A data-driven cloud classification framework based on a rotationally invariant autoencoder

Abstract:

Clouds are the biggest source of uncertainty in projections from global climate models, and this uncertainty is unlikely to be reduced in the near future. Satellite instruments have however produced a rich dataset of observations spanning several decades that can be used to improve understanding of cloud dynamics and feedbacks. Due to the large size and diversity of this dataset, conventional statistical methods are often poorly suited to extracting the relevant features used to distinguish among different types of standard cloud classes (which can also be overly restrictive). In this study we use an unsupervised deep-learning framework to automate the classification of cloud patterns and textures without any assumptions concerning artificial cloud categories. Our automated cloud-classification scheme extracts relevant cloud features from satellite radiance data, which we show to be well correlated with physical metrics. We train a fifteen-layer Convolutional Autoencoder (CAE) with rotational-transform-invariant loss to extract spatially and physically meaningful features (e.g., separation of low and high cloud as well as liquid and ice phase of cloud particles) from MODIS Terra satellite calibrated radiances (MODIS021KM, MOD02 bands 6, 7, 20, 28, 29, and 31) and cloud masks (MOD35_L2) over a total of 60,000 images. While a standard autoencoder produces different representations if trained on identical images with different rotations, our rotationally invariant approach mitigates such misclassification errors. We then apply hierarchical agglomerative clustering to the dimensionally reduced latent space representations produced by the convolutional layer to identify novel cloud types. We validate our model on 2,000 images from the period 2000 to 2018, not used in the training sample, and show by capturing rotationally invariant spatial information in the latent representation our framework is able to provide meaningful associations between images and retrieved physical parameters.