Abstract
Advanced satellite-born remote sensing instruments produce high-resolution multi-spectral data for much of the globe at a daily cadence. These datasets open up the possibility of improved understanding of cloud dynamics and feedback, which remain the biggest source of uncertainty in global climate model projections. As a step towards answering these questions, we describe an automated rotation-invariant cloud clustering (RICC) method that leverages deep learning autoencoder technology to organize cloud imagery within large datasets in an unsupervised fashion, free from assumptions about predefined classes. We describe both the design and implementation of this method and its evaluation, which uses a sequence of testing protocols to determine whether the resulting clusters: (1) are physically reasonable, (i.e., embody scientifically relevant distinctions); (2) capture information on spatial distributions, such as textures; (3) are cohesive and separable in latent space; and (4) are rotationally invariant, (i.e., insensitive to the orientation of an image). Results obtained when these evaluation protocols are applied to RICC outputs suggest that the resultant novel cloud clusters capture meaningful aspects of cloud physics, are appropriately spatially coherent, and are invariant to orientations of input images. Our results support the possibility of using an unsupervised data-driven approach for automated clustering and pattern discovery in cloud imagery.
Highlights
A DVANCED satellite-borne remote sensing instruments such as the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Terra and AquaManuscript received March 2, 2021; revised July 2, 2021; accepted July 11, 2021
This framework combines five elements to provide this capability: 1) MODIS data as a source of satellite radiance data, from which we extract cloud features that we show to be well associated with physical metrics; 2) convolutional neural networks (CNNs) [23], [24] to extract useful representations of spatial patterns from images; 3) a loss function that uses transform-invariant techniques [22] to produce latent space representations that are independent of input image orientations; 4) an autoencoder training protocol that learns from satellite data without introducing biases; and 5) hierarchical agglomerative clustering (HAC) [25] to extract novel clusters from the latent representation
When Wood and Hartmann [33] investigated the relationship between the complex textures of low clouds and the physical characteristics found within common mesoscale (1–100 km) cloud organizations, they found that four frequent mesoscale cloud patterns associated with the existence of open- and closed-cell structures, often classified as stratus or stratocumulus in the International Satellite Cloud Climatology Project (ISCCP) classification, occur in different geographical regions and have different distributions of liquid water path, suggesting that the spatial variability of low clouds is underrepresented in the standard cloud types
Summary
A DVANCED satellite-borne remote sensing instruments such as the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Terra and Aqua. We present in this article a new rotation-invariant cloud clustering (RICC) autoencoder framework capable of automating the clustering of cloud patterns and textures without any assumptions concerning artificial cloud categories This framework combines five elements to provide this capability: 1) MODIS data as a source of satellite radiance data, from which we extract cloud features that we show to be well associated with physical metrics; 2) convolutional neural networks (CNNs) [23], [24] to extract useful representations of spatial patterns from images; 3) a loss function that uses transform-invariant techniques [22] to produce latent space representations that are independent of input image orientations; 4) an autoencoder training protocol that learns from satellite data without introducing biases; and 5) hierarchical agglomerative clustering (HAC) [25] to extract novel clusters from the latent representation.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
