IAS METHOD FOR 2D DATA ANALYSIS ON PC's

CHAPTER 2: PARTICLE SHADOW CLASSIFICATION

The IAS TOUCH2D package is based on the NCAR TOUCH2D package developed by Joanne Parrish (now Joanne George) and Andy Heymsfield. The NCAR TOUCH2D package is undergoing continual evolution, but the basic parameters and techniques it employs to classify particles are described in Heymsfield and Parrish (1979). The IAS TOUCH2D package also uses a few additional measures and techniques described in Holroyd (1987), Gordon and Marwitz (1984), and helpful suggestions from W.A. Cooper (Research Aviation Facility, NCAR).

The NCAR TOUCH2D package is an example of a semi-automated scheme. A preliminary classification is assigned by a decision-tree-like computer algorithm. Then a human analyst verifies or alters the classification. The NCAR operator worked at a touch screen computer terminal displaying touch-selectable images and menus of commands and classifications. Hence came the name TOUCH2D.

Since touch-screen terminals are not available at the IAS, the NCAR software package was rewritten to be run from a monitor using cursor and mouse movements to select particles and menu items rather than the touch of a finger. The basic procedure is the same, however. Particles are sized and classified by a decision tree algorithm. They can then be displayed on a graphics screen and their size or classification altered manually.

Many classification schemes, including the original NCAR TOUCH2D, classify particle shadows into more or less specific hydrometeor categories, such as graupel, needle, dendrite, etc. The IAS TOUCH2D is less ambitious. It attempts to classify shadowgraphs into categories of increasing geometric irregularity. Since the T-28 penetrates a variety of cloud environments, particle shadows with a given degree of irregularity may correspond to different hydrometeor types in different environments. The scientist using the package must use his/her own insight into cloud microphysics to judge actual hydrometeor types from TOUCH2D classifications. There is a certain unremovable ambiguity in 2D shadowgraphs. Even two "experts" will concur in the classification of a typical image only about two-thirds of the time (Dyer et al., 1985). Therefore, a greater degree of sophistication in the IAS TOUCH2D was not sought.

The following measures are computed by TOUCH2D for use in sizing and classifying particles.

1. Heymsfield diameter: This is the diameter computed as in the original NCAR version. It is the maximum dimension of the shadow in either the vertical or horizontal direction. One diode is added to the horizontal dimension to account for the time needed to trigger the probe. For relatively regular particles that are only partially in the field of view, the particle is reconstructed making simple geometric assumptions and its "true" diameter estimated (see Heymsfield and Parrish, 1979).
2. Holroyd diameter: This measure is taken from Holroyd (1987). A least squares line is fitted through the pattern of occluded diodes and the maximum span along this line is taken to be the maximum dimension. The Holroyd diameter is usually very close to the Heymsfield diameter. If it is not, then the image is probably ambiguous and deserves extra attention.
3. Equivalent circle diameter: Diameter of a circle with the same area as the shadow.
4. Area ratio: This measure (from the original TOUCH2D) is the ratio of the observed number of shadowed diodes to the number that would be shadowed by a spherical particle with the Heymsfield diameter. Nearly spherical particles will have area ratios near 1, while very irregular particles will have area ratios of a few tenths or less.
5. Linearity: This is just the correlation coefficient from the least-squares procedure used to determine the Holroyd diameter.
6. Angle: The slope from the least squares fit.
7. Width: The standard deviation from the least squares fit.
8. Hole: The particle shadow contains a central, completely surrounded region of one or more unshadowed diodes, or it is an oval-shaped particle whose center-of-mass is located at an unshadowed diode.
9. Oval: An "oval" particle is one with a perimeter that is monotonically convex.
10. PDA: This is another measure of shadowgraph irregularity, taken from Holroyd (1987). It is the product of the number of shadowed particles on the perimeter, and the diameter, divided by the total number of shadowed diodes.
11. CHI: This measure of particle roundness is an adaptation of a technique suggested by Cooper. The position of the center of the particle shadow is estimated, and the distance from this center to each perimeter element of the shadow is computed. The variance of these radii is what is called CHI. The process is iterated to minimize CHI. The final CHI will be small for nearly round shadows and large for irregular shadows.
12. Mass: The estimation of particle mass from shadowgraphs is a highly uncertain game. The original NCAR TOUCH2D used empirically-derived power law relationships in which particle mass was given as a function of maximum dimension. Different relationships were used for different particle habits. Holroyd (1987) used an analogous procedure, but a different set of relationships. The IAS TOUCH2D package allows the user to select from among several alternatives, including these.

