Goal-Oriented Communication With Adaptive Semantic Representation Selection for Satellite Image Classification

Earth observation satellites continuously capture Earth images, generating large volumes of visual data that must be transmitted over bandwidth-limited satellite-to-ground links for remote inference. Existing deep joint source-channel coding (DJSCC) methods optimize transmission for image reconstruction, requiring complex models and significant channel resources to preserve content that is largely irrelevant to semantic inference tasks such as image classification. This paper proposes an end-to-end goal-oriented semantic communication framework for image classification over a satellite-to-ground erasure channel, where the semantic encoder, erasure-adaptive feature selection gate and task decoder are jointly optimized using only the classification objective, transmitting task-relevant latent representations rather than aiming at reconstructing the original image. The encoder conditions semantic feature extraction on the estimated erasure probability via a multi-head self-attention module, while an adaptive gate dynamically determines the number of transmitted dimensions through learned importance scores, without any predefined transmission target. Experiments on CIFAR-10 demonstrate classification accuracies of 86.32%- 89.29% over the erasure probability ϵ ∈ [0,0.8], while transmitting only 2.5%-14.4% of the 1024 latent dimensions, consistently outperforming the baseline across the evaluated erasure range.