1. Introduction
Precipitation is one of the most important meteorological phenomena that affect everyday life. Precise and accurate measurement of precipitation is essential for many applications. For instance, precipitation is a fundamental input for understanding the global water cycle, the management of water resources and transport infrastructures, planning agricultural operation, and predicting events with high societal impact like landslides and flooding [
1]. Precipitation can be measured by means of a large variety of different devices [
2]. Studies comparing different rainfall measurement devices and methods have highlighted the challenges in obtaining accurate precipitation estimates [
2,
3,
4]. Rain gauges are the most traditional instruments for providing local precipitation data and yield accurate and direct measurement of rain accumulation (in mm) in a given amount of time. Nowadays, the use of disdrometers in precipitation measurement is increasing thanks to their capability to provide the microphysical structure of hydrometeors in terms of drop size distribution and other properties rather than in terms of precipitation accumulation as provided by raingauges [
5]. There are several types of disdrometers. At the moment, thanks to its good trade-off between cost and accuracy of estimates, and to its ease of management, the most popular one is the laser disdrometer. It estimates, with a high temporal resolution, typically 1 min [
6], relevant characteristics, such as diameter and fall velocity of particles passing through a measurement area defined by a laser beam. Among the laser disdrometers, this study considers the Laser Precipitation Monitor (LPM) manufactured by Thies Clima GmbH, Germany. Several studies have been carried out in order to compare the performance of LPM with respect to other instruments for estimation of rainfall and microphysical parameters [
7,
8,
9] and for a wide range of applications in addition to the quantitative estimation of precipitation (e.g., the estimation of kinetic energy of rain [
10], the development of remote sensing retrieval techniques [
11], and the classification of solid precipitation [
12]). Also, laboratory experiments [
13] have been carried out to evaluate its accuracy in windy conditions. Therefore, LPM can be considered a reference instrument, well known both in terms of points of strength and limitations.
Another category of widely used disdrometers is that of optical imaging sensors. In general, they use a photographic sensor, operating in the visible band, to capture images of each hydrometeor crossing a sensing volume. An advantage of these instruments, with respect to laser disdrometers, is their ability to also provide information on the shape of hydrometeors in addition to their diameter and fall velocity. This ability has been used to derive information on the shape–size relations of drops, fundamental to develop and characterize the behavior of dual-polarization weather radar algorithms for quantitative rainfall estimation [
14,
15] and for investigating shapes of ice particles characterized by high variability.
The 2D video disdrometer (2DVD) [
16,
17], commercially available from Joanneum Research Forschungsgesellschaft mbH, Austria, has been adopted in many campaigns and has been used in research regarding the shape of hydrometeors. It uses two orthogonally oriented horizontal line-scan cameras to take measurements of shadows due to precipitation particles passing through a 10 cm × 10 cm area illuminated by two internal lamps. Since the viewing planes of the two cameras are separated by around 6–7 mm, it is possible to measure the fall speed from the time interval between arrival times of a drop in the two planes. Images of the two cameras are collected every 18 microseconds and are combined to produce a two-dimensional image of the precipitation particle for each camera with a resolution of ∼0.2195 mm.
In addition to 2DVD, two other optical disdrometers have been commonly used in the literature to investigate ice particles’ shapes, namely Precipitation Imaging Particles (PIP) and the Multi Angle Snowflakes Camera (MASC).
PIP [
18,
19] is composed of two parts: a single high-speed video camera and a light source to backlight the precipitation particles that pass through the open sampling volume. The camera, 2 m far from the light source, records 380 images of
pixels in a second with a pixel size of 0.1 mm by 0.1 mm that are then compressed to achieve an effective pixel size of 0.1 and 0.2 mm in the horizontal and vertical planes, respectively. Its focal plane is at 1.33 cm from the camera. MASC [
20] uses three CCD cameras separated by 36° in azimuth with a focal plane at 10 cm from the lens. The cameras look at a virtual sampling volume, defined by the intersection of the 35 mm fields of view and the 10 mm field depths of the cameras. A particle passes through the field of view of a pair of near-infrared sensors triggering the three cameras and the corresponding flashes opposite to the camera to illuminate the particle. The three cameras have a pixel size of 33.5 μm × 33.5 μm, although particles larger than 0.1 mm are retained [
21].
