Wednesday 26th June, 2013
6:00pm to 6:15pm
Lightning mapping at radio frequencies (here with VHF Lightning Mapping Array data) is typically performed by a time-of-arrival source retrieval method. Thereafter, it is common to cluster the located sources into flash-level entities (often comprised of 10^2 - 10^3 sources) using space and time separation thresholds. A previously-used clustering algorithm was a one-off implementation in Fortran, and was designed without reference to the machine learning literature. This study replaces the previous algorithm, which had been wrapped into the Python-based lmatools workflow, with the general-purpose DBSCAN implementation in Python's sklearn package. The legacy code included substantial, file format-specific, I/O boilerplate. The new code clarifies the boundary between algorithm and I/O, and promotes clean integration with the rest of the lmatools infrastructure, aiding maintainability.
A chunked, streamed processing method was developed to account for continuous data rates that may exceed 10^5 four-coordinate (space and time) source vectors per minute. The chunking method exploits known physical limits to lightning flash duration, allowing the N^2 implementation of DBSCAN in sklearn to achieve real-time processing rates within available memory. The streaming technique is expected to be useful in future work as a flexible building block for end-to-end real-time and post-processing scripts and interactive analysis tools.
The algorithm is expected to find immediate use in our analysis of data from the NSF-sponsored Deep Convective Clouds and Chemistry campaign. The open nature of the underlying clustering libraries promotes code reuse by other research groups. Accounts of source-to-flash clustering in the literature are complemented by the availability of this open, objective reference implementation for clustering of lightning mapping datasets.
Texas Tech University
Sign in to add slides, notes or videos to this session