In a recent paper “Self-Supervised Learning with Adaptive Distillation for Hyperspectral Image Classification”, Prof. Pedram Ghamisi and his team introduced a new deep learning method for analyzing hyperspectral images. Unlike the three spectral bands of red, green, and blue used in conventional images, hyperspectral images contain hundreds of narrow adjacent spectral bands. They combine two-dimensional spatial information with high-resolution spectral information, and allow detecting unique features and color signatures. Hyperspectral images are obtained by hyperspectral sensors/cameras that are widely used for Earth observation. Classification of these complex spatial-spectral data, which consists in determining the corresponding object category for each pixel, plays a key role in many remote sensing applications. Deep learning methods have a potential to significantly improve accuracy of hyperspectral image classification, but are limited by the scarcity of the labelled samples. The new method overcomes this problem by generating soft labels for unlabeled samples and using them to train the model. Soft labels are generated by calculating the spatial-spectral distances between the labelled and unlabeled samples; they reflect the probability the sample belongs to each predefined object class. Learn more…