Another optical imaging gauge is the high-speed optical disdrometer (HOD), composed of a high-speed CCD camera, an LED light with a diffuser, and a digital fiber-optic sensing unit [
22]. The camera and the light are installed at a distance of 160 cm, and the sensor is installed between the camera and the light source at the focal plane of 60 cm from the camera. The sensor captures particle images at 1000 frames per second with a resolution of
pixels. The measurement volume is defined by the vertical size and horizontal camera view frame (70 mm × 70 mm). The other dimension of resolution volume is defined by the sensor beam width (5.25 mm centered around the focal plane) instead of the field depth of the camera. The sensor is also used to trigger the camera.
Practical problems of optical disdrometers are the purchase cost and maintenance difficulties and, in some cases, commercial availability. The performance of these instruments is also critically dependent on the image processing software adopted, especially if the shape of hydrometeors is the main target of research [
21].
Thies Clima has recently made available a commercial optical disdrometer, the 3D stereo disdrometer (3DS). The purpose of this article is to evaluate the 3DS, which, as far as we know, has never been used for scientific purposes before, by comparing it with LPM, the rather popular laser disdrometer from the same manufacturer. The most innovative feature of the new disdrometer lies in capturing images of particles passing through the measurement volume defined by cameras, while a laser disdrometer measures the size and fall velocity of a raindrop that, passing through a surface defined by a laser beam, interrupts the light and causes a reduction in the intensity of the received signal. Imaging capability of 3DS is of crucial importance in providing an accurate classification of hydrometeors, especially in the case of solid precipitation. The 3DS disdrometer is expected to be helpful both regarding rain and snowfall rate, mitigating the difficulties related to the microphysical variability of solid hydrometeors [
23]. The study reported in this article aims to highlight the difference in performance of the two Thies Clima disdrometers (i.e., LPM and 3DS) in terms of hydrometeor classification, spectrograms, number of particles, particle fall speed, and rain rate. Given the different measurement principles, differences in performance are expected between the two disdrometers.
This article is organized as follows.
Section 2 deals with the description of the main characteristics of the two disdrometers.
Section 3 explains the experimental datasets used in the study, and
Section 4 describes the comparison and validation approach.
Section 5 presents the main results obtained by comparing the 3DS and LPM precipitation outputs. The imaging capabilities of 3DS are also discussed in this section. Finally,
Section 6 points out the main findings, which are summarized and commented on.
4. Comparison and Validation Approach
The comparison between LPM and 3DS is carried out based on the spectrograms of diameters and fall velocities, the precipitation classification, and the rain rate provided by the Thies Clima software, ver. 1.1220, which are part of the considered 1 min data telegrams. The 1 min spectrograms are expressed through matrices
where superscript
X identifies the disdrometer and can be LPM or 3DSD.
and
are the number of diameter and fall velocity bins, respectively, that are different for the two disdrometers (see
Table A1 and
Table A2).
Drop size distributions, which are maybe the most important expected output for hydrological applications and for developing remote sensing algorithms, are not provided by the manufacturers of disdrometers but are derived by users through additional processing that can imply filtering of spurious hydrometeors, e.g., using, a filter based on a priori fixed mask (such as in [
4,
28,
29]) or chosen adaptively (as in [
30]). In addition, LPM data are sometimes corrected using other reference instruments as in [
12].
Although evaluation of DSD retrieval methods is beyond the scope of this paper, some implications of LPM and 3DS measurements for DSD retrievals are discussed. Particular attention is focused on the fall velocity of hydrometeors. In rain, the fall velocity diameter relations from [
24] experiments, confirmed also by more recent experiments [
14], can be considered, on average, as representative of the fall behavior of raindrops at least for diameters smaller than 6 mm, although an example of noticeable deviation is reported in [
31]. For larger drops, a slightly decreasing velocity with respect to the fit to [
24] was found [
14]. In this article, in rain, a formula based on [
24], such as [
32]
can be used for reference at a height of 0 m above sea level (ASL), while modification for a generic height
h (in m) to take into account air density is according to [
33]
where
and
in kg·m
−3 are the air density at sea level and at height
h, respectively, that can be assumed according to the International Standard Atmosphere Model [
34]. Formula (
3) will be used in this study, with
m corresponding to the altitude ASL of Casale Calore, and it will be referred to as GK.
Ice particles can take different shape, density, and habits that make their velocity–size relation quite variable, although several laws are available for some reference hydrometeors. This fact allows for a hydrometeor classification based on diameter–velocity pairs [
12], and an instrumental bias in velocity estimation will result in classification error.
The second part of validation deals with the evaluation of the detection capability of 3DS with respect to LPM and the comparison of the total accumulated precipitation per event (in mm) between the two disdrometers in order to point out the performance of 3DS in providing accurate rainfall measurements. This part concerns rain events only.
