Bayesian Classification of Nonmeteorological Targets in Polarimetric Doppler Radar Measurements
This article is licensed under a 1. Introduction
Since the early history of weather radars the main target of measurements has been precipitating weather systems (see, e.g., Atlas 1990; Fabry 2015). In that context it is desirable that quantitative radar-based precipitation estimates (QPE) or forecasts (QPF) are free from major systematic bias or random errors. Thus, all sources of microwave scattering other than precipitation have conventionally been seen as undesirable disturbance called clutter (Battan 1973), and accordingly, the objective of studies on nonmeteorological targets has been bulk identification of clutter for the purpose of rejecting it from radar observations.
While relatively stationary ground clutter can be eliminated from the observations with Doppler filtering, a more challenging source of error in radar-based QPE is the presence of moving scatterers, like sea waves (sea clutter), wind turbines, ships, aircraft, military chaff, debris, birds, and even arthropods. In addition, there are commonly occurring sources of microwave signals that do not represent any material scatterers in the measured range gate sample of the radar. Such sources include external emitters (sun, other radars, microwave links, military jamming) and second-trip echoes (sometimes called “ghost echoes”; see, e.g., Battaglia et al. 2016) from targets (usually thunderstorms) outside the unambiguous measurement range.
More recently, there has been a gradual shift toward seeing the nonmeteorological component of weather radars observations as a valuable source of independent data. For example, nonpolarimetric algorithms separating representative wind tracers (precipitation and arthropods) from nonrepresentative ones such as birds have been developed for applications in vertical wind profile studies (Gauthreaux et al. 1998; Holleman et al. 2008). The aeroecological community has developed nonpolarimetric classification methods facilitating continentwide ecological research of bird migration (e.g., Dokter et al. 2011; Kelly et al. 2012; Dokter et al. 2018; Weisshaupt et al. 2018; Nilsson et al. 2019; Rosenberg et al. 2019; Kranstauber et al. 2020). All these studies classify each range ring of a PPI sweep and thus, each vertical layer, either to birds or nonbirds. Classification performance of these methods is challenged by spatially heterogeneous movement and cases where other scatterers like precipitation or insects occupy portions of the volume scan. Thus, there is an urgent need for higher spatial resolution and methodological resilience in various meteorological conditions.
The latest well-established generation, i.e., polarimetric weather radars, expands the capabilities of Doppler radars by utilizing horizontal and vertical linear polarization of the transmitted and received signals, thus providing the conventional as well as polarimetric moments of the scattering media (Table 1). The richer set of measured quantities enables better recognition of all types of scattering media, provided that the statistical properties of all target classes are known to a sufficient degree.
The radar quantities (moments) used in this study. The name and the physical unit with its expected measurement range are given for each moment, with additional comments where pertinent.
So far the primary focus of research on polarimetric data has been the separation of different hydrometeor types for increased accuracy in QPE and other meteorological applications (e.g., Lim et al. 2005; Park et al. 2009; Dolan et al. 2013; Chandrasekar et al. 2013). Likewise, QPE-related studies have presented polarimetric algorithms for more comprehensive detection and removal of nonmeteorological phenomena beyond ground clutter, such as sea clutter, chaff, birds, and insects (e.g., Zrnić and Ryzhkov 2004; Ryzhkov et al. 2005; Gourley et al. 2007; Chanthavong et al. 2010; Lakshmanan et al. 2014; Ye et al. 2015; Kilambi et al. 2018).
Some aeroecological studies on polarimetric weather radars have been conducted, including characteristic properties of birds and insects (e.g., Melnikov et al. 2015; Stepanian and Horton 2015). Jiang et al. (2013) have created a fuzzy logic polarimetric algorithm that separates birds from insects at each range gate sample in S-band radars, for the purpose of eliminating birds from radar-based wind measurements. However, their method depends on conditions that are not always valid in Finland, i.e., that insects cannot be the dominant biological scatterers at local midnight, and that Zdr of insects must be positive [for the latter, see section 2b and Melnikov et al. (2015)]. Radhakrishna et al. (2019) have tested nonpolarimetric and polarimetric fuzzy logic algorithms for pixel-by-pixel identification of birds, precipitation, and ground clutter in two selected cases. They found that polarimetric algorithms (membership functions) are more efficient in identification than nonpolarimetric ones. Machine learning approach has been used to investigate how the characteristics of six specific subgroups of bioscatterers differ from each other and from precipitation (Gauthreaux and Diehl 2020). Stepanian et al. (2016) stated that the adoption of polarimetry into well-established and quantitative biological applications is still a work in progress. Most recently Jatau et al. (2021) have created a machine learning algorithm that can identify birds and insects at each bin of S-band radars (NEXRAD), provided that a preceding classification first delineates nonmeteorological echoes from hydrometeors, as Kilambi et al. (2018) describe. The approach does not classify all common nonmeteorological target types by the same system as in the present paper, though, which is a required property for multipurpose classification product generation.
To address the shortcomings described above, we present a unified, data-driven, supervised Bayesian probabilistic method with fine-grained division of nonmeteorological classes for gate-by-gate classification of targets in polarimetric C-band radar observations. In addition to the measured moment values for a range gate the method takes into account patterns of values in the immediate neighborhood of the gate as well as the tentatively assigned class probabilities of adjacent gates, enabling the operational classifier to consistently resolve 7 meteorological and 12 nonmeteorological target classes.
The presented method enables efficient utilization of the full set of polarimetric weather radar measurements for nonmeteorological targets. Besides providing a direct estimate of classification uncertainties, the approach also enables agile product generation because any number of customized products can be derived post facto from a single analysis, in accordance with a paradigm shift in the approach for radar data processing where one user’s refuse is another user’s treasure.
2. Radar measurements
The present operational weather radar network in Finland consists of 11 polarimetric C-band systems that are maintained and owned by the Finnish Meteorological Institute (FMI). The first three polarimetric radars were implemented during the years 2009–10. Since then all radars have been equipped with Vaisala/Sigmet Inc. RVP model signal processors and their respective IRIS software (Vaisala 2014, 2017). FMI has archived all radar network measurements in the past 20 years.
In this study the training data for classification were manually selected from operational archived radar scans, as subsegments of full PPI sweeps that consist of a 360° antenna rotation at a constant elevation angle. Typical elevation angles used were the lowest one in the range 0.3°–0.5°, and the next few angles, typically 0.7°, 1.5°, and 3°. The dataset for training is described in detail in section 2b. Each PPI sweep consists of 360 azimuth angle measurements of 500 range gate samples, also called bins, between the antenna and the outer range of 250 km.
