Toward a Unified Semantic Loss Model for Deep JSCC-based Transmission of EO Imagery

Modern Earth Observation (EO) systems increasingly rely on high-resolution imagery to support critical applications such as environmental monitoring, disaster response, and land-use analysis. Although these applications benefit from detailed visual data, the resulting data volumes impose significant challenges on satellite communication systems constrained by limited bandwidth, power, and dynamic link conditions. To address these limitations, this paper investigates Deep Joint Source-Channel Coding (DJSCC) as an effective source-channel paradigm for the transmission of EO imagery. We focus on two complementary aspects of semantic loss in DJSCC-based systems. First, a reconstruction-centric framework is evaluated by analyzing the semantic degradation of reconstructed images under varying compression ratios and channel signal-to-noise ratios (SNR). Second, a task-oriented framework is developed by integrating DJSCC with lightweight, application-specific models (e.g., EfficientViT), with performance measured using downstream task accuracy rather than pixel-level fidelity. Based on extensive empirical analysis, we propose a unified semantic loss framework that captures both reconstruction-centric and task-oriented performance within a single model. This framework characterizes the implicit relationship between JSCC compression, channel SNR, and semantic quality, offering actionable insights for the design of robust and efficient EO imagery transmission under resource-constrained satellite links.

💡 Research Summary

The paper addresses the pressing challenge of transmitting high‑resolution Earth Observation (EO) imagery from low‑Earth‑orbit (LEO) satellites to ground stations under severe bandwidth, power, and dynamic channel constraints. Traditional separate source‑ and channel‑coding pipelines (e.g., JPEG2000 + LDPC) are inefficient in such environments because the separation theorem’s assumptions rarely hold when the channel conditions fluctuate rapidly. To overcome this, the authors explore Deep Joint Source‑Channel Coding (DJSCC), a neural‑network‑based paradigm that learns an end‑to‑end mapping jointly optimizing compression and robustness to channel noise.

Two complementary notions of “semantic loss” are investigated. The first is reconstruction‑centric: the authors evaluate how the visual fidelity of the reconstructed image degrades as a function of compression ratio (ρ) and channel signal‑to‑noise ratio (γ). They report standard image quality metrics—PSNR, SSIM, and MSE—across a grid of ρ ∈ {2, 4, 6, 8, 12} and γ ∈ {−6,…,8 dB}. Results show a clear trade‑off: higher compression and lower SNR lead to substantial drops in PSNR (down to ~30 dB) and SSIM (below 0.7), confirming that both parameters jointly dictate reconstruction quality.

The second notion is task‑oriented: the DJSCC decoder is coupled with a lightweight vision transformer, EfficientViT, to perform land‑use/land‑cover classification directly on the reconstructed images. Here, semantic loss is measured by downstream performance metrics—accuracy, precision, and recall—rather than pixel‑level similarity. Experiments on the EuroSAT dataset (27 000 images, 64 × 64 pixels, 13 spectral bands) reveal that, even under aggressive compression (ρ = 12), high SNR (γ ≥ 4 dB) preserves classification accuracy above 95 %. Conversely, when γ falls below 0 dB, accuracy deteriorates sharply, underscoring the importance of channel quality for task performance.

The core contribution is a unified, two‑dimensional semantic loss model that simultaneously captures reconstruction‑centric and task‑oriented degradations. The proposed functional form (Equation 1) expresses the loss ξ as a sum of terms each comprising a sigmoid of SNR (capturing channel distortion) multiplied by an exponential of compression ratio (capturing source distortion). The model parameters (μ₀…μ₆) are learned via gradient descent (Algorithm 1), with carefully tuned learning rates to ensure convergence. Compared against two prior one‑dimensional models—a generalized sigmoid (G‑Sigmoid) and a sum‑of‑exponentials (Sum‑Exp)—the unified model achieves dramatically lower fitting errors (average MSE between measured and fitted values reduced from ~0.14–0.25 to <0.04). This demonstrates that jointly modeling ρ and γ yields a far more accurate predictor of semantic quality.

Methodologically, the study is thorough: the EuroSAT dataset provides realistic multispectral EO imagery; the compression ratios span a wide range, and the SNR values cover both favorable and severely degraded channel conditions. Both image‑level (PSNR, SSIM, MSE) and classification‑level (accuracy, precision, recall) metrics are reported, enabling a holistic assessment of the trade‑offs. The gradient‑based fitting procedure is detailed, and the resulting parameter tables (Tables IV and V) illustrate the model’s adaptability to different metrics.

In the discussion, the authors highlight practical implications. The unified loss model can serve as a design tool for satellite communication engineers: given a target downstream task accuracy, the model can predict the minimum SNR and maximum permissible compression ratio, guiding link‑budget allocation and on‑board encoder configuration. Moreover, the use of EfficientViT—a model with modest computational and memory footprints—makes the task‑oriented pipeline feasible for on‑board inference or for rapid ground‑station processing after reception.

The conclusion reiterates that the proposed data‑fitting framework provides a comprehensive approach to semantic communication with DJSCC in EO scenarios, outperforming existing models and offering actionable insights for system design. Future work is suggested to extend the model to more complex channel models (e.g., fading, multi‑path) and additional downstream tasks such as object detection or change detection, as well as to validate the approach on real satellite hardware.

Overall, the paper makes a significant contribution by bridging the gap between source‑channel coding theory and application‑level performance in satellite‑based Earth observation, delivering a unified semantic loss model that can guide the development of robust, bandwidth‑efficient, and task‑aware communication systems.