradation, whereas more recent works have moved toward active relay design. For example, [6] introduced residual compensation to reduce distortion accumulation in multi-hop image transmission. On the relay side, [7] leveraged semantic-state maintenance for efficient forwarding, [8] enhanced robustness via predictive feature extraction, and [9] extended semantic relaying to multi-source scenarios through relay-side fusion of critical semantic information. Cross-layer design has also been investigated in recent years, [10] revealed the coupling between semantic encoding and routing constraints, [11] proposed a task-oriented semantic delivery framework based on local topological information for heterogeneous satellite networks, and [12] employed hybrid DeepJSCC with adaptive compression-and-forward operations to combat performance degradation over unstable links. While these studies provide valuable insights, the problem of developing a unified distributed framework for dynamic LEO satellite networks remains relatively underexplored, especially under local observations where forwarding decisions need to jointly consider routing and semantic coding.
To address this issue, this paper proposes GraphJSCR, a graph-based joint routing and semantic coding algorithm for multi-hop semantic transmission in dynamic LEO satellite networks. The main contributions can be summarized as follows: First, we formulate multi-hop semantic transmission in LEO satellite networks as a unified joint decision problem integrating routing, relay processing, and semantic coding control under local observations. Second, we design a graph-representation-based distributed decision framework that exploits local topological and transmission states for adaptive hop-by-hop optimization. Third, we build an evaluation mechanism and demonstrate the effectiveness of the proposed method in jointly improving semantic quality and transmission efficiency. II. P ROBLEM F ORMULATION A. System Architecture As illustrated in Fig. 1, we consider multi-hop semantic image transmission over a dynamic LEO satellite network. At time slot t, the constellation is modeled as a time-varying directed graph G(t) = V, E(t) , (1) where V denotes the set of satellites and E(t) denotes the set of available inter-satellite links (ISLs). Let N = |V| be the number of satellites. A semantic transmission task is initiated at a source satellite s ∈ V and terminated at a destination satellite d ∈ V. At the source node, the input image is first transformed into semantic features by the semantic encoder and then organized into packetized semantic payloads for multi-hop transmission over the dynamic LEO network. After reception at the destination, the semantic payload is restored and fed into the semantic decoder to reconstruct the semantic content. In our work, the semantic encoder-decoder is developed based on the SwinJSCC [4] framework, namely SwinJSCC w/ SA&RA. Compared with conventional bit-level forwarding, the considered system additionally allows relay-side semantic processing at relay satellites. Therefore, the end-to-end transmission performance is jointly determined by the network path, relay processing behavior, and semantic payload budget along the multi-hop route. B. Relay Processing and Semantic Budget Adaptation Let zk denote the encoded semantic latent representation of task k. Before transmission, zk is packetized into semantic payload units and injected into the LEO network. For each relay satellite, the outgoing semantic payload can either be directly forwarded or further processed before transmission to the next hop. We use mi to denote the relay processing mode/level at relay node i, and use ci to denote the semantic transmission budget allocated to the outgoing hop. The transmission budget can be interpreted as the effective semantic channel budget (e.g., Channel-C in the implementation), which controls the
