Home – Home – Jornals – Atmospheric and Oceanic Optics 2026 number 2

This article addresses the problem of cloud detection in hyperspectral satellite images using an interpretable neural network classifier for partial cloudiness. For effective solution, the preliminary selection of spectral channels and derived features is performed using decision trees trained with labeled satellite data of the HYPERION sensor. The selected channels and features are then used for building a convolutional neural network based on a modified Unet architecture. Modifications to the original Unet architecture enable simplifying the network structure, avoiding overfitting, assessing the importance of spatial and spectral features, analyzing classification results, and explaining decision-making processes. Feature selection and evaluation of their importance are critical stages in developing adequate machine learning and deep learning models combined with the analysis of their generalization ability. The suggested feature selection method is based on iterative training of decision trees to identify significant features in terms of classification accuracy. The operation of the convolutional neural network is interpreted and the importance of spatial and spectral features is assessed by evaluating Shapley vectors. The results of testing a neural network with HYPERION images made over three surface types (ocean, vegetation, and urbanized territory) are presented; its accuracy and commission and omission errors are estimated. The model enables semantic segmentation of images with thin clouds with accuracy over 95% in selected spectral bands and with selected features. The importance of input features, caused by their distribution across spectral channels and the relative positions of pixels in an image, for the detection of thick and thin clouds in hyperspectral satellite images is analyzed. The presented neural network model is designed for working with limited data volumes, enables applying augmentation, and can be used to assess the importance of selected spectral channels and spatial features.