Cloud Recognition in Hyperspectral Satellite Images Using an Explainable Machine Learning Model

Problem of developing algorithm based upon neutral networks and machine learning to find clouds on hyperspectral images are under consideration. It is required that the network is not a "black box," but allows an analysis of the reasons for decision making and classification results. Presented hybrid model includes decision tree trained to overcast recognition (model 1) on pre-selected features of an image in combination with convolutional neural network (model 2). Model 2 uses the result of model 1 and brightness in a selected band of an image. Model 1 finds cloud cores, and model 2 finds cloud edges. Results of testing the hybrid model on data of HYPERION sensor are presented. Data obtained over three surface types (ocean, plant, and urban region) are considered. Overall accuracy, as well as commission and omission errors are assessed. It is shown that the hybrid model can find 85% cloud pixels, only if the neural network is trained on an image where the contrast attains a maximum in the same spectral band. The results of this work can be applied to solve the general problem of analyzing and processing multispectral satellite images and further in environmental science and monitoring of changes in vegetation, ocean and glaciers.