Figure 5 - available via license: Creative Commons Attribution 4.0 International
Content may be subject to copyright.
Illustration of the nonlinear SVM. (a) A nonlinearly separable data set in the input space X , involving two classes (squares and circles). (b) Projection on a 3-D space H by the kernel K(x i , x j ) = x i , x j 2 . (c) Linear classification in the projected space H (filled dots are support vectors). (d) Corresponding nonlinear decision function in the original space X . Adapted from Volpi et al. (2013).
Source publication
![Figure 6. Evolution of K [–] as a function of the training set size....](publication/307700295/figure/fig6/AS:403277033295872@1473160145379/Evolution-of-K-as-a-function-of-the-training-set-size-The-solid-red-line-indicates_Q320.jpg)
The first hydrometeor classification technique based on
two-dimensional video disdrometer (2DVD) data is presented. The method provides
an estimate of the dominant hydrometeor type falling over time
intervals of 60 s during precipitation, using the
statistical behavior of a set of particle descriptors as input, calculated
for each particle image. T...
Similar publications
Deep learning has recently emerged as one of the most popular and powerful paradigms for learning tasks. In this paper, we present a deep learning approach to 3D shape classification using convolutional neural networks. The proposed framework takes a multi-stage approach that first represents each 3D shape in the dataset as a 2D image using the bag...
Automatic fall detection in videos could enable timely delivery of medical service to the injured elders who have fallen and live alone. Deep ConvNets have been used to detect fall actions. However, there still remain problems in deep video representations for fall detection. First, video frames are directly inputted to deep ConvNets. The visual fe...
Human action recognition is an important recent challenging task. Projecting depth images onto three depth motion maps (DMMs) and extracting deep convolutional neural network (DCNN) features are discriminant descriptor features to characterize the spatiotemporal information of a specific action from a sequence of depth images. In this paper, a unif...
In this work, automated abnormality detection using keypoint information from Speeded-Up Robust feature (SURF) and Scale Invariant Feature Transform (SIFT) descriptors in chest Radiographic (CR) images is investigated and compared. Computerized image analysis using artificial intelligence is crucial to detect subtle and non-specific alterations of...
Lung cancer is one of the leading causes of death in the world. Although early detection of lung tumours (nodules) can remarkably diminish the mortal rate, precise detection of them is not always possible by visual inspection of the computerised tomography images. Since nodules with different sizes have non-uniform shape and brightness, texture att...
Citations
... Using assumptions about sedimentation velocity or an aggregation and riming model as a reference, the particle mass-size and/or massdensity relationship can also be inferred from in situ observations (Tiira et al., 2016;Pet-tersen et al., 2020;Tokay et al., 2021;Leinonen et al., 2021;Vázquez-Martín et al., 2021a). Various attempts have been made to classify particle types and identify active snowfall formation processes using various machine learning techniques (Nurzyńska et al., 2013;Grazioli et al., 2014;Praz et al., 2017;Hicks and Notaroš, 2019;Leinonen and Berne, 2020;Del Guasta, 2022;Maherndl et al., 2023b); these classifications are needed to support quantification of snowfall formation processes (Grazioli et al., 2017;Moisseev et al., 2017;Dunnavan et al., 2019;Pasquier et al., 2023). In situ observations have also been used to characterize particle size distributions Fitch and Garrett, 2022), to investigate sedimentation velocity and turbulence of hydrometeors (Garrett et al., 2012;Garrett and Yuter, 2014;Li et al., 2021;Vázquez-Martín et al., 2021b;Takami et al., 2022), and for model evaluation (Vignon et al., 2019). ...
The open-source Video In Situ Snowfall Sensor (VISSS) is introduced as a novel instrument for the characterization of particle shape and size in snowfall. The VISSS consists of two cameras with LED backlights and telecentric lenses that allow accurate sizing and combine a large observation volume with relatively high pixel resolution and a design that limits wind disturbance. VISSS data products include various particle properties such as maximum extent, cross-sectional area, perimeter, complexity, and sedimentation velocity. Initial analysis shows that the VISSS provides robust statistics based on up to 10 000 unique particle observations per minute. Comparison of the VISSS with the collocated PIP (Precipitation Imaging Package) and Parsivel instruments at Hyytiälä, Finland, shows excellent agreement with the Parsivel but reveals some differences for the PIP that are likely related to PIP data processing and limitations of the PIP with respect to observing smaller particles. The open-source nature of the VISSS hardware plans, data acquisition software, and data processing libraries invites the community to contribute to the development of the instrument, which has many potential applications in atmospheric science and beyond.
