MMF-Net: multi-model fusion network for hyperspectral image classification based on transfer learning

Deep learning has demonstrated outstanding performance in hyperspectral image (HSI) classification. However, its generalization ability is limited by the high cost of annotations and strong cross-scene heterogeneity. To address these challenges, this paper proposes a multi-model fusion network (MMF-Net) based on transfer learning. First, a linear mapping layer is used to convert HSI data into a three-channel representation to mitigate the modality gap between the source domain (ImageNet) and the target domain (HSI). Then, three heterogeneous pre-trained models - Inception, VGG16, and Xception - are fine-tuned, with a channel attention mechanism incorporated to enhance discriminative features. Finally, high-level semantic features from multiple models are fused to construct a joint spectral-spatial-semantic representation, and a lightweight logistic regression classifier is employed for efficient pixel-wise classification. Experimental results show that the proposed method achieves good classification performance and generalization ability even with limited samples.