The detection capability is analyzed by means of contingency tables together with POD (Probability of Detection), FAR (False Alarm Ratio), and ACC (Accuracy) indices. Each contingency table contains the following parameters: the number of ‘hits’ minutes , when precipitation is detected by both the 3DS and the LPM; the number of ‘false alarm’ minutes , when precipitation is detected by the 3DS but not by the LPM; the number of ‘missing’ minutes , when precipitation is detected by the LPM but not by the 3DS; and the number of ‘reject’ minutes , when both disdrometers do not detect precipitation. POD, FAR, and ACC are statistical indices that6 can be computed from contingency tables. The POD index measures the probability that 3DS correctly detects precipitation when precipitation is actually present (i.e., detected by the LPM). The FAR index is calculated as the ratio between the number of no-rain minutes wrongly categorized as rain minutes (false positives) and the total number of actual no-rain minutes.
The ACC index shows how close measurements by 3DS are to the LPM:
The rainfall accumulation per event is computed by considering the value provided by the Thies Clima software for all the rainy minutes between the beginning and the end of the event itself. Then, a quantitative comparison is conducted by calculating the correlation coefficient, RMSE (Root Mean Square Error), NMAE (Normalized Mean Absolute Error), and NB (Normalized Bias). The correlation coefficient measures the degree to which the two sets of data are linearly related. It assumes values in the range from
to
, where
indicates the strongest possible agreement and 0 the strongest possible disagreement. Root Mean Square Error allows to evaluate how concentrated the data are around the line of best fit,
where
Y is the vector representing 3DS data and
X is the vector of LPM data. The Normalized Mean Absolute Error is a normalization of the mean absolute error (MAE):
N being the sample size. The Normalized Bias provides information on the quality of the difference of the two datasets: negative NB values indicate an underestimation of 3DS with respect to LPM, while positive NB indicates overestimation:
A qualitative analysis of 3DS images is finally conducted for liquid and solid hydrometeors. It should be noted that dependency of LPM measurements from horizontal wind has been noticed in experiments [
35] and justified by the Computational Fluid Dynamics simulations [
13], pointing out the possible dependency of the performance LPM from wind intensity and also by direction because of the non-symmetric shape of the instrument. Such effects are expected to be more pronounced in light rain and in snow. In fact, for snow, some authors prefer not to rely on LPM measurement if horizontal wind exceeds a given threshold (it is 7 m·s
−1 in [
36]). The geometry of 3DS, also non-symmetric, enables thinking about the influence of horizontal wind, although it is expected to be different from that of LPM. In this paper, we do not consider the likely different effect of wind on the two devices, which deserves different analysis methods, such as laboratory experiments.
6. Conclusions
Imaging disdrometers have become quite popular in field campaigns, but, also due to their price, higher than that of laser disdrometers, their application is limited to research environments. Thies Clima has applied the concept of imaging disdrometers to a more affordable off-the-shelf product, named 3D stereo disdrometer. This paper has compared this instrument with the laser disdrometer of the same manufacturer (the Laser Precipitation Monitor). LPM can be considered a known instrument, being used in many research papers, while 3DS is a new device and, in the literature, there are few papers that deal with its data. The data used for the comparison were collected at a site in the Appennine range in Central Italy and report mostly rain precipitation. The two instruments are compared based only on outputs provided by the software of the manufacturer. In general, the 3DS detects more particles than the LPM, especially those with smaller diameters. This behavior is probably due to the different sampling modes of the two disdrometers and is different depending on diameters of drops. The 3DS in rain reports a lower percentage of superterminal drops than LPM. Both the disdrometers are in agreement with the reference size–velocity relation for rain drops with diameter less than 3 mm, while, for larger particles, underestimation of fall velocity occurs. The underestimation seems to be more pronounced for LPM, and the variability among the estimates is more spread around the median value. The two disdrometers agree quite well with each other in terms of rain rate and accumulated precipitation, although 3DS underestimates high rain rates with respect to LPM. The imaging capabilities of 3DS are of high interest in particular for identification and characterization of snow, ice, and mixed precipitation. However, to make such images useful for research, an improvement in the instruments is recommended. In fact, for the moment, the software provides the user with only a limited sample (from one to four particles per minute) of rather low-resolution images (12 × 12 pixel matrix) regardless of the dimensions of the hydrometeors). More images that show all (or a significant portion of) the particles that pass through the measuring area each minute can be useful and of high interest for precipitation microphysics-related studies. Improvement regarding the resolution of the images made available to the user is also recommended.