Addressing the challenges of insufficient modeling of complex nonlinear mixing effects by traditional physical models in hyperspectral unmixing (HU) and the weak generalization capability of purely data-driven methods due to the lack of physical priors, this study proposes an unmixing method (M2S2) that integrates mechanistic constraints with multi-scale spectral-spatial joint learning. The aim is to resolve the failure of hyperspectral sub-pixel analysis caused by nonlinear mixing effects in complex scenarios. Accordingly, this paper first constructs physical constraint conditions for the endmember abundances, including non-negativity, sum-to-one, and sparsity. Subsequently, distinct core modules are designed to hierarchically extract features from both spectral and spatial data, leveraging multi-scale pooling structures and multi-channel attention mechanisms to highlight the saliency of hyperspectral mixed information and enhance the model’s unmixing performance. Comparative experiments on multiple datasets demonstrate that the proposed M2S2 method improves hyperspectral unmixing performance by approximately 6.8% compared to other methods. The findings validate the effectiveness of combining physical mechanistic priors with deep learning models and conducting feature enhancement through multi-scale spectral-spatial joint learning, offering a novel approach for addressing the challenges of nonlinear hyperspectral unmixing.