amount of semantic information carried by the outgoing payload. Different relay levels correspond to different processing intensities on the semantic representation. When relay processing is activated, the incoming semantic payload is processed through a relay pipeline including bitstream parsing, latent feature recovery, importance-aware channel pruning, latent re-quantization, and bitstream repacking. Denote the relay operator at node i by z̃k,i = Rmi ,ci (zk,i ) ,
(2)
where zk,i and z̃k,i are the incoming and outgoing semantic representations, respectively. The operator Rmi ,ci (·) adapts the semantic payload to downstream link and congestion conditions by jointly controlling relay-side processing and semantic budget allocation. C. Queue Model Each satellite maintains one receiving queue and multiple sending queues associated with its available outgoing directions. Newly arrived packets are first cached in the receiving queue and then dispatched to the corresponding sending queue according to the forwarding decision. Let qi,p (t) denote the length of the sending queue associated with port p on satellite i at time slot t. The queue evolution is modeled as n o + qi,p (t + 1) = min qi,p (t) − oi,p (t) + zi,p (t), qmax , (3) where zi,p (t) and oi,p (t) denote the number of packets entering and leaving queue (i, p) during time slot t, respectively, [·]+ = max(·, 0), and qmax is the maximum queue capacity. D. Delay Model For a packet transmitted from satellite i to satellite j at time slot t, the propagation delay is given by prop τi,j (t) = di,j (t)/c0 ,
(4)
where di,j (t) is the Euclidean distance between the two satellites and c0 is the speed of light. Let Ri,j (t) denote the achievable transmission rate of link (i, j) at time slot t, and let bout k,i (t) denote the outgoing payload size of task k after possible relay-side semantic processing and budget adaptation at node i. Then, the transmission delay is tx τk,i,j (t) = bout k,i (t)/Ri,j (t),
(5)
The queuing delay of task k at the selected sending queue q (i, p) is denoted by τk,i,p (t), which depends on the instantaneous queue occupancy and service condition of the corresponding outgoing port. In addition, if relay-side semantic processing is enabled at node i, an extra processing delay proc τk,i (t) is incurred. Therefore, the per-hop delay of task k on hop (i, j) can be written as prop q proc tx τk,i,j (t) = τi,j (t) + τk,i,j (t) + τk,i,p (t) + τk,i (t),
(6)
Let Pk denote the multi-hop path traversed by task k. Then the cumulative end-to-end transmission delay is X Dk = τk,i,j (t), (7) (i,j)∈Pk
Semantic Decoder
Dequantization
Reassembly
Deserialization
Depacketization
Packetization
Serialization
Blocking
Quantization
Semantic Encoder
GAT-based Representation Learning Sub-graph Inputs
• • •
Node Features Local Topology packet / semantic states
GAT Encoder
• •
Multi-head attention Neighbor message passing
Neighbor Aggregation
• •
Weighted feature fusion Representative node update
Node Embedding
• •
Latent state encoding Input to policy actor
Relaylevel
0
128
0
Node 68
1
64
0
Node 05
2
96
1
...
...
...
...
Node xx
x
xx
x
Queue/Load Status
Packet State
Current/Destination Info
PPO-Actor
Port Queue Status
Semantic State Measured SNR
Estimated Quality Proxy
Forward to Next Hop
Channel -C
Node 23
Importance-aware Channel Pruning
Nexthop
Link Status
Latent Feature Recovery
Local Topology
Destinati on
Node Features
Bitstream Parsing
Action Instruction Table
Network State
Bitstream Repacking
GAT
States Collection
Latent Re-quantization
Relay Processing Pipeline
Fig. 1. Overall framework of the proposed GraphJSCR method.
E. Problem Formulation Let Ik and Iˆk denote the source image and the reconstructed image of semantic task k, respectively. To characterize endto-end semantic delivery performance, we define a normalized semantic quality score Qsem = N Φ(Ik , Iˆk ) , 0 ≤ Qsem ≤ 1, (8) k k where Φ(·, ·) denotes the selected semantic quality evaluator and N (·) denotes a normalization mapping to [0, 1]. Let K denote the set of semantic transmission tasks. The considered problem aims to jointly optimize routing and semantic processing so as to reduce cumulative transmission delay while improving end-to-end semantic quality. We formulate the objective as 1 X min EΠ λd Dk + λs (1 − Qsem k ) Π |K| k∈K
s.t.
