On the importance of spatial and spectral features in the problem of cloud recognition for satellite images

Hyperspectral images are data sets, usually obtained from satellites, capable of capturing light spectra in numerous narrow ranges, which increases their informative value compared to standard color images. Models based on neural networks and machine learning are widely used for hyperspectral image analysis, in particular, in cloud recognition tasks. Despite their high classification accuracy, in many cases they are quite difficult to interpret. In most cases, this makes it difficult to analyze the results of their application, which makes it relevant to create algorithms for selecting significant features using explainable machine learning models.

The large amount of data in hyperspectral images is combined with correlations along neighboring spectral lines, which makes it necessary to select the most informative features. In this work, a feature selection method based on a classification algorithm is proposed, utilizing the existing labeling of the original satellite images acquired from the HYPERION sensor. The significance of the selected features was assessed from the point of view of cloud recognition quality for selected categories of images using an algorithm for iteratively excluding groups of features with high correlation. The final classifier was trained using significant spectral channels and derived features for three types of surfaces (ocean, vegetation, urbanized territory). A model based on a neural network is proposed for recognizing thick and thin cloud cover using selected spectral channels, and its performance is analyzed with consideration of the influence of spatial features on the cloud recognition results.