PhD Defense by Jannik Höller

Title: Development of a Deep Learning Approach for the Segmentation of Convective Cold Pools in Satellite-Observable Data

Abstract: Convective cold pools (CPs) result from convective downdrafts that are horizontally deflected when reaching Earth's surface. While spreading, CPs influence the distribution of low-level moisture and can trigger new convection by lifting warmer ambient air. Organizing convection, CPs are directly linked to extreme events such as mesoscale convective systems and thus crucial for the understanding of the underlying dynamics. Despite the socio-economic importance of such weather extremes, the role of CPs in convective organization is not yet sufficiently understood.

One of the main challenges is to accurately track dynamic CP gust fronts and link them to triggered convection. Although idealized models help us continuously disentangle the underlying mechanisms, they have so far been unable to capture the causal chains of CPs and lack observational benchmarks. Especially over tropical land, where CPs are ubiquitous, CP observations are mostly limited to event-based case studies due to the lack of relevant high-resolution data.

The goal of the present thesis is to (i) identify a detectable signature of CPs in geostationary satellite data, which is globally available with good spatio-temporal resolutions, and (ii) devise a corresponding approach for detecting CP regions.

In manuscript I, we address (i) by analyzing the space-borne signatures of 4218 CPs which we identified from almost 43 years of high-resolution near-surface data collected from twelve automatic weather stations spanning equatorial Africa. The identified CPs are accompanied by a mean decrease in satellite-derived brightness temperature of around 30 K. Moreover, in the majority of cases, CP gust fronts coincide with maximum brightness temperature decrease rates.

Modern deep learning methods can learn spatial patterns at various scales, making them well-suited to detect the identified CP signatures in satellite imagery and address (ii). However, their training requires substantial amounts of annotated ground truth data, which, for satellite data, can only be generated manually. To facilitate the generation of annotated data, we thus simulate cloud and rainfall fields based on different environmental conditions and automatically annotate the simulated scenes using a novel CP detection and tracking algorithm (CoolDeTA), presented in manuscript II. As CoolDeTA utilizes both thermodynamic and dynamic variables to detect and track CPs, the boundaries of the identified CPs align well with satellite-observable cloud signatures.

Using the annotated simulation scenes, we train convolutional neural networks for the segmentation of CPs in cloud and rainfall fields in manuscript III. Applied to an unknown test set, the trained neural networks achieve pixel accuracies of around 94% and successfully detect 83% of the CPs. In an additional case study, based on a simulation setup with realistic boundary conditions for a day over West Africa, the developed method confirmed its promising performance, further demonstrating its potential with respect to real satellite data. In the future, our method may open for large-scale observational studies of mesoscale CP dynamics and convective organization.