Detecting Activity of Tropical Cyclones with the Unsupervised Maximally Divergent Interval Algorithm

Kurzfassung

The goal of this master thesis is the detection of tropical cyclones (TCs) by means of a multivariate unsupervised machine learning algorithm, the Maximally Divergent Intervals (MDI) method. Compared to traditional approaches of TC detection, it does not rely on hard, grid-size dependent thresholds of variables. Based on ERA-Interim reanalysis and the TC database IBTrACS, a labeled data set was created that allows the verification of the detections. The MDI algorithm is applied to the Gulf of Mexico during the hurricane seasons 2000 to 2010. First, the ideal initial settings were elaborated: this showed that the Kullback-Leibler divergence should be used to identify anomalies, no embedding should be applied and that the detections should be shifted one time step forward. It was subsequently found that the algorithm achieves the best detection skill with a mean Average Precision (mAP) of 0.537 when applied univariately to the relative vorticity at 850 hPa. A multivariate application involving other variables did not improve the mAP. In order to minimize the false alarm ratio, soft variable thresholds in wind speed of 8 ms-1 and relative vorticity of 1 · 10-5 s-1 have been introduced. Since the algorithm assigns scores to its detections, an additional score threshold of 1150 was defined. These measures reduced the false alarm rate to 0.162. Allover, a detection scheme with a precision of 0.838 and a probability of detection of 0.455 was designed. Beyond that, the MDI algorithm proved to be suitable for estimating the strength of TC activity: the sum of the scores of individual hurricane seasons correlates statistically significantly (r = 0:9) with the accumulated cyclone energy (ACE). In general, the algorithm is recommended for detecting less specific anomalies that are still more unexplored in terms of their nature than TCs.