... (Tiira et al., 2016;von Lerber et al., 2017;Pettersen et al., 2020;Tokay et al., 2021;Leinonen et al., 2021;Vázquez-Martín et al., 2021a). Various attempts have been made to classify particle types and identify active snowfall formation processes using various machine learning techniques (Nurzyńska et al., 2013;Grazioli et al., 2014;Praz et al., 2017;Hicks and Notaroš, 2019; 30 Leinonen and Berne, 2020;Del Guasta, 2022); these classifications are needed to support quantification of snowfall formation processes (Grazioli et al., 2017;Moisseev et al., 2017;Dunnavan et al., 2019;Pasquier et al., 2023). In situ observations have also been used to characterize particle size distributions (Kulie et al., 2021;Fitch and Garrett, 2022), investigate sedimentation velocity and turbulence of hydrometeors (Garrett et al., 2012;Garrett and Yuter, 2014;Li et al., 2021;Vázquez-Martín et al., 2021b;Takami et al., 2022), and for model evaluation (Vignon et al., 2019). ...
The open source Video In Situ Snowfall Sensor (VISSS) is introduced as a novel instrument for the characterization of particle shape and size in snowfall. The VISSS consists of two cameras with LED backlights and telecentric lenses that allow accurate sizing and combine a large observation volume with relatively high resolution and a design that limits wind disturbance. VISSS data products include per-particle properties and integrated particle size distribution properties such as particle maximum extent, cross-sectional area, perimeter, complexity, and – in the future – sedimentation velocity. Initial analysis shows that the VISSS provides robust statistics based on up to 100,000 particles observed per minute. Comparison of the VISSS with collocated PIP and Parsivel instruments at Hyytiälä, Finland, shows excellent agreement with Parsivel, but reveals some differences for the PIP (Precipitation Imaging Package) that are likely related to PIP data processing and limitations of the PIP with respect to observing smaller particles. The open source nature of the VISSS hardware plans, data acquisition software, and data processing libraries invites the community to contribute to the development of the instrument, which has many potential applications in atmospheric science and beyond.
... K dp in particular is an estimated variable affected by mean errors on the order of 30 % (Grazioli et al., 2014a). An additional ground-based source of information for this event is provided by a two-dimensional video disdrometer, 2DVD (for more information about this instrument at this location, see Grazioli et al., 2014b), which was deployed on the opposite side of the Davos valley with respect to MXPol (46.830 • N,9.810 • E; 2543 m a.m.s.l.). The 2DVD measures the size and fall velocity of hydrometeors captured within its measurement area of 11 cm × 11 cm. ...
The change in wind direction and speed with height, referred to as vertical wind shear, causes enhanced turbulence in the atmosphere. As a result, there are enhanced interactions between ice particles that break up during collisions in clouds which could cause heavy snowfall. For example, intense dual-polarization Doppler signatures in conjunction with strong vertical wind shear were observed by an X-band weather radar during a wintertime high-intensity precipitation event over the Swiss Alps. An enhancement of differential phase shift (Kdp > 1◦ km−1) around −15 ◦C suggested that a large population of oblate ice particles was present in the atmosphere. Here, we show that ice–graupel collisions are a likely origin of this population, probably enhanced by turbulence. We perform sensitivity simulations that include ice–graupel collisions of a cold frontal passage to investigate whether these simulations can capture the event better and whether the vertical wind shear had an impact on the secondary ice production (SIP) rate. The simulations are conducted with the Consortium for Small-scale Modeling (COSMO), at a 1 km horizontal grid spacing in the Davos region in Switzerland. The rime-splintering simulations could not reproduce the high ice crystal number concentrations, produced too large ice particles and therefore overestimated the radar reflectivity. The collisional-breakup simulations reproduced both the measured horizontal reflectivity and the ground-based observations of hydrometeor number concentration more accurately (∼ 20 L−1). During 14:30–15:45 UTC the vertical wind shear strengthened by 60 % within the region favorable for SIP. Calculation of the mutual information between the SIP rate and vertical wind shear and updraft velocity suggests that the SIP rate is best predicted by the vertical wind shear rather than the updraft velocity. The ice–graupel simulations were insensitive to the parameters in the model that control the size threshold for the conversion from ice to graupel and snow to graupel.
