Abstract
Hyperspectral imaging is a rich source of data, allowing for multitude ofeffective applications. On the other hand such imaging remains challengingbecause of large data dimension and, typically, small pool of availabletraining examples. While deep learning approaches have been shown to besuccessful in providing effective classification solutions, especially for highdimensional problems, unfortunately they work best with a lot of labelledexamples available. To alleviate the second requirement for a particulardataset the transfer learning approach can be used: first the network ispre-trained on some dataset with large amount of training labels available,then the actual dataset is used to fine-tune the network. This strategy is notstraightforward to apply with hyperspectral images, as it is often the casethat only one particular image of some type or characteristic is available. Inthis paper, we propose and investigate a simple and effective strategy oftransfer learning that uses unsupervised pre-training step without labelinformation. This approach can be applied to many of the hyperspectralclassification problems. Performed experiments show that it is very effectivein improving the classification accuracy without being restricted to aparticular image type or neural network architecture. An additional advantageof the proposed approach is the unsupervised nature of the pre-training step,which can be done immediately after image acquisition, without the need of thepotentially costly expert's time.