C1 : (i, j) ∈ E(t),
∀(i, j) ∈ Pk ,
C2 : 0 ≤ qi,p (t) ≤ qmax ,
∀i, p, t,
C3 : mi ∈ M,
∀i,
ci ∈ C,
C4 : TTLk (t) ≥ 0,
(9)
∀k, t,
C5 : λd ≥ 0, λs ≥ 0, λd + λs = 1. where Π denotes the overall forwarding-and-processing policy, λd ≥ 0 and λs ≥ 0 are weighting coefficients, and λd + λs = 1. M and C denote the feasible relay-mode set and semanticbudget set, respectively. III. T HE P ROPOSED G RAPH JSCR M ETHOD A. GAT-based State Representation Learning As shown in Fig. 1, each packet-holding satellite can only access local observations and one-hop neighboring information, which makes direct hop-by-hop decision making susceptible to local optima. To enhance local perception under partial
observability, we construct a relay-aware subgraph centered at the current forwarding node and employ a graph attention network (GAT) to aggregate neighboring information. Specifically, for the current node i, let Ni (t) denote its one-hop neighboring set at time slot t, and let xj (t) denote the input feature vector of node j ∈ Ni (t) ∪ {i}. The subgraph input jointly includes network states, packet states, and semantic states. The attention coefficient from node j to node i is computed as eij (t) = LeakyReLU a⊤ [Wxi (t) ∥ Wxj (t)] , (10) exp(eij (t)) , u∈Ni (t)∪{i} exp(eiu (t))
αij (t) = P
(11)
where W is the learnable projection matrix, a is the attention vector, ∥ denotes concatenation. The hidden representation of node i is then obtained by weighted neighborhood aggregation: X hi (t) = σ αij (t)Wxj (t) , (12) j∈Ni (t)∪{i}
where σ(·) is a nonlinear activation function. The resulting embedding hi (t) provides a compact representation that implicitly captures local topology variation, queue/load propagation trends, and semantic transmission context. This GAT-enhanced representation is then fed into the downstream decision network for hop-by-hop joint optimization. B. GraphJSCR Decision Network To model the hop-by-hop joint decision process under local observations, we formulate multi-hop semantic forwarding in the dynamic LEO network as a partially observable Markov decision process (POMDP).
1) State Space: At time slot t, the current packet-holding node i receives a local observation pkt sem oi (t) = onet (t) , i (t), oi (t), oi
(13)
where onet i (t) contains network-related states such as node features, local topology, link status, and queue/load status; opkt i (t) contains packet-related states such as current/destination infor(t) contains semanticmation and port queue status; and osem i related states such as measured SNR and semantic quality proxy. After one-hop graph aggregation, the input to the policy network is written as si (t) = oi (t), hi (t) ,
(14)
Therefore, the decision state jointly reflects local network dynamics, packet context, and semantic transmission conditions. 2) Action Space: Instead of flattening all forwarding choices into a single label, GraphJSCR adopts a factorized joint action design: relay c ai (t) = ahop (t), a (t), a (t) , (15) i i i c where ahop i (t) denotes next-hop selection; ai (t) denotes the semantic channel budget, chosen from C = {64, 96, 128}, where each value determines the semantic payload capacity; arelay (t) denotes the relay processing level, where 0 means i direct forwarding and 1 means relay processing before nexthop transmission. The action can be interpreted as an instruction tuple that jointly determines where the packet should be forwarded, how much semantic transmission budget should be allocated, and whether relay-side semantic processing should be activated. 3) Reward Function: The reward is designed to jointly encourage efficient forwarding and high semantic fidelity. Let rt denote the immediate reward at time slot t. It is defined as
rt = rttr + βsem rtsem ,
(16)
where rttr is the transmission-related reward and rtsem is the semantic-quality reward, βsem is a weighting coefficient. The transmission-related reward is further written as rttr = rtprog + rtsucc − rtfail ,
(17)
where rtprog is a shaping term for forwarding progress, rtprog = ωh ∆hop − ωd pdelay − ωq pqueue − ωl ploop ,where ∆hop denotes t t t t t the reduction of the residual distance to the destination after taking the current action, and pdelay , pqueue , and ploop denote t t t the penalties associated with delay growth, queue congestion, and loop risk, respectively. rtsucc is the terminal reward for successful delivery, and rtfail is the penalty for failed forwarding events such as TTL expiration, queue overflow, or unavailable outgoing links. rtsem is activated when a semantic session is successfully completed.