Measured data contain static metadata for each sweep, followed by the measured values of the quantities, commonly called moments, at each bin (Table 1). In addition to these moments the system also provides the HydroClass hydrometeor classification product that is not used as an input to the classifier of this study. For a more detailed description of the measuring process we refer to Vaisala (2014, 2017).
a. Principles of training and class definitions
The primary objective of this study is the classification of nonmeteorological echoes in single range gate samples, since methods for fine-grained classification of echo types listed in Tables 2–6 are only available for hydrometeors. The reason or this choice of scope is that numerous studies have compared polarimetric radar measurements to in situ observations of hydrometeors and hence the polarimetric scattering characteristics of precipitation types are already rather well known (e.g., Ryzhkov et al. 2005; Dolan et al. 2013; Bechini and Chandrasekar 2015; Grazioli et al. 2015; Roberto et al. 2017). In our training dataset precipitation classes were defined based on such a priori knowledge of polarimetric moments as well as on ground-based observations of temperature, dewpoint, and precipitation type from automatic weather stations (AWS). However, our estimates for those classes should be considered roughly indicative only since continuous recording of the size and water phase of hydrometeors was not available for any of the radars, and since most in situ measurements (AWS and synoptic observations) were located well below the lowest beam.
Meteorological target classes defined in the training set. Class index i is used in this paper to denote the respective class where the usage of the full name would be cumbersome.
Terrain target classes defined in the training set. Class index i is used in this paper to denote the respective class where the usage of the full name would be cumbersome.
Zoogenic target classes defined in the training set. Class index i is used in this paper to denote the respective class where the usage of the full name would be cumbersome.
Anthropogenic target classes defined in the training set. Class index i is used in this paper to denote the respective class where the usage of the full name would be cumbersome.
Immaterial target classes defined in the training set. Class index i is used in this paper to denote the respective class where the usage of the full name would be cumbersome.
It should be pointed out that we classify meteorological systems primarily by the type of scattering medium instead of their dynamical characteristics. For example, a sea breeze front is not considered a class despite being present in the training dataset; instead, it is classified as “insects” or “birds” since those are the scatterers making the front detectable (Puhakka et al. 1986). However, this rule could not be applied consistently for hydrometeors which were divided into two subtypes according to the appearance of the precipitation pattern and intensity in the radar image, i.e., widespread and convective precipitation for two water phase types, wet precipitation (i.e., rain and partly melted snow), and snow (i.e., dry solid precipitation). Subjective division into convective and nonconvective precipitation is necessarily fuzzy, especially in cases of embedded convection. However, any statistical methodology would have required considerable effort (e.g., Wang et al. 2021). Simple rules like the 40 dBZ threshold (Steiner et al. 1995) do not work in the cool climate of Finland where large majority of convective cells never reach values of 40 dBZ or more.
We include a full set of precipitation classes in the training set because the available fuzzy logic meteorological classifiers do not provide the kind of probability distributions that are required by Bayesian inference. While the presence of precipitation in the training set is required for completeness, the above mentioned issues suggest that methods based on in situ measurements should be preferred over the method in this paper when classifying precipitation types, with the caveat that regularly occurring shallow precipitation phenomena in Finland, snow.bl and drizzle in Table 2, are categories not covered by other studies. Because of weak Ze and similar υ and echo patterns they are inseparable from birds and insects in nonpolarimetric data (most insects and precipitation are passive tracers, and all exhibit annular patterns). They are also not available as subclasses of the more general categories rain and snow in the existing polarimetric hydrometeor classifiers, e.g., the Vaisala’s HydroClass system. Thus, it would be preferable to use the method in this paper to separate insects and birds from very weak precipitation. For other precipitation types existing hydrometeor-specific polarimetric classification schemes should take precedence.
The classes used in the training set and short characterizations of each class are given in Tables 2–6.
b. Training set
Selection of each training sample was based on visual inspection of all the PPI moments listed in Table 1 and applying the class criteria given in Tables 2–6. A region (annulus sector) of a PPI image was selected to be a training case if visual assessment suggested that it contained only echoes belonging to a single class. In the selection of regions of single-type hydrometeors published polarimetric characteristics of hydrometeor types, temperature at ground level and vertical temperature soundings from the nearby sounding stations Jokioinen and Tallinn (WMO numbers 02963 and 20638, respectively) were also used. For the identification and separation of insects from other echo sources, we consulted wind and temperature data from radiosondes near the respective radar sites, with maximal temporal offset of 6 h and typically less than an hour. The selection of observation dates was in principle random, though the selected cases of a single scatterer type were not completely independent when temporally proximate observations were chosen. Still, several samples from different radars, elevation angles, and successive scans were typically picked when a representative case was found.
A graphical user interface was implemented to facilitate case selection. The interface provides a convenient tool for browsing radar volumes stored in the FMI archive, plotting PPI scans of all moments listed in Table 1 and defining the training dataset which consists of annulus sector selections and their respective metadata describing their contents and possibly setting additional selection constraints for individual quantities (e.g., Ze > 20 dBZ and Zdr > 0.5 dB).
An exemplary radar image used for constructing the training dataset is shown in Fig. 1, displaying the moments Ze, υ, Zdr, and Φdp at the elevation angle 0.3°. The radar is located in a sector of warm air mass (10°C temperature at 850 hPa) where the airflow in the boundary layer arrives approximately from south-southeast (SSE). This time of the year and the synoptic situation are typical for massive invasions of insects, especially aphids, in Finland (Nieminen et al. 2000; Leskinen et al. 2011). The warm period is short, since a cold front is approaching from west at a distance of approximately 150 km. As can be seen in Ze the front generates patches of convective rain with Ze up to 40–50 dBZ. Accordingly, a training data selection is chosen SW of the radar and marked as belonging to the class “wet.shower,” with an additional constraint that only bins with Ze > 15 dBZ constitute the class sample. In Φdp we can see two sectors of values significantly larger than in the neighboring azimuth angles (in SSW and NW directions). These sectors lie behind cells with high Ze, which indicates enhanced attenuation, i.e., the class “prec.attenuated.” This categorization is further supported by Zdr that exhibits negative values in the same sectors (see, e.g., Bringi et al. 2001).
