A machine learning based approach to melting layer detection

Abstract

Stratiform precipitation is one of the most important precipitation systems in the mid-latitudes. To gain understanding of the melting layer, which is part of a stratiform precipitating system, an algorithm which is able to select melting layer data from large datasets is required. The goal of this thesis work is to create and deliver a program which is able to performthis work. For development the available data was separated into three groups: no melting layer, only melting layer and the rest. Based on the second group, typical melting layer signatures are analysed. In particular the different signatures (reflectivity, polarimetric and Doppler) occur at different heights. Based on the knowledge gained from the analysis and focussing on the reflectivity and polarimetric signatures, three different approaches were taken to detect and characterise the melting layer fromthe data. The first approach is based on an existing method in literature, this approach acts as the reference method. The second approach is based on image processing techniques, while the third approach is based on machine learning techniques. The second approach is later abandoned because of limitations of the techniques investigated. The third approach, machine learning, is the main contribution of this thesis work. For analysis of the performance of the different algorithms an annotated test dataset is created which represents the entire dataset. The performance is determined by the melting layer detection probability using a confusion matrix and by the determination of errors of the melting layer boundaries. The reference method proved to show an extremely low false positive rate (0%) on the test dataset. This means that if the method detected a melting layer, it is almost certain that there is one. The overall detection probability was 80%. The method fails in detecting the melting layer when the peak reflectivity is below the threshold (30dBZ) used in the method. The detected upper melting layer boundary of the reference method is on average 365 m lower compared to the ground truth. The lower boundary is on average 149 m higher than the ground truth. This means that the reference method only selects a part of the entire thickness of the ML. The correlation between the lower boundary and the ground truth is higher than the upper boundary (0.987 vs. 0.965). The proposed machine learning method has a detection performance of almost 94%, which is higher than the reference method, but some false positives occur (3%). The upper boundary is on average 20 m above the ground truth, while the lower boundary is 66 m lower than the ground truth. The machine learned method thus overestimates the thickness of the ML. The correlation between the lower boundary and the true boundary is 0.976 and the correlation between the upper boundary and the true boundary is 0.966. The upper boundaries of both methods have a very similar correlation, while the lower boundary of the machine learning method has a lower correlation. However, despite the lower correlation of the lower boundary and the introduction of false positives, the machine learning method is an improvement over the reference method since it has a significantly higher detection probability and it captures the entire thickness of theML much better. It is therefore suited forML analysis.