Algorithm 1 Training Procedure of the Proposed GraphJSCR 1: Initialize actor parameters θ, critic parameters ϕ, and rollout buffer D 2: for each training episode do 3: Reset the dynamic LEO environment and initialize semantic tasks 4: for each environment step do 5: Observe local state st of the packet-holding node 6: Construct the relay-aware subgraph and obtain GAT embedding ht c relay 7: Generate joint action at = (ahop ) from t , at , at policy πθ 8: Execute at , and obtain reward rt and next state st+1 9: Store (st , at , rt , st+1 ) and related information in D 10: if the rollout horizon is reached or the episode terminates then 11: Compute return and advantage targets from D 12: for each PPO update epoch do 13: Sample mini-batches from D 14: Update actor θ using the clipped surrogate objective 15: Update critic ϕ using the value loss 16: end for 17: Clear rollout buffer D 18: end if 19: end for 20: end for 21: Output the trained shared GraphJSCR policy
C. Training Strategy of GraphJSCR GraphJSCR is trained end-to-end in the dynamic LEO simulation environment through repeated environment interaction and policy optimization. At each step, the current node observes its local state, performs a forward pass through the actor network, executes the selected joint action, and receives the resulting reward. The collected rollout samples are then used to update the actor and critic. To stabilize policy learning, we adopt Proximal Policy Optimization (PPO) with the clipped surrogate objective h i Lclip = Et min rt (θ)Ât , clip(rt (θ), 1 − ϵ, 1 + ϵ)Ât , (18) where πθ (at | st ) rt (θ) = , (19) πθold (at | st ) Ât is the advantage estimate, and ϵ is the clipping coefficient. The full training objective further includes a value regression term and an entropy regularization term: L = −Lclip + cv Lvalue − ce H πθ , (20) where Lvalue is the critic loss, H(πθ ) denotes the policy entropy, and cv and ce are weighting coefficients. During training, actor and critic parameters are updated for multiple epochs using minibatch rollout samples. After
convergence, the learned GraphJSCR policy can be deployed in a fully distributed manner, where each packet-holding satellite reuses the shared policy to adaptively coordinate next-hop selection, semantic budget allocation, and relay-side processing under dynamic topology and local observations. IV. E VALUATION A. Experiment Setup All experiments are conducted in ns-3.41, where the simulator is connected to the learning agent through the ns3-ai Gym interface. We consider a pure ISL scenario with a Walker-like constellation consisting of 10 orbital planes and 7 satellites per plane at an orbital altitude of 570 km. The maximum queue length and packet time-to-live are 600 packets and 16 hops, respectively. A random channel perturbation with a standard deviation of 1 dB is introduced. In addition, a slow timevarying jitter with an amplitude of 2.0 dB is imposed, the channel correlation horizon is set to 2.0 s, and the link failure rate is 0.05. For semantic transmission, the payload chunk size is 1200 bytes and the semantic frame interval is 6s. The default evaluation uses a single-session regime, with each flow generating at most one semantic session per episode. Semantic performance is evaluated on selected images from the DIV2K dataset. We use SSIM and a CLIP-based image-image cosine similarity metric [13] to measure structural fidelity and semantic consistency, respectively. We also report session drop rate as an auxiliary reliability metric for failures caused by congestion, TTL expiration, or unavailable forwarding. For policy training, we adopt PPO with a learning rate of 5 × 10−5 , a discount factor of 0.99, a rollout horizon of 256, a clip ratio of 0.2, 4 update epochs, and a mini-batch size of 128. The entropy and value loss coefficients are set to 0.05 and 0.5, respectively. The weighting coefficient βsem is 1. B. Baseline Algorithms We consider two groups of baselines, corresponding to backend comparison and routing-oriented comparison. For backend comparison, we evaluate the proposed GraphJSCR against classical semantic transmission baselines, including JPEG2000+LDPC and DeepJSCC. This group is used to compare complete end-to-end semantic transmission schemes under the same dynamic LEO environment and task setting, rather than to isolate the coding module alone. In addition, to assess the contribution of relay-aware semantic processing, we compare the full GraphJSCR scheme with a reduced variant using plain SwinJSCC without relay strategy. For routing-oriented comparison, the semantic frontend is fixed to SwinJSCC. On this basis, we compare different decision mechanisms, including the proposed GraphJSCR, GraphPR [3], and DQN-IR [14], in order to isolate the gain brought by joint routing, relay processing, and semantic budget adaptation. For ablation, we test two reduced GraphJSCR variants—without source-C control and without relay process-
Fig. 2. Convergence Analysis.