Training interface displaying a multiple selection case for Anjalankoski radar at 1415 UTC 17 May 2012, with the radar moments (a) Ze, (b) υ, (c) Zdr, and (d) Φdp shown, and four representative class sample sectors selected.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
In Ze and Zdr a wide eastward region of weak echoes with variable negative and positive values of Zdr (in the approximate range of −2 to 4 dB) can be seen. The variation correlates with the azimuth angle, being on average smallest in SSE and largest in east-northeast (ENE). This supports the assessment that the weak echo consists of migrating insects whose bodies are commonly oriented along the SSE to north-northwest (NNW) axis (Melnikov et al. 2015). The hypothesis is further supported by the following observations:
-
In Ze a part of the echo consists of narrow maxima oriented along the SSE–NNW direction. Such structures are typical in a warm sunny boundary layer (as is in this case east of the radar) and consist of convective roll vortices (or finelines) where insects are converging with the wind into the elongated lines of updrafts (e.g., Puhakka and Saarikivi 1986; Wainwright et al. 2017).
-
The radial Doppler velocity image shows that the υ field is relatively smooth in the texture, i.e., apart from the rather densely occurring jumps across the Nyquist frequency (±7.6 m s−1) the pattern is not noisy in the individual bin neighborhood as is the case with birds (Holleman et al. 2008). It is also known that no major songbird daytime migration takes place in Finland in mid-May.
-
The copolar correlation coefficient ρhv (image not shown) exhibits values in the approximate range of 0.7–0.9 that are too low for other potential classes like drizzle.
It is also known from AWS observations that the weather is fair in the selected region. The final conclusion is thus that the class insects is the best categorization of the selected data.
The example given above illustrates the reasoning applied to all of the approximately 2000 cases of training data. Categorization was naturally easier and faster in cases when multiple elevations or temporally consecutive samples were picked to represent a class in a particular area. Altogether, the training set represents about 4 months of full-time work.
3. Methods
As shown in Fig. 2, the overall organization of the classifier system consists of three semiautonomous subsystems. The training subsystem converts the training dataset into class specific probability density functions (PDFs). The auxiliary data subsystem determines a priori values by applying simple directives to any observation-related additional data that are available. The actual classifier is implemented as a modular pipe with the following successive operations: preprocessing, consisting of equalization of data moments and generation of additional channels through filtering; estimation of class probabilities by Bayesian inference, first based on the given class PDFs and a priori values and then based on class probabilities in a limited neighborhood (spatial correlation); and finally composition of the desired product.
Data-flow diagram of the classifier system with the subsystems for training (blue), classification (green), and a priori value determination (red), respectively.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
Data equalization is done by applying the cumulative distribution function FX(x) = P(X ≤ x) of the entire training set to each moment separately. Thus, the modified distribution of data for each moment is uniform (0, 1), which greatly increases the accuracy of single-class distribution models of a given resolution for pronouncedly nonuniformly distributed moments like ρhv.
a. Class probabilities
The training set of a class consists of single-bin measurements of the values of radar moments. This set of measurement vectors {X} can be considered a random sampling of the probability density function p(x) of the respective class. The choice of method for estimating p(x) from {X} directly affects the accuracy of classification. In this study we have chosen the naive Bayes (NB) model as the operational method because of its computational simplicity combined with the ability to approximate arbitrary density distributions. In the NB model the multivariate probability density is calculated as a product of independent one-dimensional probability densities,
b. Filters
In addition to single-bin values, we consider statistical characteristics of adjacent bins through manually designed filters that are applied to the measured data to create additional data channels. In polar coordinates, a filter operator OH is applied to a (2H + 1) × (2H + 1) bin aperture of the original moment where H is the aperture half-width in bins, and the result is stored as the value of the center bin of the respective derived channel. In our implementation of the filtering system an arbitrary number of operations can be chained together to create more complex filtering effects. Our choice of basic operators is informed by known statistical features of the target classes in this study. In the following equations the measured bin values in the aperture xij, where the indices i and j are relative to the center bin, are assumed to have been normalized into the range [0, 1] prior to filtering. Bins without a valid value (either not measured or measured but undetected) in the aperture are excluded from the filter response calculation. Filter response for a nonresponding center bin is similarly nonresponse.
Exemplary signal responses for filters operating on the equalized radar moment Zt in (a) are shown in (b) for the filter F0(Zt), (c) for F1(Zt), and (d) for F2(Zt), respectively, and the response for the moment υ in (e), is shown in (f) for the filter F3(υ). The original radar moments Zt and υ are taken from the same observation as in Fig. 1.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
c. Missing data
Weather radar data are conspicuously incomplete in a way that must be taken into account when analyzing them. The primary source of nonresponse is a lack of scattering media within a particular volume of space. This is marked as undetected return signal in the data format, in contrast to a lack of data when a measurement was not made in the first place. This sparseness of available data is taken into account by only classifying bins with responding Zt.
Another significant source of nonresponse is fully artificial, i.e., the Doppler filtering applied by the signal processor. While preemptive rejection of clutter is justified in the context of weather observations, it severely affects the classification of nonmeteorological targets.
This loss of data cannot be easily compensated, and thus it is to be expected that classification accuracy between low-responding classes suffers in proportion. Our system takes this kind of data nonresponse into account by having a separate classifier trained for each combination of responding and nonresponding moments in per-bin data vectors of the training set (Fig. 4), with a binary search tree that minimizes the expected number of response evaluations. This approach maximizes the available per-bin information by turning the response pattern of a data vector into a metavariable for classifier selection.
The optimal binary search tree for selecting a classifier based on moment response. Starting from the root, measurement vectors are checked for individual moment response (solid path for responding and dotted path for nonresponding) to determine the correct classifier, here numbered from 0 through 6.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
d. Spatial correlation
While information from multiple bins can be incorporated by additional inferences, the task soon becomes intractable because correlations between all the included bins should also be taken into account. The number of required operations thus grows exponentially which is prohibitively costly for anything beyond a trivial number of bins. This is addressed by doing the inferences in a five-dimensional linearly uncorrelated subspace of rb as determined by principal component analysis (PCA).
e. Performance evaluation
The chosen classification method does not depend on internal parameter tuning that would necessitate a validation step separate from testing. Furthermore, we also avoid partitioning data into separate training and testing sets by utilizing the standard k-folding method (Breiman and Spector 1992) with k = 5. In this approach the full set of training cases is randomly divided into k partitions and each partition in turn is used as the testing set while the remaining k − 1 partitions are assigned as the training set.
Because of limitations in manuscript length, in this paper we only evaluate classification performance, i.e., determining the most probable class, and leave the issue of calibration of absolute probabilities in class detection outside the scope of the study. We evaluate the accuracy and recall of different classifier configurations using the F1 score (van Rijsbergen 1979). The proportion correct score (PC; Finley 1884) is used to evaluate the accuracy separately, and the Heidke skill score (HSS; Heidke 1926) and Hanssen and Kuipers discriminant [HK; Hanssen and Kuipers 1965; also known as the Peirce skill score (Peirce 1884)] are used to evaluate classification skill (see appendix).
