SDFL-FC: Semisupervised Deep Feature Learning With Feature Consistency for Hyperspectral Image Classification

Semisupervised deep learning methods (DLMs) can mitigate the dependence on large amounts of labeled samples using a small number samples. However, for semisupervised feature (SDFL), quality extracted features cannot be well ensured without certain amount To address this issue, we develop SDFL method with consistency (SDFL-FC) hyperspectral image (HSI) classification. The SDFL-FC first adopts convolutional neural network (CNN) to extract spectral–spatial HSI and then uses fully connected layers (FCLs) model consistency. Moreover, two constraints that enforce both single pixel (FCS) group pixels (FCG) are introduced obtain representative discriminative features. FCS is achieved by generative adversarial (GAN) regularization, which reconstruct original data from FCG based assumption should have similar characteristics within superpixel, embedded in each FCL. final FCL outputs class labels, cross-entropy (CE) loss calculated samples, while losses all training (both unlabeled). integrates FCS, FCG, CE into unified objective function customized iterative optimization algorithm optimize it. Experiments demonstrate outperform related state-of-the-art classification methods.