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

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

Problem Statement
Modern Earth Observation (EO) systems generate high-resolution imagery essential for applications like environmental monitoring, disaster response, and land-use analysis. However, the increasing resolution and acquisition frequency of EO sensors create data volumes that strain satellite communication systems, which are constrained by limited bandwidth, strict power budgets, and dynamic channel conditions. Traditional transmission methods, which separate source coding (e.g., JPEG2000) and channel coding (e.g., LDPC), often fail to optimize performance under these constraints, particularly when the separation theorem does not hold. While Deep Joint Source–Channel Coding (DJSCC) offers a promising end-to-end alternative, there is a lack of tractable models to quantify "semantic loss." Existing models often treat source and channel distortions as independent or fail to capture the two-dimensional relationship between compression parameters and channel quality, making it difficult to balance image reconstruction fidelity with downstream task performance.

Methodology
The authors propose a data-driven approach to model semantic loss in DJSCC-based EO transmission systems. The study utilizes the EuroSAT dataset, a large-scale benchmark of 27,000 Sentinel-2 satellite images (64x64 pixels, 13 spectral bands) across 10 land-use classes.

The experimental framework evaluates two complementary aspects of DJSCC performance:

1. Reconstruction-Centric: Measuring visual fidelity using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Mean Squared Error (MSE) under varying compression ratios ($\rho$) and channel Signal-to-Noise Ratios (SNR, $\gamma$).
2. Task-Oriented: Integrating the DJSCC framework with a lightweight, application-specific classifier, EfficientViT, to measure downstream task accuracy, precision, and recall.

To model the relationship between the system parameters ($\rho$ and $\gamma$) and the performance metrics, the authors employ a unified semantic loss model defined as a non-linear function. Since standard curve-fitting tools could not handle the required multi-dimensional structure, the authors developed a custom gradient descent optimization algorithm (Algorithm 1) to estimate the model parameters ($\mu_{i,j}$). The proposed model (Equation 1) is a two-dimensional function that jointly depends on both the compression ratio and the channel SNR.

Key Contributions

• Unified Semantic Loss Framework: The paper introduces a novel model that captures both reconstruction-centric and task-oriented performance within a single mathematical framework. Unlike previous works that model distortion as a sum of independent components or use one-dimensional functions of SNR, this model explicitly characterizes the implicit, coupled relationship between DJSCC compression settings, channel conditions, and semantic quality.
• Empirical Validation: The study provides extensive empirical results across six different metrics (PSNR, SSIM, MSE, Accuracy, Precision, Recall) and two evaluation aspects (image reconstruction and classification) using real-world EO data.
• Comparative Analysis: The proposed model is benchmarked against existing semantic loss models, specifically the Generalized Sigmoid model [9] and the Sum-of-Exponential model [10].

Results
Extensive experiments demonstrate that the proposed unified model significantly outperforms existing approaches.

• Fitting Accuracy: Table III shows the average Mean Squared Error (MSE) between measured data and fitted values. For a compression ratio of $\rho=12$, the proposed model achieved an MSE of 0.0407, compared to 0.2474 for the Generalized Sigmoid model and 0.0521 for the Sum-Exp model. Similar improvements were observed at $\rho=8$.
• Dimensionality: While existing models are restricted to one-dimensional functions of SNR (requiring fixed compression ratios for fitting), the proposed model successfully captures the two-dimensional behavior of the system, allowing for simultaneous optimization of compression and channel parameters.
• Visual Fit: Figures 2 and 3 illustrate that the proposed curve fitting aligns closely with empirical data across all tested scenarios, including both the raw DJSCC-AWGN case and the EfficientViT-DJSCC-AWGN case.

Significance
The paper claims that this unified semantic loss framework offers actionable insights for the design of robust and efficient EO imagery transmission systems under resource-constrained satellite links. By providing a tractable model that links compression parameters and channel conditions to application-level outcomes, the work addresses a critical gap in semantic communication system design. It enables engineers to better predict how transmission impairments affect both the visual quality of reconstructed images and the accuracy of downstream AI tasks, facilitating the development of bandwidth-efficient and task-aware EO communication systems. The authors position this work as an extension of their prior research on semantic communication using the DVB-S2 standard, now specifically tailored to the complexities of DJSCC in Earth Observation.