4. Results
a. Class affinity
The phenomenological division of scatterers into hydrometeors, terrain, zoogenic, anthropogenic, and immaterial is not directly related to their measured characteristics in radar data. The measured differences can, however, be visualized by flattening the class dissimilarity matrix [Eq. (15)] with the MDS algorithm as shown in Fig. 5. A cursory examination of the plot affirms that meteorological classes do indeed form a homogeneous group in their measurable characteristics as well, whereas all other groupings are heterogeneous to some degree.
MDS visualization of class affinity. Class indices refer to the definitions given in Tables 2–6 and colors indicate phenomenological grouping of classes.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
b. Moment relevance
We assess the specific relevance of each moment for each class by calculating the momentwise mean dissimilarity through Eq. (15). The results of this exercise are shown in Fig. 6. It can be seen that the moments υ and σ are relatively nonindicative of a particular target class. Furthermore, the spot filter F2(Zt) is only relevant for the mostly spotlike target classes ships, aircraft, and noise, which is again an expected result. Beyond these observations, these results can be used to determine the relevant subset of moments for each class. This is pertinent information if one wants to construct a classification product for some specific class or a subset of classes.
The relevance of moments for the purpose of classification. The area of the circles is proportional to the mean dissimilarity of the PDF of the respective class and moment in comparison to other classes, weighted by the respective coverage. Class indices are those given in Tables 2–6, with grouping indicated by alternating background color.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
c. Classifier performance
To estimate the performance of the classifier, it was tested with the k-folding procedure in four different configurations: first as a basic NB classifier using only the measured moments, then as a basic NB classifier using the full set of measured moments and derived channels, then using multiple data response-based NB classifiers with the full set of measured moments and derived channels, and finally as the complete classifier with all the above mentioned methods as well as postclassifier pass of the spatial inference module. The respective performance scores are plotted in Fig. 7, showing that each additional method noticeably increases the overall skill of the classifier. A visualization of the full confusion matrix of the most feature-complete version of the classifier (i.e., configuration d) is shown in Fig. 8.
A comparison of the F1 score, proportion correct score (PC), Heidke skill score (HSS), and Hanssen and Kuipers discriminant (HK) of the classifier when using only polarimetric moments (a in the legend/key), polarimetric moments and derived channels (b), all moments, channels and response-based subclassifiers (c), and all moments, channels, subclassifiers and spatial correlation (d).
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
Confusion matrix of the NB classifier with all the available methods enabled. The area of the circles is proportional to the probability that a measurement assigned to a class is classified to the respective class. Reference circles in background color correspond to random chance, with the smaller of the actual and reference circles drawn over the other. Class indices are those given in Tables 2–6, with grouping indicated by alternating background color.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
d. Classification example
As an independent case example, not used in the training of the method described above, we will now present a fully classified image, i.e., the most probable class at each bin, of the lowest elevation PPI scan (0.3°) of the Kuopio (KUO) radar at 2245 UTC 19 May 2020 (Fig. 9) and assess the results manually.
The lowest elevation PPI scan of the Kuopio radar at 2245 UTC 19 May 2020 used as an independent test case, with radar moments (a) Zt, (b) υ, (c) Zdr, and (d) ρhv shown.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
When looking at the total reflectivity Zt in Fig. 9 the first rather obvious assessment is that the echo pattern consists of the following major components:
-
A somewhat extended and irregular area surrounding the radar site up to the distance of 50–70 km (red) exhibits features characteristic of ground clutter. This interpretation is confirmed by the fact that the pattern is not present in other moments where Doppler filtering has been applied.
-
An annulus with a spotty outer edge and values mostly in the interval from −10 to 0 dBZ, surrounding the radar up to the distance of 120–130 km (blue) most likely corresponds to nocturnally migrating songbirds (passerines), since the pattern as well as both the date and local time (0145 LT) are congruent with the seasonal and diurnal occurrence of such a migration. However, the shape of the pattern and Zt alone cannot exclude insects, drizzle, or snow grain.
-
The irregular green and yellow patches of echoes at distances from 50 to 240 km suggest typical shapes and intensities of mostly weak convective precipitation instead of widespread Ns. This is also supported by the synoptic situation explained below.
When assessing the tentative classification of the second component as birds, we first look at υ where the Doppler velocity pattern inside the second component is quite variable and does not match the wind field indicated by the third component. It also contains smaller-scale spotty variance that is typical for birds (Holleman et al. 2008) but not for precipitation or chaff. Relatively low ρhv also confirms that the blue annulus in Zt cannot be drizzle or snow grain that, according to the Finnish AWS comparisons during the training represent high values of ρhv, which we thus exclude, leaving birds and insects as possible explanations for the component.
The blue spots around the outer edge of the annulus (i.e., at a higher altitude) in particular exhibit highly variable velocities, a feature that is more congruent with birds—corresponding to a rather homogeneous mass in the annulus and solitary flocks above it—than with insects. Interestingly, ρhv and Zdr exhibit azimuthal patterns of maxima and minima that are perpendicular between the two moments (axis of maximum approximately 170°–350° in Zdr and 90°–250° in ρhv). These patterns may suggest common body orientation of the animals, a phenomenon that can be pronounced for insect migration [see the training case example and Melnikov et al. (2015)] and at least in S-band for birds (Radhakrishna et al. 2019). Despite this, the interpretation still strongly leans toward small birds due to the measured moment intensities and the recognized pattern.
Finally, the tentative classification is supported by external weather conditions. At the time of observation a cold polar air mass prevailed in northern Europe. According to a nearby sounding the temperature was close to 0°C at ground and −4°C at 850 hPa. The freezing level was at the altitude of 700–800 m and the maximum temperature above ground was +3°C. The wind was weak and from the NE sector in the lowest 1.5 km. Such weather conditions in mid-May are not at all favorable to insect ingress migration in Finland and the local nocturnal insects (moths) are not flying in large numbers in such temperatures. Nocturnally migrating passerines, on the other hand, are in a hurry to reach their breeding grounds further north and their behavior is indifferent to temperature variations above approximately 5°–10°C below zero. Based on these considerations we conclusively classify the second component as birds.
