Camera-assisted plant diagnosis can place image acquisition and neural feature extraction at a field terminal while an edge server performs the final classification. The uplink then transports an intermediate feature vector; for fixed quantization resolution and modulation order, the number of transmitted feature dimensions determines the modulation-symbol count. With imbalanced training classes, a single global feature ranking can allocate most transmitted coordinates to feature responses associated with the majority classes. This paper develops a class- routed encoder representation and a class-conditioned feature- index selection rule for a fixed-size transmitted subset. The D encoder coordinates are partitioned into class-associated groups. A routing loss concentrates the feature magnitude of class-c samples in the group assigned to class c. After training, classifier coefficients and class-conditioned feature magnitudes form a relevance matrix, from which a round-robin rule selects K unique coordinates. Experiments use eight tomato-leaf conditions, D = 128, K = 32, quadrature phase-shift keying (QPSK), an additive white Gaussian noise (AWGN) link at a signal- to-noise ratio (SNR) of 20 dB, and five independent random seeds. Relative to global saliency selection, the paired macro-F1 difference is +0.0068 ± 0.0242, with a 95% confidence interval of [−0.0232, 0.0369]. The mean F1 difference over four minority classes is +0.0199. The principal observed effect is class-wise discriminative information retention under a fixed transmitted subset size.

wireless edge inference, split inference, intermediate feature transmission, class imbalance, feature selection, class-wise F1.