
Variability of Drop Size Distributions: Noise and Noise Filtering in Disdrometric Data
Author(s) -
GyuWon Lee,
Isztar Zawadzki
Publication year - 2005
Publication title -
journal of applied meteorology
Language(s) - English
Resource type - Journals
eISSN - 1520-0450
pISSN - 0894-8763
DOI - 10.1175/jam2222.1
Subject(s) - spurious relationship , statistics , mathematics , regression , noise (video) , linear regression , robust regression , regression analysis , simple linear regression , sample size determination , computer science , artificial intelligence , image (mathematics)
Disdrometric measurements are affected by the spurious variability due to drop sorting, small sampling volume, and instrumental noise. As a result, analysis methods that use least squares regression to derive rainfall rate–radar reflectivity (R–Z) relationships or studies of drop size distributions can lead to erroneous conclusions. This paper explores the importance of this variability and develops a new approach, referred to as the sequential intensity filtering technique (SIFT), that minimizes the effect of the spurious variability on disdrometric data. A simple correction for drop sorting in stratiform rain illustrates that it generates a significant amount of spurious variability and is prominent in small drops. SIFT filters out this spurious variability while maintaining the physical variability, as evidenced by stable R–Z relationships that are independent of averaging size and by a drastic decrease of the scatter in R–Z plots. The presence of scatter causes various regression methods to yield different best-fitted R–Z equations, depending on whether the errors on R or Z are minimized. The weighted total least squares (WTLS) solves this problem by taking into account errors in both R and Z and provides the appropriate coefficient and exponent of Z = aRb. For example, with a simple R versus Z least squares regression, there is an average fractional difference in a and b of Z = aRb of 17% and 14%, respectively, when compared with those derived using WTLS. With Z versus R regression, the average fractional difference in a and b is 19% and 12%, respectively. This uncertainty in the R–Z parameters may explain 40% of the “natural variability” claimed in the literature but becomes negligible after applying SIFT, regardless of the regression methods used.