In hyperspectral image (HSI) classification, precise fusion classification within the same domain remains challenging due to spectral variability, noise interference, and class imbalance. Although UNet can effectively extract multi-scale features, its robustness, global dependency modeling, and cross-layer feature interaction are still limited in complex scenarios. To address these issues, this study proposes a multi-head cross-layer interaction UNet (MHCLI-UNet). By incorporating a multi-head self-attention (MHSA) mechanism, the network dynamically weights key spectral and spatial features, enhancing global context modeling while suppressing irrelevant noise. Additionally, a cross-layer interaction module (CLIM) is designed to adaptively fuse shallow detail features and deep semantic features through bidirectional information transfer. It is beneficial to capture fine-grained spatial-spectral characteristics. Experimental results on the datasets of MUUFL Gulfport and Houston demonstrate that MHCLI-UNet achieves significant improvements in classification accuracy and robustness.