Semantic Segmentation of Hyperspectral Imaging Using Convolutional Neural Networks

neural networks in hyperspectral imaging helps to get through the obstruction to solving data analysis, classification, and segmentation problems. There are problems, such as vegetations analysis in agriculture, which cannot be solved using classic RGB images due to lack of information. Applying neural networks to hyperspectral images is a sophisticated problem. The aim of this study is to examine concerns about using convolutional neural networks for the semantic segmentation of hyperspectral data. The following problems were considered: large spatial resolution, the influence of neural network's input size on accuracy and performance; hyperspectral data preprocessing, the influence of dimensionality reduction and brightness equalization; neural network architecture influence on analyzing hyperspectral imaging. Also, the accuracy of neural networks was compared to classic approaches: multinominal logistic regression, random forest algorithm, discriminant analysis. As the result of the study the importance of choosing neural network's architecture and hyperspectral data preprocessing methods are discussed.