Semantic communication is increasingly recognized as a promising paradigm for enhancing the communication capabilities of wireless systems in the 6G era. Existing deep learning (DL)-based semantic methods typically enhance system robustness through module-centric strategies, where additional components are integrated into the model. Due to the significant computational overhead, these strategies are often unsuitable for resource-limited systems, such as the emerging low-altitude wireless networks (LAWNs). Moreover, most existing methods are optimized for fixed channel models and lack architectural adaptability across diverse environments, leading to repeated retraining and increased maintenance complexity. To address these challenges, we propose a novel Dual-Branch Architecture (DBA) for semantic communication. DBA employs a training-only auxiliary decoder branch to provide noise-free supervision for the main decoder, thereby enhancing robustness without adding deployment or inference overhead. Specifically, we introduce a contrastive learning mechanism to align the outputs of the noisy and noise-free branches, reinforcing semantic consistency. Building on this, we further propose the Extensible Multi-Channel Architecture (EMCA), a multi-branch design that incorporates multiple decoder branches optimized for different channel models and jointly trains them with a shared encoder and auxiliary branch, improving scalability without duplicating model parameters. Simulation results demonstrate that DBA and EMCA consistently outperform existing baselines in terms of semantic fidelity across additive white Gaussian noise (AWGN), Rayleigh, and Rician fading channels, without incurring any additional inference cost. Additionally, experiments conducted using a software-defined radio (SDR)-based platform with universal software radio peripheral (USRP) further validate the robustness and practicality of the proposed methods in practical environmental applications.