Using Self-Organizing Maps to Identify Coherent CONUS Precipitation Regions: Application to Extremes Aggregation
Abstract
Extreme precipitation events have major societal impacts. These events are rare by definition and can have small spatial scale making statistical analysis difficult; both factors are mitigated by combining events over a region. A methodology is presented to objectively define "coherent" regions where data points have matching annual cycles within each region. Regions are found by training a self-organizing map (SOM) on the annual cycle of precipitation for each grid point across the contiguous United States (CONUS). This allows the regions to be agnostic about extremes data when used to aggregate extreme events. There is no agreed upon method of choosing the number of nodes in a SOM and is often application dependent, though statistical tests of distinguishability exist (Johnson 2013). Useful numbers of regions (nodes) in this application balance two conflicting preferences: larger regions contain more events and thereby have more robust statistics, but more compact regions allow weather patterns associated with extreme events to be aggregated confidently. No single criterion is suitable for this. We must define novel criteria to assess these properties for each region: having many more events than occur at a single grid point, connectedness, and compactness. We create a Regional Extremes Ratio (RER) which compares regional events to grid cell events. RER favors large regions that necessarily contain many events. To measure compactness we use the nondimensional ratio of the square root of a region's area to its perimeter, penalizing long thin regions. Connectedness has two aspects. First, we want to minimize the number of separate areas that comprise each region. Second, each region should be a single large area and any other areas be individually and collectively small. To measure this connectedness attribute, two metrics are designed. Optimization between, on the one hand, connectedness and compactness, and on the other, RER identifies 12-15 regions over the CONUS. Our methodology is applicable across datasets whether gridded or station. Results will be shown from both reanalysis and gridded observational data. The precipitation regions so obtained align with large scale geographical features, are readily interpretable, and are found to have superior coherence to the NCA 9-region archetype (Karl & Knight 1998).
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMGC43D1351S
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1626 Global climate models;
- GLOBAL CHANGE;
- 1942 Machine learning;
- INFORMATICS;
- 4313 Extreme events;
- NATURAL HAZARDS