# Outlier Detection In Depth Of Snow Data

Article

## College

College of Science

## Presentation Type

Oral Presentation

## Abstract

In many places in the United States, buildings need to be built to withstand extreme snow events, without making the construction overly expensive. Often, fitting a probability distribution based on annual maximum measurements of snow depth (or snow water equivalent) will be used in an extreme value analysis. Because the maximum annual snow depths are used in fitting the probability distributions, it is crucial that those maximum values are legitimate. Manually searching through about 100 different snow measurement locations scattered in four different states, there are four patterns that large, but legitimate, maximum snow values tend to follow. The patterns are characterized by the degree of build up to, or build down from, the maximum observation. We use these patterns as part of a two-step filtering process. The first step of the filter naively flags potential outliers. The second step then looks through each potential outlier and compares the set of days around the max to the four patterns previously identified. Any potential outliers that closely follow one of the four patterns are not thrown out, but those that do not follow any pattern are removed. This method of outlier removal protects the probability distribution fitting process from anomalous high values while still ensuring that buildings are designed to withstand true, extreme snow load events.

## Start Date

4-9-2020 12:00 PM

4-9-2020 1:00 PM

## Share

COinS

Apr 9th, 12:00 PM Apr 9th, 1:00 PM

Outlier Detection In Depth Of Snow Data

In many places in the United States, buildings need to be built to withstand extreme snow events, without making the construction overly expensive. Often, fitting a probability distribution based on annual maximum measurements of snow depth (or snow water equivalent) will be used in an extreme value analysis. Because the maximum annual snow depths are used in fitting the probability distributions, it is crucial that those maximum values are legitimate. Manually searching through about 100 different snow measurement locations scattered in four different states, there are four patterns that large, but legitimate, maximum snow values tend to follow. The patterns are characterized by the degree of build up to, or build down from, the maximum observation. We use these patterns as part of a two-step filtering process. The first step of the filter naively flags potential outliers. The second step then looks through each potential outlier and compares the set of days around the max to the four patterns previously identified. Any potential outliers that closely follow one of the four patterns are not thrown out, but those that do not follow any pattern are removed. This method of outlier removal protects the probability distribution fitting process from anomalous high values while still ensuring that buildings are designed to withstand true, extreme snow load events.