There was a general tendency observed in these studies for the effective density of these particles to decrease with increasing size. That is, the mass of the particle became a smaller and smaller fraction of the mass of a liquid sphere with a diameter equal to the particle maximum dimension as particle size increased. An example from one such investigation is shown in Fig. 3, taken from Locatelli and Hobbs (1974).

Most T-28 2D-C particle images are less than 1 mm in diameter. Tests with early versions of the IAS TOUCH2D showed that the power law mass relationships used by the NCAR TOUCH2D and by Holroyd (1987) seemed to give very low effective densities for sub-millimeter particles, much lower than might be expected from an extrapolation based on Fig. 3 to smaller sizes.

Fig. 3: Density versus diameter for lump graupel. The height of the shaded region is the average density of the lump graupel in the given diameter range.

In addition, these relationships required that particle images be classified into hydrometeor types in order to choose the correct power law for mass estimation, and the IAS TOUCH2D was evolving away from doing this.

A heuristic mass estimation technique was developed for the IAS TOUCH2D, based on particle size and shadow area ratio. The mass of a particle is estimated as the mass of a liquid sphere with diameter equal to particle maximum dimension, multiplied by area ratio raised to the 3/2 power. The motivation for this form was that area ratio represents the irregularity of the particle projected onto two dimensions, and that area ratio to the 3/2 power is a heuristic estimate of the three-dimensional irregularity of the particle. This three-dimensional irregularity becomes an effective density. When multiplied by the mass of a regular particle (i.e., a sphere), a reasonable estimate of irregular particle masses is derived.

There is no rigorous theoretical justification for this procedure. Limited comparisons show that masses estimated using this technique are within the scatter of those predicted by the appropriate published power law relationships employed by NCAR TOUCH2D and by Holroyd (1987) for particles larger than 1 mm, and somewhat larger for particles smaller than 1 mm, as was desired.

It should be noted that the accuracies generally claimed for mass loading estimations based on a sequence of 2D shadows are generally 2x [i.e., uncertainties of the order of a factor of 2; e.g., Holroyd (1987)]. Higher accuracies would not be expected based on the scatter shown in empirical studies of particle masses like those of Locatelli and Hobbs (1974). The uncertainty in mass loading estimates will go down as more and more particles are grouped for each mass loading estimate. More quantitative descriptions of this uncertainty and how it depends on particle number and sample volume are given in Detwiler et al. (1993).

Most of these measures can be displayed with the images. An example of the default TOUCH2D image display format is shown in Fig. 4.

Fig. 4: Standard TOUCH2D image display format. Lines 3-5 in the upper left corner describe the alphanumeric data which follow. Each image is immediately below its numeric data.

The measures described above are used in a decision tree algorithm to classify particle shadows into one of 12 categories. Five of these categories are considered artifact categories. Images classified into any one of the artifact categories are not likely to represent actual cloud hydrometeors.

NOTE: The examples listed on Figs. 5-13 were produced by an older version of the IAS TOUCH2D programs. Although somewhat different from the current output format, they provide examples of habit types.

The five artifact categories are:

1. Zero times: These are particles for which the elapsed time from the preceding particle is 0. A zero elapsed time is an indication that probe is not functioning properly. Zero time particles are almost always also zero-element ("zero-Y") particles, for which the probe was triggered but no shadowed diodes were found on the following scan of the array (see below).
2. Broken: These shadows consist of two or more separated parts. If these represent actual particles in the cloud observed with a properly-functioning probe, the decision tree algorithm will fail to properly classify them. If, as often happens, some bits get stuck on or off in the image buffer and rearrange the bit pattern in the end-of-particle slices, TOUCH2D will not be able to tell where one shadow frame ends and the next begins. Several separate particles may end up combined together in one long frame and this composite "particle" will be classified as "broken". See Fig. 5.

 

 


Fig. 5: Examples of "broken" particle shadowgraphs from a 2D-C probe.
 

3. Streakers: Water collected on probe arm tips will occasionally separate off into streamers and droplets that break loose and fall back through the sample volume. They will not have decelerated to rest by the time they pass the diode array and so will be moving more slowly relative to the probe (attached to the aircraft moving ~100 m s^-1 through the cloud) than the true cloud hydrometeors. This produces elongated "streaker" shadows. An example is shown in Fig. 6. The IAS TOUCH2D classes a shadow as a streaker if its horizontal dimension is more than six times its vertical dimension. This follows a rule-of-thumb developed from years of experience at the University of Wyoming.

 

 


Fig. 6: Examples of 2D-C "streakers". The "streaker" decision is made before the "hole" decision, so streakers with holes will be classed as streakers.
 