(a)
(b)
Fig. 3. Semantic quality comparison of different schemes under varying SNR conditions. (a) CLIP score. (b) SSIM score.
ing—to isolate the gains of semantic budget adaptation and relay-aware processing under heavier load. C. Results and Analysis Convergence Analysis: Fig. 2 compares the training convergence behaviors of DQN-IR, GraphPR, and GraphJSCR. It can be observed that GraphJSCR converges significantly faster to a higher reward level and exhibits smaller oscillations than DQN-IR and GraphPR during training. This improvement stems from our proposed joint optimization, which helps the policy better balance transmission efficiency and semantic fidelity in dynamic multi-hop LEO networks. Semantic Quality under Different SNRs: Fig. 3(a) and Fig. 3(b) compare the semantic consistency and structural fidelity of different schemes under varying SNRs. GraphJSCR achieves the best or near-best CLIP score across the whole SNR range, with a particularly clear advantage in the lowand medium-SNR regimes. In terms of SSIM, GraphJSCR also shows superior performance at low SNRs and significantly outperforms other methods in this regime. Although JPEG2000+LDPC performs best at high SNRs, GraphJSCR still maintains competitive quality while avoiding unnecessary semantic payload overhead. This indicates that GraphJSCR prioritizes semantic quality preservation under poor channel conditions, while under favorable channels it tends to reduce transmission load without sacrificing reconstruction quality. Considering both SSIM and CLIP, GraphJSCR provides a more robust and balanced semantic transmission solution across different channel conditions. Load Sensitivity under Increasing Session Load: Fig. 4(a) and Fig. 4(b) show the average delivery latency and session
bustness for semantic image transmission in challenging lowSNR multi-hop scenarios. V. C ONCLUSION
(a)
(b)
Fig. 4. Load sensitivity comparison under different semantic session loads. (a) Average end-to-end delivery latency. (b) Session drop rate.
This paper proposed GraphJSCR, a graph-based joint routing and semantic coding framework for multi-hop semantic transmission in dynamic LEO satellite networks. By modeling the forwarding process as a partially observable sequential decision problem, the proposed method jointly optimizes nexthop selection, relay processing, and semantic transmission budget based on local topology, queue, packet, and semantic states. Simulation results showed that GraphJSCR achieves faster convergence, better semantic fidelity under varying SNR conditions, and stronger robustness under increasing session loads than benchmark methods. R EFERENCES
Fig. 5. Qualitative reconstruction comparison at SNR = 3 dB. GraphJSCR employs adaptive source-channel selection over candidate channel budgets {64, 96, 128}, and the policy selects C = 128 at this SNR. DeepJSCC uses latent channel number c = 8 with an inference patch size of 128 × 128. JPEG2000+LDPC uses compression rate r = 16 with LDPC parameters (n = 128, dv = 2, dc = 4) and a maximum of 40 decoding iterations.
