A hybrid deep learning framework for SEM-based air pollutant analysis: Mamba integration and GAN-augmented training

Air pollution poses severe threats to public health and ecological stability, making accurate analysis of airborne pollutant composition increasingly vital. In this paper, we propose a novel deep learning framework for efficient classification of pollutant components based on microscopic or spectral images. The proposed model integrates the recent Mamba mechanism, a state space model (SSM) architecture known for its superior long-range dependency modeling and linear computational complexity, into the image classification pipeline. By leveraging convolutional layers for local feature extraction and Mamba blocks for global semantic representation, our approach significantly improves both detection accuracy and inference speed compared to traditional CNN or Transformer-based baselines. To address the challenge of limited labeled data, we further introduce a generative adversarial network (GAN)-based data augmentation strategy. A CGAN is trained to synthesize realistic SEM-like particulate images, which are then incorporated into the training set to expand the training dataset. This integration of generative modeling effectively mitigates overfitting and strengthens the model's ability to generalize across varied pollutant types and imaging conditions. Experimental results on benchmark demonstrate the model's effectiveness in identifying common airborne constituents.