Turning our attention to the third component, ρhv and Zdr suggest that the initial nonpolarimetric classification into precipitation does not appear to be completely valid. Both moments clearly indicate that precipitation occurs only in the SW and NE sectors where Zdr is close to zero and ρhv close to one. The otherwise similar patches in the NW sector must be something else. The Doppler velocity in υ exhibits a less regular pattern of radial velocities but it is still clearly discernible that the unidentified echoes in the NW sector are moving with the prevailing wind, congruent with the velocity pattern of precipitation and incongruent with the velocity pattern of the adjacent birds. Based on the motion, observed structure, and moment values (especially very low ρhv and high Zdr) we choose chaff as the most probable target class for the NW echoes. This assessment agrees with that of Zrnić and Ryzhkov (2004) and was later verified by the Finnish Air Force, confirming that chaff had been released before the observation in that region.
When comparing this manual classification to the output of the classifier in Fig. 10 it can be seen that the patterns of ground clutter, birds, precipitation, and chaff coincide with those suggested by the human interpreter. The classification of precipitation appears somewhat messy, especially at the weak edges. Measuring and properly classifying the weakest precipitation echoes (approximately below 5 dBZ) remains a common challenge, since the signal-to-noise ratio is low and the computation of polarimetric moments suffers from large relative errors. Such reflectivity levels are not significant in hydrological applications but can be pertinent for, e.g., visibility estimation in the cases of drizzle, snow grain, and weak snowfall (Dixon et al. 2004). On the other hand, the cores of the convective cells with reflectivity factors of approximately 15–30 dBZ (panel Zt in Fig. 9) are properly classified. These known issues aside, there exists an excellent overall agreement between the two interpretations.
Classification results for Kuopio radar at 2245 UTC 19 May 2020. Probability distributions for the class groups (a) hydrometeors, (b) terrain, and (c) birds, and the classes (d) insects and (e) chaff, and (f) a detection likelihood weighted composite image of all class probabilities. The colors of the composite are as in Fig. 5.
Citation: Journal of Atmospheric and Oceanic Technology 39, 10; 10.1175/JTECH-D-21-0177.1
5. Discussion
Existing polarimetric classification schemes utilized in QPE can separate different types of hydrometeors rather well (e.g., Lim et al. 2005; Park et al. 2009; Dolan et al. 2013; Chandrasekar et al. 2013) and also detect regions of nonprecipitating echoes (e.g., Gourley et al. 2007; Ye et al. 2015; Kilambi et al. 2018). However, relatively little effort has been dedicated to fine-grained classification of nonmeteorological echoes (Stepanian et al. 2016). On the level of individual range gate measurements (bins) such classifications are just emerging (Jatau et al. 2021). To our knowledge this paper represents the first comprehensive statistical study of the bin-based polarimetric characteristics of all common nonmeteorological echo types, for C-band weather radars in particular. In addition to seven meteorological classes, our approach successfully pairs the measured polarimetric moments of nonmeteorological echoes at each bin with 4 top-level (terrain, zoogenic, anthropogenic, and immaterial) and 12 fine-grained subclasses through Bayesian inference based on a representative set of manually collected and classified training data.
The chosen method provides remarkably consistent and accurate per-bin classification of nonmeteorological scatterers. The computational complexity—and thus the cost of classification—of the chosen NB approach is roughly equivalent to the more commonly used fuzzy logic classifiers. However, in addition to the classification product itself NB also provides well-substantiated estimates for individual class detection probabilities through statistical modeling of the underlying multivariate probability distributions. Thus, a separate method for product accuracy estimation is not required. Another desirable feature of the NB approach is the consistent and comprehensive usage of all the measured moments in the classification process even for targets with less well-defined scattering characteristics since the method derives the moment distributions directly from the training data.
The training dataset used for estimating class probability distributions covers a wide range of regularly occurring cases. It is a strong indicator of the validity of the chosen approach that this large-scale assignment of class identities yields a consistent classification on a per-bin basis, leaving only room for some gross systematic error in the class assignment process. Systematic misidentification of a class in the training, however, would lead to erroneous assignment of at least two classes, since all common echo types are included in the classification. Accordingly, values of polarimetric moments of wrongly assigned classes would not fit to already published class properties (Tables 3–6). Proper assignment was also supported by careful analysis of meteorological and environmental circumstances.
Detection of target classes can in principle be affected by several factors including flawed methodology, insufficiently specific data, improper cross contamination in the training set, or inadequate coverage of the training set. For instance, it is a known fact that the quality of polarimetric moments decreases closer to the level of noise at the edges of weak echoes and at isolated pixels (Ryzhkov et al. 2005; Kilambi et al. 2018). In the context of nonmeteorological target classification the relative lack of independent in situ reference measurements of the scattering media poses a significant challenge since it is not possible to arrange a massive campaign that could measure the in situ physical conditions at the randomly occurring locations of each class (Zrnić and Ryzhkov 1998). In combination with the far greater dimensionality of the data this largely rules out approaches based on theoretical partitioning of the measurement space, i.e., the kind of membership functions used in common meteorological fuzzy logic classifiers of hydrometeors. Furthermore, there is a growing need for weather radar products providing information on observation uncertainty and quality (Saltikoff et al. 2019; Jurczyk et al. 2020) which is most readily met by probabilistic classification methods.
The NB approach is both mathematically and computationally simple enough that we have been able to verify that the produced model is a fair approximation of the underlying probability distributions, thus excluding methodological failure from potential explanations. Despite careful selection of the training data some cross contamination (e.g., blended migration of insects and birds) is almost inevitable but our assessment in the chosen framework is that the resulting inaccuracy is negligible. Major random sources of error are not present as can be seen in Fig. 8.
Despite the issues mentioned in the previous paragraphs the current performance of the classification system described in this paper is sufficient for operational use in automatic detection of various nonmeteorological target classes. Services using classified image products provided by FMI have been successfully implemented by various end users. We expect the methodology to be applicable to other C-band networks outside FMI but local retraining may be required depending on the radar systems (e.g., manufacturer, antenna sites) and environmental conditions such as precipitation climate, topography and characteristics of animal migrations. A complete retraining would be required for applying the method for radar systems operating at other wavelengths.
Although it is beyond the scope of this study, judicious determination of a priori values for each bin and class has the potential to noticeably enhance the accuracy of the classification by utilizing heterogenous external knowledge like local terrain type, time of day or year, or ground temperature. Moderate gains in classification accuracy can be also anticipated from the ongoing refinement and augmentation of the teaching set and the external a priori rules. Other potential avenues for increased accuracy within the current framework would be the implementation of class specific pattern filters for targets with known characteristic features, and the expansion of the considered bin neighborhood from two-dimensional images to three-dimensional radar volumes or even four-dimensional time series of radar volumes.
