ACTA TECHNICA NAPOCENSIS ELECTRONICS AND TELECOMMUNICATIONS
URI permanent pentru această colecțiehttps://oasis.utcluj.app/handle/123456789/447
Navigare
Articol SPATIAL-SPECTRAL CLASSIFICATION OF HYPERSPECTRAL DATA WITH CONTROLLED DATA SEPARATION(Terebes Romulus, 2023-06-22) MICLEA Andreia Valentina; ABRUDAN MihaelaExploring spatial-spectral data frequently involves classifying hyperspectral images using convolutional neural networks. Due to the high complexity of the data and the scarcity of available training samples, hyperspectral image classification presents significant difficulties. In the context of supervised classification, we find that traditional experimental designs are frequently misused in the spectral-spatial processing context, resulting in unfair or biased performance evaluation, particularly when training and testing samples are selected at random from the same dataset. Under these circumstances, the dependence caused by the overlap between training and testing samples may be artificially increased, in breach of the data independence assumption upheld by supervised learning theory. In order to prevent an unbiased classification result, we present in this paper a controlled strategy designed to minimize the overlap between the samples present in the training and the testing data sets. The proposed controlled sampling strategy ensures a more trustworthy generalization of the CNN model by minimizing the issues present in the random sampling approach, such as the inability to determine whether or not an increase in classification accuracy is due to the spatial information incorporated into a classifier or to an increase in the overlap between training and testing data sets. Experiments performed with a wavelet CNN on different HSIs, namely Indian Pines, Pavia University, and Salinas, ensure the generalization of the data under the assumption that the training and data sets are independent from one another, based on a controlled strategy. Considering the high dimension of the HSI image, as a pre-processing step, the evaluation of the proposed framework is done by PCA and FA methods.