4. Holes: Small droplets streaming from the probe tips will occasionally be spherical enough to be considered as possibly true hydrometeors. However, they will also be close to one probe arm or the other and generally slightly out of focus. The shadows of such particles will have bright spots near their centers. See Fig. 7. Particles classed in this category are likely, then, to be shed droplets and not true cloud hydrometeors.

 

 



Fig. 7: Examples of "hole" images from a 2D-C.
 

5. Out-of-view: Shadows with length greater than five times height and occluding one or the other end element typically represent partially shadowed very large particles. There is no reliable way to reconstruct "true" size from such a small portion of shadow, so these shadows are classed into an artifact category. The current TOUCH2D logic will generally assign particles to the "streaker" category rather than "out-of-view".

 

 

The remaining shadow categories can usually be presumed to represent actual cloud hydrometeors. They are:
 

6. Zero-Y: The probe stays in a wait state until one or more diodes are shadowed. It then begins to rapidly scan the diode array to build up a shadow image of the passing particle. It is possible for a small particle to be just big enough to trigger the probe, but to be entirely past the diode array by the time of the next array scan. The size to assign to such a particle is uncertain. TOUCH2D assigns it a diameter equal to the effective size of one diode (nominally 25 µ m for the T-28 2D-C).

 

 

In most instances, zero-element particles, if they have valid elapsed times, are probably actual cloud particles. They may be large cloud droplets or small ice particles in the case of the 2D-C. Unfortunately, it is not possible to use 2D-C data to distinguish between these two types of particles. in the case of 2D-P data, the zero-element particles are on the order of 200 µ m diameter, and may be either liquid or ice.

Comparison between data obtained in overlapping size ranges of FSSP, 2D-C, and 2D-P probes mounted on the same aircraft shows that the 2D probes usually underestimate concentrations in their respective zero-element size categories (e.g., Gordon and Marwitz, 1984).

7. Tiny: Particles less than 4 diodes in diameter for the 2D-C, or 3 diodes in diameter for the 2D-P, are classed as "tiny". These shadows have such low relative resolution that very little can be said about their shapes. See Fig. 8.

 

 


Fig. 8: Examples of "tiny" shadowgraphs from a 2D-P.

8. Circular: These are almost perfectly circular shadows (raindrops, most frequently). See Fig. 9.



Fig. 9: Examples of "circular" images from a 2D-C.

9. Nearly-circular: Not quite perfectly circular (raindrops or graupel). See Fig. 10.

 

 


Fig. 10: Examples of "nearly circular" images from a 2D-P.

10. Regular: Generally convex, but maybe a little rough or non-circular. (Graupel might often fall in this category.) See Fig. 11.

 

 


Fig. 11: Examples of "regular" shadowgraphs from a 2D-C (top) and a 2D-P (bottom).
 

11. Irregular: Snowflakes, like dendrites and spatials, aggregates, etc., often end up in this category. See Fig. 12.



Fig. 12: Examples of "irregular" images from a 2D-C.

12. Linear: Columns and needles, typically. See Fig. 13.

 

 


Fig. 13: Examples of "linear" particle images from a 2D-P (upper left) and a 2D-C (all others).

The decision tree algorithm for non-artifact images, as of the date on this bulletin, is as follows:

FOR UNRECONSTRUCTED SHADOWS

If (Area Ratio 0.75 and PDA 5.5) then
if (OVAL and CHI < 0.1)
HABIT = CIRCULAR
else
HABIT = NEARLY CIRCULAR
else
if (Area Ratio 0.5 and PDA 10) then
HABIT = REGULAR
else
if (least squares correlation < 0.4) then
HABIT = IRREGULAR
else
HABIT = LINEAR

FOR RECONSTRUCTED SHADOWS

If (Area Ratio 0.95) then
HABIT = CIRCULAR
else
if (Area Ratio 0.75) then
HABIT = NEARLY CIRCULAR
else
HABIT = REGULAR

The cutoffs and criteria utilized in this decision tree are entirely empirical, based on experience gained processing a limited range of cloud penetration data from the T-28 2D-C and a very limited range of T-28 2D-P data taken in Alabama thunderstorms in 1986. It is entirely possible that any individual user might wish to alter this scheme. Appropriate changes can be made in subroutine ANDY_HABIT (file ANDY_HAB.C) within the TOUCH2D package.

Introduction

Chapter 1 - Discussion of Probes

Chapter 3 - PRELIM2D Program

Chapter 4 - TOUCH2D Program

Chapter 5 - ANALYZE and ANELAP

References