Overall, the greatest immediate benefit of the chosen approach is comprehensive individual identification of various types of nonmeteorological echoes which has so far received little attention in literature. It will also assist in obtaining more accurate separation of precipitation from other echoes. Furthermore, the approach opens new avenues for research and applications, e.g., in regional and local-scale environmental monitoring. This includes topics like aerial biomass flow by birds and insects, and monitoring of ducting conditions of electromagnetic propagation over the sea by data from weather radar networks. Concerning the latter application that specifically relies on clutter detection, it must be noted that ground clutter is predominantly missing in the products that have passed Doppler filtering in the signal processor. However, occurrence of anomalous ground clutter in Zt as well as ship and sea clutter in υ and in polarimetric moments is a good indicator of anomalous propagation in ducting conditions (Bebbington et al. 2007).
Furthermore, while conventional and commonly available hydrometeor class products for fine-grained subdivision of precipitation (e.g., rain, dry snow, wet snow, graupel, hail) must be considered preferable over the results of our implementation, echoes from boundary layer snow (i.e., shallow weak snowfall or snow grain) and drizzle are not specifically covered by the existing classifications of hydrometeor types, and thus the presence of the respective classes in our system is a novel and useful feature for the estimation of visibility.
Considering the prospects of aeroecological utilization of weather radars, it has been proposed that the traditional approach of “clutter rejection” in QPE should be replaced with “echo information” in multidisciplinary weather radar applications (Peura and Koistinen 2006; Chilson et al. 2012; Radhakrishna et al. 2019). In practice, it is suggested that after an initial class assessment all measured information from the various scattering media should be saved, not rejected, as the complete dataset supports several interdisciplinary applications in weather services and beyond. Additional challenges can also arise from irreversible quality control (QC) procedures a posteriori (Shamoun-Baranes et al. 2021). The method presented in this paper can easily solve issues in application-dependent QC by providing per-bin probabilistic assessment of utility for, e.g., QPE or animal quantification.
Referring to the already extensive scientific utilization of weather radars in bird migration studies the method described in this paper, together with the method of Jatau et al. (2021), offers a significant improvement in the spatial resolution of migration analysis, from averaged layers above each radar to individual bins. While layers are suitable for continental-scale analysis (e.g., Nilsson et al. 2019), bin-based analysis enables regional and local analyses of biomass flow. Also, radar-based monitoring of the local vertical density distribution of migrating birds is valuable for aviation since bird collisions are a serious threat for aircraft (van Gasteren et al. 2019). Similarly, risk assessment studies in the wind energy sector could also potentially leverage weather radar data when planning new wind farms. In the same way the ability to warn farmers of migrations of pest insects is a new potential application for weather radar networks (Leskinen et al. 2011).
These kinds of emerging applications demonstrate that the quality of radar measurements cannot be specified as a single, universal metric in terms of QPE but depends on the application, the measured target type, and on the specificity of the customers’ requirements instead. Thus, ideally, the process of including or rejecting data should happen in the product generation phase according to the requirements of each specific application and the needs of the end users instead of irreversibly in the signal processing phase. The present approach provides a first step toward novel opportunities for such a broader suit of bin-based applications.
Acknowledgments.
Part of this research was funded through the 2017–18 Belmont Forum and BiodivERsA joint call for research proposals, under the BiodivScen ERA-Net COFUND programme, and with the funding organizations Academy of Finland (aka 326315), Swiss National Science Foundation (SNF 31BD30_184120), Belgian Federal Science Policy Office (BelSPO BR/185/A1/GloBAM-BE), Netherlands Organisation for Scientific Research (NWO E10008), and National Science Foundation (NSF 1927743).
Data availability statement.
The metadata XML file defining the training set of the classifier used in this paper, and the confusion matrices for different classifier configurations are available at https://fmi.b2share.csc.fi/records/734f69f590c040069f0901d105872db9. The FMI radar volumes referenced by the training set are provided by the Amazon AWS S3 service at http://fmi-opendata-radar-volume-hdf5.s3-website-eu-west-1.amazonaws.com/.
APPENDIX
Definitions
REFERENCES
Atlas, D., Ed., 1990: Radar in Meteorology: Battan Memorial and 40th Anniversary Radar Meteorology Conference. 1st ed. Amer. Meteor. Soc., 806 pp.
Battaglia, A., K. Mroz, S. Tanelli, F. Tridon, and P.-E. Kirstetter, 2016: Multiple-scattering-induced “ghost echoes” in GPM DPR observations of a tornadic supercell. J. Appl. Meteor. Climatol., 55, 1653–1666, https://doi.org/10.1175/JAMC-D-15-0136.1.
Battan, L. J., 1973: Radar Observation of the Atmosphere. University of Chicago Press, 324 pp.
Bebbington, D., S. Rae, J. Bech, B. Codina, and M. Picanyol, 2007: Modelling of weather radar echoes from anomalous propagation using a hybrid parabolic equation method and NWP model data. Nat. Hazards Earth Syst. Sci., 7, 391–398, https://doi.org/10.5194/nhess-7-391-2007.
Bechini, R., and V. Chandrasekar, 2015: A semisupervised robust hydrometeor classification method for dual-polarization radar applications. J. Atmos. Oceanic Technol., 32, 22–47, https://doi.org/10.1175/JTECH-D-14-00097.1.
Breiman, L., and P. Spector, 1992: Submodel selection and evaluation in regression: The X-random case. Int. Stat. Rev., 60, 291–319, https://doi.org/10.2307/1403680.
Bringi, V. N., T. D. Keenan, and V. Chandrasekar, 2001: Correcting C-band radar reflectivity and differential reflectivity data for rain attenuation: A self-consistent method with constraints. IEEE Trans. Geosci. Remote Sens., 39, 1906–1915, https://doi.org/10.1109/36.951081.
Chandrasekar, V., R. Keränen, S. Lim, and D. Moisseev, 2013: Recent advances in classification of observations from dual polarization weather radars. Atmos. Res., 119, 97–111, https://doi.org/10.1016/j.atmosres.2011.08.014.
Chanthavong, V., J. Holmes, R. Keränen, D. Paris, J. Selzler, A. Siggia, and T. Stordell, 2010: Mitigation of sea clutter and other non-stationary echoes based on general purpose polarimetric echo identification. Sixth European Conf. on Radar in Meteorology and Hydrology, Sibiu, Romania, EUMETSAT.
Chilson, P. B., and Coauthors, 2012: Partly cloudy with a chance of migration: Weather, radars, and aeroecology. Bull. Amer. Meteor. Soc., 93, 669–686, https://doi.org/10.1175/BAMS-D-11-00099.1.
