U-Net Segmentation for the Detection of Convective Cold Pools From Cloud and Rainfall Fields

Abstract

Convective cold pools (CPs) are known to mediate the interaction between convective rain cells and thereby help organize thunderstorm clusters, in particular mesoscale convective systems and extreme rainfall events. Unfortunately, the observational detection of CPs on a large scale has so far been hampered by the lack of relevant large-scale nearsurface data. Unlike numerical studies, where high-resolution near-surface fields of relevant quantities such as virtual temperature and winds are available and frequently used to detect cold pools, observational studies mainly identify CPs based on surface time series. Since research vessels or weather stations measure these time series locally, the characterization of cold pools from observations is limited to regional or station-based studies. To eventually enable studies on a global scale, we here develop and evaluate a methodology for the detection of CPs that relies only on data that (i) is globally available and (ii) has high spatio-temporal resolution. We trained convolutional neural networks to segment CPs in cloud and rainfall fields from high-resolution cloud resolving simulation output. Such data is not only available from simulations, but also from geostationary satellites that fulfill both (i) and (ii). The networks make use of a U-Net architecture, a common choice for image segmentation due to its strength in learning spatial correlations at different scales. Based on cloud and rainfall fields only, the trained networks systematically identify CP pixels in the simulation output. Our methodology may thus open for reliable global CP detection from space-borne sensors. As it also provides information on the spatial extent and the relative positioning of CPs over time, our method may offer new insight into the role of CPs in convective organization.

Detecting Cold Pool Family Trees in Convection Resolving Simulations

“…The code for the simple CP offspring model is licensed under MIT and published on GitHub https://github.com/Shakiro7/coldPool-detection-and-tracking. The CP detection and tracking algorithm (Cool-DeTA) was used in version 1.0 and is licensed under Creative Commons Attribution 4.0 International (Hoeller, 2023). CoolDeTA makes use of a watershed algorithm (van der Walt et al, 2014) and a k-means algorithm (Pedregosa et al, 2011).…”

“…Moreover, as CoolDeTA considers dynamic CP signatures, the identified CP boundaries align with the cloud patterns associated with cold pools in satellite images or corresponding simulation output. CoolDeTA thus offers a systematic and objective ground truth labeling for artificial intelligence methods that detect cold pools from simulated cloud fields and that potentially pave the way for future satellite‐based CP observations (Hoeller et al., 2024).…”

Characteristics of Station‐Derived Convective Cold Pools Over Equatorial Africa

“…While significant decreases in brightness temperature alone are not a sufficient criterion for identifying cold pool gust fronts, the concurrence with spatiotemporal patterns like radially spreading cloud arcs and rapidly expanding deep convection (Text S2, Figure S3 in Supporting Information ) could aid neural networks, such as those developed by Hoeller, Fiévet, et al. (2024), in limiting the number of false positive detections. Even in cases where not all parts of a cold pool gust front exhibit brightness temperature drops (Figure S2b in Supporting Information ), the presence of such spatiotemporal patterns may enable the neural networks to accurately track the gust front.…”

“…Recent idealized cloud‐resolving and large‐eddy simulations have provided new insight into CP structure and dynamics, such as on required mesh resolution (Fiévet et al., 2022), moisture rings (Drager et al., 2020; Langhans & Romps, 2015), or general interaction mechanisms (Haerter et al., 2020; Meyer & Haerter, 2020; Tompkins, 2001a; Torri et al., 2015), and triggered a range of simplified conceptual models (Böing, 2016; Haerter, 2019; Haerter et al., 2019; Niehues et al., 2022; Nissen & Haerter, 2021), which may help elucidate organizing mechanisms. New methods of CP detection in numerical studies have also been developed which help automatize the tracking of GFs and their interactions (Fournier & Haerter, 2019; Gentine et al., 2016; Henneberg et al., 2020; Hoeller, Fiévet, & Haerter, 2024; Hoeller, Fiévet, et al., 2024; Torri & Kuang, 2019).…”

U‐Net Segmentation for the Detection of Convective Cold Pools From Cloud and Rainfall Fields

“…The neural network was implemented using PyTorch version 1.11.0 (Paszke et al., 2019). Both the code and the data sets for the training and testing of the neural networks were used in version 1.0 and are licensed under Creative Commons Attribution 4.0 International (Hoeller, Fiévet, Engelbrecht, & Haerter, 2023). The idealized simulations used for training and testing are run with the cloud‐resolving three‐dimensional atmosphere simulator System for Atmospheric Modeling (SAM; Khairoutdinov & Randall, 2003), version 6.11.…”