3-D-ANAS: 3-D Asymmetric Neural Architecture Search for Fast Hyperspectral Image Classification

Hyperspectral images (HSIs) provide abundant spectral and spatial information, playing an irreplaceable role in land-cover classification. Recently, based on deep learning (DL) technologies, increasing number of HSI classification approaches have been proposed, which demonstrate promising performance. However, previous studies suffer from two major drawbacks: 1) the architecture most DL models is manually designed, relies specialized knowledge, relatively tedious. Moreover, classifications, datasets captured by different sensors physical properties. Correspondingly, need to be designed for datasets, further increases workload designing architectures 2) mainstream framework a patch-to-pixel framework. The overlap regions patches adjacent pixels are calculated repeatedly, computational cost time cost. In addition, accuracy sensitive patch size, artificially set extensive investigation experiments. To overcome issues mentioned above, we first propose 3-D asymmetric neural network search algorithm leverage it automatically efficient classifications. By analyzing characteristics HSIs, specifically build decomposition space, where information processed with convolutions. Furthermore, new fast framework, i.e., pixel-to-pixel has no repetitive operations reduces overall Experiments three public networks our (3-D-ANAS) achieve competitive performance compared several state-of-the-art methods, while having much faster inference speed. Code available at: https://github.com/hkzhang91/3D-ANAS .