Dixon, M., R. M. Rasmussen, and S. Landolt, 2004: Short-term forecasting of airport surface visibility using radar and ASOS. 11th Conf. on Aviation, Range and Aerospace, Hyannis, MA, Amer. Meteor. Soc., 10.6, https://ams.confex.com/ams/11aram22sls/webprogram/Paper82036.html.
Dokter, A. M., F. Liechti, H. Stark, L. Delobbe, P. Tabary, and I. Holleman, 2011: Bird migration flight altitudes studied by a network of operational weather radars. J. Roy. Soc. Interface, 8, 30–43, https://doi.org/10.1098/rsif.2010.0116.
Dokter, A. M., and Coauthors, 2018: Seasonal abundance and survival of North America’s migratory avifauna determined by weather radar. Nat. Ecol. Evol., 2, 1603–1609, https://doi.org/10.1038/s41559-018-0666-4.
Dolan, B., S. A. Rutledge, S. Lim, V. Chandrasekar, and M. Thurai, 2013: A robust C-band hydrometeor identification algorithm and application to a long-term polarimetric radar dataset. J. Appl. Meteor. Climatol., 52, 2162–2186, https://doi.org/10.1175/JAMC-D-12-0275.1.
Fabry, F., 2015: Radar Meteorology: Principles and Practice. Cambridge University Press, 272 pp.
Finley, J. P., 1884: Tornado prediction. Amer. Meteor. J., 1, 85–88.
Gauthreaux, S., and R. Diehl, 2020: Discrimination of biological scatterers in polarimetric weather radar data: Opportunities and challenges. Remote Sens., 12, 545, https://doi.org/10.3390/rs12030545.
Gauthreaux, S., D. S. Mizrahi, and C. G. Belser, 1998: Bird migration and bias of WSR-88D wind estimates. Wea. Forecasting, 13, 465–481, https://doi.org/10.1175/1520-0434(1998)013<0465:BMABOW>2.0.CO;2.
Gourley, J. J., P. Tabary, and J. P. du Chatelet 2007: A fuzzy logic algorithm for the separation of precipitating from nonprecipitating echoes using polarimetric radar observations. J. Atmos. Oceanic Technol., 24, 1439–1451, https://doi.org/10.1175/JTECH2035.1.
Grazioli, J., D. Tuia, and A. Berne, 2015: Hydrometeor classification from polarimetric radar measurements: A clustering approach. Atmos. Meas. Tech., 8, 149–170, https://doi.org/10.5194/amt-8-149-2015.
Hannesen, K., and A. Weipert, 2012: A probability-based sea clutter suppression method for polarimetric weather radar systems. Int. Symp. on Weather Radar and Hydrology, Exeter, United Kingdom, IAHS, 52–57.
Hanssen, A. W., and W. J. A. Kuipers, 1965: On the relationship between the frequency of rain and various meteorological parameters. Meded. Verh., 81, 2–15.
Heidke, P., 1926: Berechnung des erfolges und der güte der windstärkevorhersagen im sturmwarnungsdienst. Geogr. Ann., 8, 301–349.
Holleman, I., H. van Gasteren, and W. Bouten, 2008: Quality assessment of weather radar wind profiles during bird migration. J. Atmos. Oceanic Technol., 25, 2188–2198, https://doi.org/10.1175/2008JTECHA1067.1.
Huuskonen, A., M. Kurri, H. Hohti, H. Beekhuis, H. Leijnse, and I. Holleman, 2014: Radar performance monitoring using the angular width of the solar image. J. Atmos. Oceanic Technol., 31, 1704–1712, https://doi.org/10.1175/JTECH-D-13-00246.1.
Jatau, P., V. Melnikov, and T.-Y. Yu, 2021: A machine learning approach for classifying bird and insect radar echoes with S-band polarimetric weather radar. J. Atmos. Oceanic Technol., 38, 1797–1812, https://doi.org/10.1175/JTECH-D-20-0180.1.
Jiang, Y., Q. Xu, P. Zhang, K. Nai, and L. Liu, 2013: Using WSR-88D polarimetric data to identify bird-contaminated Doppler velocities. Adv. Meteor., 2013, 769275, https://doi.org/10.1155/2013/769275.
Jurczyk, A., J. Szturc, and K. Ośródka, 2020: Quality-based compositing of weather radar derived precipitation. Meteor. Appl., 27, e1812, https://doi.org/10.1002/met.1812.
Kelly, J. F., J. R. Shipley, P. B. Chilson, K. W. Howard, W. F. Frick, and T. H. Kunz, 2012: Quantifying animal phenology in the aerosphere at a continental scale using NEXRAD weather radars. Ecosphere, 3, 16, https://doi.org/10.1890/ES11-00257.1.
Kilambi, A., F. Fabry, and V. Meunier, 2018: A simple and effective method for separating meteorological from nonmeteorological targets using dual-polarization data. J. Atmos. Oceanic Technol., 35, 1415–1424, https://doi.org/10.1175/JTECH-D-17-0175.1.
Koistinen, J., 2000: Bird migration patterns on weather radars. Phys. Chem. Earth, 25B, 1185–1193, https://doi.org/10.1016/S1464-1909(00)00176-3.
Kranstauber, B., W. Bouten, H. Leijnse, B.-C. Wijers, L. Verlinden, J. Shamoun-Baranes, and A. M. Dokter, 2020: High-resolution spatial distribution of bird movements estimated from a weather radar network. Remote Sens., 12, 635, https://doi.org/10.3390/rs12040635.
Lakshmanan, V., C. Karstens, J. Krause, and L. Tang, 2014: Quality control of weather radar data using polarimetric variables. J. Atmos. Oceanic Technol., 31, 1234–1249, https://doi.org/10.1175/JTECH-D-13-00073.1.
Leskinen, M., and Coauthors, 2011: Pest insect immigration warning by an atmospheric dispersion model, weather radars and traps. J. Appl. Entomol., 135, 55–67, https://doi.org/10.1111/j.1439-0418.2009.01480.x.
Lim, S., V. Chandrasekar, and V. N. Bringi, 2005: Hydrometeor classification system using dual-polarization radar measurements: Model improvements and in situ verification. IEEE Trans. Geosci. Remote Sens., 43, 792–801, https://doi.org/10.1109/TGRS.2004.843077.
Melnikov, V. M., M. J. Istok, and J. Westbrook, 2015: Asymmetric radar echo patterns from insects. J. Atmos. Oceanic Technol., 32, 659–674, https://doi.org/10.1175/JTECH-D-13-00247.1.