... including De and fall speed (Vf) (e.g., Kruger and Krajewski 2002;Kubo et al. 2009;Grazioli et al. 2014;Gavrilov 2015). The manufacturer-developed algorithm determines De for a hydrometeor from the measured volume, assuming that each of the projections is from an ellipsoid. ...
We performed a detailed analysis of ground-based data to investigate changes in the morphological properties and particle size distribution of precipitation particles as they fall through the melting layer (ML). In July 2013, we started continuous precipitation monitoring in Sapporo (Japan) with a two-dimensional video disdrometer, electrical balance-type snow gauge, and X-band marine radar. We used data collected from 09:43 to 10:40 Japan Standard Time (JST) on 10 March, 2015 for analysis, when the bright band progressively descended to the ground surface and precipitation intensity was moderate and approximately steady (∼10 mm hr ⁻¹ ). We found that the aggregation of aggregates in the upper half of the ML did not necessarily result in large raindrops. Almost all of the snow particles with a melted diameter ( D m ) ≥ 4 mm broke up before they melted into raindrops of equivalent size. The apparent one-to-one relationship between melting snow particles and raindrops held for particles with 2 < D m < 3 mm. Most small raindrops were generated by the successive breakup of melting 22 particles in the lower half of the ML.
... They provide the size distribution and falling speed of hydrometers, but they give no direct information about the shape. The evolution of disdrometers into 2D disdrometers gave access to some shape indications about hydrometeors (Grazioli et al., 2014). A further evolution of disdrometers into imaging disdrometers, such as the Snowflake Video Imager (SVI) (Newman et al., 2009), provided realistic images of the crystals. ...
Studying precipitation at very high latitudes is difficult because of the
harsh environmental conditions that limit the external activity of humans
and instruments, especially in the polar winter. The direct monitoring of
ice crystal habits and size distribution in Antarctic precipitation is
important for the validation of the algorithms used for retrieving
precipitation from ground-based and satellite-borne radar instruments and
for the improvement of the climatological modelling of polar areas. This paper
describes an automated device (ICE-CAMERA) specifically developed for the
imaging, measurement, and classification of ice precipitation on the
Antarctic high plateau. The instrument gives detailed information on
precipitation on an hourly basis. The article provides a description of the
device and its image processing software. Starting in 2014, the instrument
has operated almost unattended all year round at Concordia station, Antarctica (75∘ S, 123∘ E, 3220 m altitude).
... These descriptors included particle size, shape, and textural information and were all extracted during preprocessing. Praz et al. (2017) validated the classification method through comparison with two-dimensional video disdrometer (2DVD) data by Grazioli et al. (2014). Grazioli et al. (2014) used 2DVD data to employ a support vector machine (SVM) for classification tasks. ...
... Praz et al. (2017) validated the classification method through comparison with two-dimensional video disdrometer (2DVD) data by Grazioli et al. (2014). Grazioli et al. (2014) used 2DVD data to employ a support vector machine (SVM) for classification tasks. The SVM was trained over 1-min averages to predict the dominant hydrometeor type during precipitation events, which were visually labeled into eight dominant hydrometeor classes. ...
A vast amount of ice crystal imagery exists from a variety of field campaign initiatives that can be utilized for cloud microphysical research. Here, nine convolutional neural networks are used to classify particles into nine regimes on over 10 million images from the Cloud Particle Imager probe, including liquid and frozen states and particles with evidence of riming. A transfer learning approach proves that the Visual Geometry Group (VGG-16) network best classifies imagery with respect to multiple performance metrics. Classification accuracies on a validation dataset reach 97% and surpass traditional automated classification. Furthermore, after initial model training and preprocessing, 10 000 images can be classified in approximately 35 s using 20 central processing unit cores and two graphics processing units, which reaches real-time classification capabilities. Statistical analysis of the classified images indicates that a large portion (57%) of the dataset is unusable, meaning the images are too blurry or represent indistinguishable small fragments. In addition, 19% of the dataset is classified as liquid drops. After removal of fragments, blurry images, and cloud drops, 38% of the remaining ice particles are largely intersecting the image border (≥10% cutoff) and therefore are considered unusable because of the inability to properly classify and dimensionalize. After this filtering, an unprecedented database of 1 560 364 images across all campaigns is available for parameter extraction and bulk statistics on specific particle types in a wide variety of storm systems, which can act to improve the current state of microphysical parameterizations.
