FactoFormer: Factorized Hyperspectral Transformers with Self-Supervised Pretraining

Hyperspectral images (HSIs) contain rich spectral and spatial information.Motivated by the success of transformers in the field of natural languageprocessing and computer vision where they have shown the ability to learn longrange dependencies within input data, recent research has focused on usingtransformers for HSIs. However, current state-of-the-art hyperspectraltransformers only tokenize the input HSI sample along the spectral dimension,resulting in the under-utilization of spatial information. Moreover,transformers are known to be data-hungry and their performance relies heavilyon large-scale pretraining, which is challenging due to limited annotatedhyperspectral data. Therefore, the full potential of HSI transformers has notbeen fully realized. To overcome these limitations, we propose a novelfactorized spectral-spatial transformer that incorporates factorizedself-supervised pretraining procedures, leading to significant improvements inperformance. The factorization of the inputs allows the spectral and spatialtransformers to better capture the interactions within the hyperspectral datacubes. Inspired by masked image modeling pretraining, we also devise efficientmasking strategies for pretraining each of the spectral and spatialtransformers. We conduct experiments on six publicly available datasets for HSIclassification task and demonstrate that our model achieves state-of-the-artperformance in all the datasets. The code for our model will be made availableat https://github.com/csiro-robotics/factoformer.