MSC-StyleACGAN:A style‑based GAN with multiscale attention for hyperspectral data augmentation in tobacco mildew detection

Accurate and nondestructive detection of mildew in tobacco leaves is crucial for ensuring quality and safety across postharvest processing stages. Hyperspectral imaging (HSI) offers sensitive spectral information for mildew detection. However, deep learning–based classification models are highly dependent on large and well-balanced datasets, which are difficult to obtain in practice due to severe class imbalance and inter-cultivar variability. In this article, we propose MSC-StyleACGAN, a novel style-based conditional generative adversarial network (GAN) for data augmentation of HSI spectra. First, a novel hybrid conditioning mechanism is proposed, which combines direct input concatenation for global consistency with hierarchical style modulation via a mapping network and Adaptive Instance Normalization (AdaIN) for feature control. Second, a lightweight multiscale channel attention (MSCA) module, composed of parallel depthwise-separable convolutions with channel attention, is employed to capture global spectral trends and narrow absorption features characteristic of mildew. Finally, the discriminator adopts the auxiliary classifier GAN (AC-GAN) framework and integrates WGAN-GP to enhance training stability. Experimental results on real tobacco datasets demonstrate that the proposed MSC-StyleACGAN approach generates higher quality HSI data than state state-of-the-art GAN, and improves mildew classification performance.