To address performance degradation in semantic communication (SC) systems operating over time-frequency doubly-selective channels in high-mobility environments, this paper introduces a novel framework that integrates multiple-input multiple-output orthogonal time frequency space (MIMO-OTFS) modulation with semantic-aware resource allocation. The proposed approach intelligently maps semantic features to sub-channels based on both their relative importance and the instantaneous quality of each sub-channel. In this scheme, an entropy-based model quantifies the significance of individual semantic components, while the Kendall correlation coefficient is used to measure the consistency between feature importance and sub-channel conditions. The encoder and decoder are jointly trained using a composite loss function that jointly optimizes image classification accuracy, reconstruction fidelity, and sub-channel matching effectiveness. Simulation results show that, under high-mobility channels, the proposed MIMO-OTFS-SC scheme with semantic-aware sub-channel allocation reconstructs images more accurately than the orthogonal frequency division multiplexing-based semantic communication baseline.