Three-Dimensional Convective Updraft Cell Segmentation Using Deep Learning

Convection is the driving force of atmospheric circulation. They are related to cloud formation, precipitation, and many major weather events. The exact nature of the basic building blocks of convection, i.e., the individual convective updraft cells, has not been thoroughly explored because of the difficulties involved in accurately identifying them and the challenges in observing the vertical air velocity in true 3D form. Here we take the first step of segmenting and tracking individual updraft cells using cloud-resolving model simulations. In this study, we use a modified 3D-Mask R-CNN deep neural network architecture for convective cell segmentation. The instance masking capacity of the 3D-Mask R-CNN of each object in the problem domain is particularly valuable in various computer vision applications such as medical imaging, robotics, electron microscopy imaging, satellite image analysis, LiDAR data analysis, etc. The dataset used in this study was generated from a spectral bin microphysical scheme-based simulation of convective events near Darwin, Australia, during the monsoon period of the TWP-ICE field campaign, where the horizontal resolution was 1 km with varying vertical resolution and the outputs were saved every 10 minutes. Each geographical point is treated as a 3D pixel, with vertical air velocity dictating image intensity. The training dataset is refined through threshold-based image segmentation (Voronoi-Otsu Labeling) and human expert intervention, ensuring accurate contours for each convective cell. The corner points of bounded boxes are identified via Breadth-First Search (BFS). Our 3D-Mask R-CNN model maintains the foundational elements of the original architecture, including a backbone network for feature extraction and a Region Proposal Network (RPN) for identifying potential cell regions. The model learns not only to classify cells and predict their bounding boxes but also to generate a 3D mask for each convective cell. This mask delineates precise spatial dimensions, encapsulating the cell's volumetric extent.

Through instance masks, 3D-Mask R-CNN facilitates a comprehensive comprehension of convective cell contours, unveiling crucial insights into their shape and boundaries. This holistic understanding opens avenues for analyzing diverse weather phenomena related to atmospheric convection and their lifecycles. This not only allows for better grasp of the intricate interplay between atmospheric dynamics and precipitation processes, but also improve global climate models where cumulus parameterizations are necessary for long-term simulations.