... For example, Zhang et al. (2022) utilized an object-detector model for predicting the 3D coordinates of relatively localized particles, while Shao et al. (2020) showed that a U-net can accomplish the same objective with similar performance. Other efforts have focused on performing classification tasks with holograms, where decision-tree approaches (Grazioli et al., 2014;Bernauer et al., 2016) and convolutional neural networks have been investigated (Zhang et al., 2018;Xiao et al., 2019;Touloupas et al., 2020;Wu et al., 2020). For example, Touloupas et al. (2020) used a vanilla CNN model for classifying objects detected in the holograms as either artifact, water droplet, or ice particle. ...
HOLODEC, an airborne cloud particle imager, captures holographic images of a fixed volume of cloud to characterize the types and sizes of cloud particles, such as water droplets and ice crystals. Cloud particle properties include position, diameter, and shape. We present a hologram processing algorithm, HolodecML, that utilizes a neural segmentation model, GPUs, and computational parallelization. HolodecML is trained using synthetically generated holograms based on a model of the instrument, and predicts masks around particles found within reconstructed images. From these masks, the position and size of the detected particles can be characterized in three dimensions. In order to successfully process real holograms, we find we must apply a series of image corrupting transformations and noise to the synthetic images used in training. In this evaluation, HolodecML had comparable position and size estimation performance to the standard processing method, but improved particle detection by nearly 20\% on several thousand manually labeled HOLODEC images. However, the improvement only occurred when image corruption was performed on the simulated images during training, thereby mimicking non-ideal conditions in the actual probe. The trained model also learned to differentiate artifacts and other impurities in the HOLODEC images from the particles, even though no such objects were present in the training data set, while the standard processing method struggled to separate particles from artifacts. The novelty of the training approach, which leveraged noise as a means for parameterizing non-ideal aspects of the HOLODEC detector, could be applied in other domains where the theoretical model is incapable of fully describing the real-world operation of the instrument and accurate truth data required for supervised learning cannot be obtained from real-world observations.
... More recently, accurate and high-resolution depictions of snowflakes could be obtained with imagers like the snow video imager/particle image probe (Newman et al., 2009) or with the multi-angle snowflake camera (MASC; Garrett et al., 2012). The availability of actual images has promoted the development and rapid improvement of several automatic hydrometeor classification techniques (Grazioli et al., 2014;Gavrilov et al., 2015;Praz et al., 2017;Leinonen and Berne, 2020) adapted to the data of these sensors. While the accuracy of the measurements of fall velocity provided by those instruments is often hampered by wind and turbulence (Nešpor et al., 2000;Garrett and Yuter, 2014;Fitch et al., 2021), the added value in terms of microphysical characterization is significant. ...
This paper presents a method named 3D-GAN, based on a generative adversarial network (GAN), to retrieve the total mass, 3D structure and the internal mass distribution of snowflakes. The method uses as input a triplet of binary silhouettes of particles, corresponding to the triplet of stereoscopic images of snowflakes in free fall captured by a multi-angle snowflake camera (MASC). The 3D-GAN method is trained on simulated snowflakes of known characteristics whose silhouettes are statistically similar to real MASC observations, and it is evaluated by means of snowflake replicas printed in 3D at 1:1 scale. The estimation of mass obtained by 3D-GAN has a normalized RMSE (NRMSE) of 40 %, a mean normalized bias (MNB) of 8 % and largely outperforms standard relationships based on maximum size and compactness. The volume of the convex hull of the particles is retrieved with NRMSE of 35 % and MNB of +19 %. In order to illustrate the potential of 3D-GAN to study snowfall microphysics and highlight its complementarity with existing retrieval algorithms, some application examples and ideas are provided, using as showcases the large available datasets of MASC images collected worldwide during various field campaigns. The combination of mass estimates (from 3D-GAN) and hydrometeor classification or riming degree estimation (from independent methods) allows, for example, to obtain mass-to-size power law parameters stratified on hydrometeor type or riming degree. The parameters obtained in this way are consistent with previous findings, with exponents overall around 2 and increasing with the degree of riming.