drop rate under different semantic session loads, where the load level is controlled by varying the number of concurrent flows while keeping the semantic frame interval and packetization setting fixed. As the offered load increases, all methods suffer from higher queueing delay and reduced delivery reliability. Nevertheless, GraphJSCR-full consistently maintains the lowest overall latency and the most favorable robustness–latency tradeoff in terms of session success. The ablation results further show that removing adaptive sourceC control leads to the most noticeable degradation, while removing relay processing mainly hurts robustness under high load. Qualitative Reconstruction Comparison: Fig. 5 presents a visual comparison at SNR = 3 dB. Under this low-SNR condition, both DeepJSCC and JPEG2000+LDPC suffer from evident quality degradation. DeepJSCC exhibits texture blurring and loss of fine structural details, whereas JPEG2000+LDPC shows artifact contamination and local structural distortion. By contrast, GraphJSCR preserves the overall scene semantics much more faithfully, achieving the highest SSIM and CLIP scores among the compared schemes. Notably, this quality advantage is obtained with only 931 transmitted packets, which is comparable to JPEG2000+LDPC (908 packets) and far smaller than DeepJSCC (4588 packets). This visual evidence demonstrates that the proposed method provides stronger ro-
[1] E. Lagunas, S. Chatzinotas, and B. Ottersten, “Low-earth orbit satellite constellations for global communication network connectivity,” Nature Reviews Electrical Engineering, vol. 1, no. 10, pp. 656–665, 2024. [2] E. Bourtsoulatze, D. B. Kurka, and D. Gündüz, “Deep joint sourcechannel coding for wireless image transmission,” IEEE Transactions on Cognitive Communications and Networking, vol. 5, no. 3, pp. 567–579, 2019. [3] Y. Ran, Y. Ding, S. Chen, J. Lei, and J. Luo, “Fully-distributed dynamic packet routing for LEO satellite networks: A GNN-enhanced multiagent reinforcement learning approach,” IEEE Transactions on Vehicular Technology, vol. 74, no. 3, pp. 5229–5234, Mar. 2025. [4] K. Yang, S. Wang, J. Dai et al., “SwinJSCC: Taming swin transformer for deep joint source-channel coding,” IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 1, pp. 90–104, 2024. [5] C. Bian, Y. Shao, H. Wu, E. Ozfatura, and D. G”und”uz, “Processand-forward: Deep joint source-channel coding over cooperative relay networks,” IEEE Journal on Selected Areas in Communications, vol. 43, no. 4, pp. 1118–1134, Apr. 2025. [6] B. Xie, J. Park, Y. Wu, W. Zhang, and T. Quek, “Multi-hop parallel image semantic communication for distortion accumulation mitigation,” arXiv preprint arXiv:2510.26844, 2025. [7] E. Arda, E. Kutay, and A. Yener, “Semantic forwarding for next generation relay networks,” in 2024 58th Annual Conference on Information Sciences and Systems (CISS), 2024, pp. 1–6. [8] W. Lin, Y. Yan, L. Li et al., “Semantic-forward relaying: A novel framework toward 6G cooperative communications,” IEEE Communications Letters, vol. 28, no. 3, pp. 518–522, 2024. [9] J. Chen, S. Yang, T. T. Chan et al., “SemFusion: Multi-source semantic information fusion and communication,” in 2024 International Wireless Communications and Mobile Computing (IWCMC). IEEE, 2024, pp. 1740–1745. [10] B. Guo, Z. Xiong, B. Wang et al., “Semantic communication-aware end-to-end routing in large-scale LEO satellite networks,” in 2024 IEEE International Conference on Metaverse Computing, Networking, and Applications (MetaCom). IEEE, 2024, pp. 137–142. [11] R. Gao, B. Zhang, Q. Zhang, and Z. Yang, “Task-oriented semantic delivery in large-scale heterogeneous satellite networks: A localtopological-information-dependable deep learning approach,” IEEE Internet of Things Journal, vol. 12, no. 21, pp. 45 895–45 908, Nov. 2025. [12] C. Bian, Y. Shao, and D. G”und”uz, “A deep joint source-channel coding scheme for hybrid mobile multi-hop networks,” IEEE Journal on Selected Areas in Communications, vol. 43, no. 7, pp. 2543–2559, Jul. 2025. [13] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in Proceedings of the 38th International Conference on Machine Learning, vol. 139, Jul. 2021, pp. 8748–8763. [14] P. Zuo, C. Wang, Z. Yao, S. Hou, and H. Jiang, “An intelligent routing algorithm for LEO satellites based on deep reinforcement learning,” in Proc. IEEE 94th Vehicular Technology Conference (VTC2021-Fall), 2021, pp. 1–5.