Convolutional Neural Networks (CNNs) are widely used in acoustic scene classification (ASC) for their ability to extract informative features by fusing time-frequency and channel information. However, in resource-constrained environments, CNN-based ASC models often suffer from limited accuracy and robustness. To address this, we propose an Attention-Fused Depthwise Separable Network (AFDS-Net), which leverages ERes2NetV2 as its backbone and enhances performance through two key components. First, a Depthwise Feature Fusion (DFF) block is developed to facilitate the precise aggregation of local features within the residual block, leveraging an improved attention mechanism to preserve fine-grained feature representation effectively. Second, a Channel Feature Enhancement (CFE) block integrates Depthwise Separable Convolution (DSConv) with a Squeeze-and-Excitation (SE) mechanism for adaptive cross-channel weighting. Building on these, a Bottom-up Two-stage Feature Fusion (BTFF) enhances time-frequency modeling for short audio by synergizing DFF and CFE. Experiments on the TAU 2020 dataset show that AFDS-Net achieves 75.3% accuracy with a comparable parameter count, delivering a 50.2% absolute improvement over the baseline and strong generalization across datasets.