... More recently, accurate and high-resolution depictions of snowflakes could be obtained with imagers like the Snow Video Imager/Particle Image Probe (Newman et al., 2009) or with the Multi-Angle Snowflake Camera (MASC Garrett et al., 2012). The availability of actual images has promoted the development and rapid improvement of several automatic hydrometeor classification techniques (Grazioli et al., 2014;40 Gavrilov et al., 2015;Praz et al., 2017; adapted to the data of these sensors. While the accuracy of the measurements of fall velocity provided by those instruments is often hampered by wind and turbulence (Nešpor et al., 2000;Garrett and Yuter, 2014;Fitch et al., 2021), the added value in terms of microphysical characterization is significant. ...
This paper presents a method named 3D-GAN, based on a generative adversarial network (GAN), to retrieve the total mass, 3D structure and the internal mass distribution of snowflakes. The method uses as input a triplet of binary silhouettes of particles, corresponding to the triplet of stereoscopic images of snowflakes in free fall captured by a Multi-Angle Snowflake Camera (MASC). 3D-GAN is trained on simulated snowflakes of known characteristics whose silhouettes are statistically similar to real MASC observations and it is evaluated by means of snowflake replicas printed in 3D at 1 : 1 scale. The estimation of mass obtained by 3D-GAN has a normalized RMSE (NRMSE) of 40 %, a mean normalized bias (MNB) of 8 % and largely outperforms standard relationships based on maximum size and compactness. The volume of the convex hull of the particles is retrieved with MNRSE of 35 % and MNB of +19 %. In order to illustrate the potential of 3D-GAN to study snowfall microphysics and highlight its complementarity with existing retrieval algorithms, some application examples and ideas are provided, using as showcases the large available datasets of MASC images collected worldwide during various field campaigns. The combination of mass estimates (from 3D-GAN) and hydrometeor classification or riming degree estimation (from independent methods) allows for example to obtain mass-to-size power law parameters stratified on hydrometeor type or riming degree. The parameters obtained in this way are consistent with previous findings, with exponents overall around 2 and increasing with the degree of riming.
... A variety of attempts have been made to classify snowflakes (Nakaya and Sekido 1936;Magono and Lee 1966;Korolev and Sussman 2000;Grazioli et al. 2014;Vazquez-Martin et al. 2020). As in our previous work (Hicks and Notaro s 2019), we chose to use the scheme adopted by Praz et al. (2017) for training and testing of their multinomial logistic regression snowflake classifier. ...
We present improvements over our previous approach to automatic winter hydrometeor classification by means of convolutional neural networks (CNNs), using more data and improved training techniques to achieve higher accuracy on a more complicated dataset than we had previously demonstrated. As an advancement of our previous proof-of-concept study, this work demonstrates broader usefulness of deep CNNs by using a substantially larger and more diverse dataset, which we make publicly available, from many more snow events. We describe the collection, processing, and sorting of this dataset of over 25,000 high-quality multiple-angle snowflake camera (MASC) image chips split nearly evenly between five geometric classes: aggregate, columnar crystal, planar crystal, graupel, and small particle. Raw images were collected over 32 snowfall events between November 2014 and May 2016 near Greeley, Colorado and were processed with an automated cropping and normalization algorithm to yield 224x224 pixel images containing possible hydrometeors. From the bulk set of over 8,400,000 extracted images, a smaller dataset of 14,793 images was sorted by image quality and recognizability (Q&R) using manual inspection. A presorting network trained on the Q&R dataset was applied to all 8,400,000+ images to automatically collect a subset of 283,351 good snowflake images. Roughly 5,000 representative examples were then collected from this subset manually for each of the five geometric classes. With a higher emphasis on in-class variety than our previous work, the final dataset yields trained networks that better capture the imperfect cases and diverse forms that occur within the broad categories studied to achieve an accuracy of 96.2% on a vastly more challenging dataset.
