In real-world scenarios, hyperspectral images (HSIs) often exhibit imbalanced class distributions, leading models to favor classes with more samples and to underperform on minority classes. Existing HSI classification methods typically fail to fully exploit multi-scale and multi-level features, and often produce redundant representations, limiting their classification performance. To address these challenges, we propose a novel multi-task learning framework based on dense feature pyramid network and multi-scale attention (DFPN-MSA) for class-imbalanced HSI classification. Our framework features three key innovations: (1) an efficient Multi-Scale Feature Extraction Module (MSFEM) that employs asymmetric convolutions to significantly reduce parameters while maintaining feature discriminability at different scales; (2) a Spectral-Spatial Attention Module (SSAM) that effectively suppresses feature redundancy through dual-branch weighting in spatial and spectral dimensions; (3) a Dense Feature Pyramid Network (DFPN) that enhances multi-level feature reuse and integration through dense connections and pyramid-structured decoding. To specifically address class imbalance, we introduce a reconstruction-based auxiliary task and propose a novel Reconstruction Loss-Weighted Focal Loss (RL-WFL) that dynamically adjusts classification weights based on sample-specific difficulty. Experiments on three public datasets demonstrate that DFPN-MSA outperforms state-of-the-art methods in both classification accuracy and computational cost.