Using Scikit-Learn to Increase the Precision of an Automated Mesoscale Convective System Segmentation and Tracking Procedure

This research evaluates and demonstrates the utility of a mesoscale convective system (MCS) segmentation, classification, and tracking framework implemented in Python. In particular, the use of select scikit-learn algorithms is explored for the purpose of augmenting the process of extracting qualifying rainfall clusters from radar reflectivity images (i.e., segmentation). Segmentation is performed by identifying contiguous or semi-contiguous regions of deep, moist convection that are organized on a horizontal scale of at least 100 km. Classification is performed by first compiling a database of thousands of precipitation clusters, and then subjectively assigning each sample one of the following labels: 1) midlatitude MCS; 2) unorganized convective cluster; 3) tropical system; 4) synoptic system; and 5) ground clutter and/or noise. The attributes of each sample, along with their assigned label, are then used to train scikit-learn algorithms. Results using a testing dataset suggest that the algorithms can distinguish between MCS and non-MCS samples with high specificity, precision, and sensitivity. To track segmented clusters through time, a custom tracking algorithm spatiotemporally associates classified segments using three-dimensional labeling. Various reflectivity and prediction probability thresholds are explored to determine an approach that captures the majority of actual MCS events while ignoring non-MCS events. The results suggest that machine learning can add value to an MCS-tracking approach by limiting the amount of false-positive (non-MCS) samples that are not removed by segmentation alone.