Nieminen, M., M. Leskinen, and J. Helenius, 2000: Doppler radar detection of exceptional mass-migrations of aphids into Finland. Int. J. Biometeor., 44, 172–181, https://doi.org/10.1007/s004840000064.
Nilsson, C., and Coauthors, 2019: Revealing patterns of nocturnal migration using the European weather radar network. Ecography, 42, 876–886, https://doi.org/10.1111/ecog.04003.
Park, H. S., A. V. Ryzhkov, D. S. Zrnić, and K.-E. Kim, 2009: The hydrometeor classification algorithm for the polarimetric WSR-88D: Description and application to an MCS. Wea. Forecasting, 24, 730–748, https://doi.org/10.1175/2008WAF2222205.1.
Peirce, C. S., 1884: The numerical measure of the success of predictions. Science, 4, 453–454, https://doi.org/10.1126/science.ns-4.93.453.b.
Peura, M., and J. Koistinen, 2006: Quality information in processing weather radar data for varying user needs. Proc. Fourth European Conf. on Radar in Meteorology and Hydrology, Barcelona, Spain, Copernicus, 563–566.
Puhakka, T., and P. Saarikivi, 1986: Doppler radar observations of horizontal roll vortices in Finland. University of Helsinki Dept. of Meteorology Rep., 18 pp.
Puhakka, T., J. Koistinen, and P. L. Smith, 1986: Doppler radar observation of a sea breeze front. Preprints, 23rd Conf. on Radar Meteorology, Snowmass, CO, Amer. Meteor. Soc., JP198–JP201.
Radhakrishna, B., F. Fabry, and A. Kilambi, 2019: Fuzzy logic algorithms to identify birds, precipitation, and ground clutter in S-band radar data using polarimetric and nonpolarimetric variables. J. Atmos. Oceanic Technol., 36, 2401–2414, https://doi.org/10.1175/JTECH-D-19-0088.1.
Roberto, N., L. Baldini, E. Adirosi, L. Facheris, F. Cuccoli, A. Lupidi, and A. Garzelli, 2017: A support vector machine hydrometeor classification algorithm for dual-polarization radar. Atmosphere, 8, 134, https://doi.org/10.3390/atmos8080134.
Rosenberg, K. V., and Coauthors, 2019: Decline of the North American avifauna. Science, 366, 120–124, https://doi.org/10.1126/science.aaw1313.
Ryzhkov, A. V., T. J. Schuur, D. W. Burgess, P. L. Heinselman, S. E. Giangrande, and D. S. Zrnić, 2005: The Joint Polarization Experiment: Polarimetric rainfall measurements and hydrometeor classification. Bull. Amer. Meteor. Soc., 86, 809–824, https://doi.org/10.1175/BAMS-86-6-809.
Saltikoff, E., and Coauthors, 2019: OPERA the radar project. Atmosphere, 10, 320, https://doi.org/10.3390/atmos10060320.
Shamoun-Baranes, J., and Coauthors, 2021: Weather radars’ role in biodiversity monitoring. Science, 372, 248, https://doi.org/10.1126/science.abi4680.
Steiner, M., R. A. Houze Jr., and S. E. Yuter, 1995: Climatological characterization of three-dimensional storm structure from operational radar and rain gauge data. J. Appl. Meteor. Climatol., 34, 1978–2007, https://doi.org/10.1175/1520-0450(1995)034<1978:CCOTDS>2.0.CO;2.
Stepanian, P. M., and K. G. Horton, 2015: Extracting migrant flight orientation profiles using polarimetric radar. IEEE Trans. Geosci. Remote Sens., 53, 6518–6528, https://doi.org/10.1109/TGRS.2015.2443131.
Stepanian, P. M., K. G. Horton, V. M. Melnikov, D. S. Zrnić, and S. A. Gauthreaux Jr., 2016: Dual-polarization radar products for biological applications. Ecosphere, 7, e01539, https://doi.org/10.1002/ecs2.1539.
Torgerson, W. S., 1958: Theory and Methods of Scaling. 99th ed. John Wiley and Sons, 460 pp.
Vaisala, 2014: RVP900 digital receiver and signal processor. Vaisala Doc. M211322EN-D, 512 pp., https://www.manualslib.com/products/Vaisala-Rvp900-3758279.html.
Vaisala, 2017: User guide IRIS radar. Vaisala Doc. M211317EN-G, 120 pp., https://www.coursehero.com/file/64870958/IRIS-Radar-Manualpdf/.
van Gasteren, H., and Coauthors, 2019: Aeroecology meets aviation safety: Early warning systems in Europe and the Middle East prevent collisions between birds and aircraft. Ecography, 42, 899–911, https://doi.org/10.1111/ecog.04125.
van Rijsbergen, C. J., and C. J. K. Van Rijsbergen, 1979: Information Retrieval. Butterworths, 208 pp.
Wainwright, C. E., P. M. Stepanian, D. R. Reynolds, and A. M. Reynolds, 2017: The movement of small insects in the convective boundary layer: Linking patterns to processes. Sci. Rep., 7, 5438, https://doi.org/10.1038/s41598-017-04503-0.
Wang, Y., L. Tang, P.-L. Chang, and Y.-S. Tang, 2021: Separation of convective and stratiform precipitation using polarimetric radar data with a support vector machine method. Atmos. Meas. Tech., 14, 185–197, https://doi.org/10.5194/amt-14-185-2021.
Weisshaupt, N., A. M. Dokter, J. Arizaga, and M. Maruri, 2018: Effects of a sea barrier on large-scale migration patterns studied by a network of weather radars. Bird Study, 65, 232–240, https://doi.org/10.1080/00063657.2018.1476457.
Weisshaupt, N., T. Lehtiniemi, and J. Koistinen, 2020: Combining citizen science and weather radar data to study large-scale bird movements. IBIS, 163, 728–736, https://doi.org/10.1111/ibi.12906.
Ye, B.-Y., G. Lee, and H.-M. Park, 2015: Identification and removal of non-meteorological echoes in dual-polarization radar data based on a fuzzy logic algorithm. Adv. Atmos. Sci., 32, 1217–1230, https://doi.org/10.1007/s00376-015-4092-0.
Zrnić, D. S., and A. V. Ryzhkov, 1998: Observations of insects and birds with a polarimetric radar. IEEE Trans. Geosci. Remote Sens., 36, 661–668, https://doi.org/10.1109/36.662746.
Zrnić, D. S., and A. V. Ryzhkov, 2004: Polarimetric properties of chaff. J. Atmos. Oceanic Technol., 21, 1017–1024, https://doi.org/10.1175/1520-0426(2004)021<1017:PPOC>2.0.CO;2.
