In the realm of semantic communication, the significance of encoded features can vary, while wireless channels are known to exhibit fluctuations across multiple subchannels in different domains. Consequently, critical features may traverse subchannels with poor states, resulting in performance degradation. To tackle this challenge, we introduce a framework called Feature Allocation for Semantic Transmission (FAST), which offers adaptability to channel fluctuations across both spatial and temporal domains. In particular, an importance evaluator is first developed to assess the importance of various features. In the temporal domain, channel prediction is utilized to estimate future channel state information (CSI). Subsequently, feature allocation is implemented by assigning suitable transmission time slots to different features. Furthermore, we extend FAST to the space-time domain, considering two common scenarios: precoding-free and precoding-based multiple-input multiple-output (MIMO) systems. An important attribute of FAST is its versatility, requiring no intricate fine-tuning. Simulation results demonstrate that this approach significantly enhances the performance of semantic communication systems in image transmission. It retains its superiority even when faced with substantial changes in system configuration.