Next: Talks/seminar slides
Up: Research Publications
Previous: Papers at Engineering Conferences
Subsections
These papers were all presented at meteorological conferences
to audiences consisting primarily of meteorologists. These
papers are typically glance over the techniques involved,
but explain the methodology and underlying scientific assumptions.
The discussion of the results is also more thorough than the
Engineering Conference papers.
Entry (32) for the 3rd annual AI competition:
We employ partial least squares regression to transform the input data
into components that have a high correlation with the variable to
be predicted and preserve the variability in the dataset.
Then, the transformed data are presented to a neural
network whose output node has a sigmoid transfer function. Along the
way, much data processing is done.
[ talk (PDF) online ]
Invited talk on the stuff described in (7)
and a slightly expanded version of (33).
[ talk (PDF) online ]
[ talk (PPT) online ]
A talk on the stuff described in (7)
and a slightly expanded version of (33).
[ talk (PDF) online ]
In (1), we introduce a new approach that breaks down the
observed and forecast fields into a mixture of Gaussians and
examine the parameters of the GMM fit to identify translation,
rotation and scaling errors.
We discuss the advantages of this method in
terms of the traditional filtering or object-based methods and
interpret resulting scores on a standard verification dataset.
[ talk (PDF) online ].
In (33), we describe the multiscale storm identification,
motion estimation, tracking and attribute extraction algorithm
in WDSS-II (segmotion). We explain the impact of the various
tuneable parameters in the algorithm and provide pointers on
how to choose those parameters to fit different tracking applications.
[ talk (PDF) online ].
In (34),
cloud-to-ground lightning data from the National Lightning Data
Network (NLDN), satellite visible and radar-derived products are used
to train a lightning prediction algorithm. The radar reflectivity
values are clustered to identify storm and real-time geometric,
lagrangian and scalar attributes of those storms are computed. A
lightning density field is "precast" to form the target decision field
to be predicted using the computed attributes. Several days of data
from the continental United States were chosen to obtain a seasonally
and geographically diverse dataset for training. The trained system is
used to predict lightning density and the predicted lightning density
field is advected to produce a 30-minute nowcast field. The skill of
the resulting algorithm is evaluated against a
steady-state prediction with motion correction.
[ paper (PDF) online ].
[ talk (PDF) online ].
In (35), we built three different AI models to predict the category
(frozen, liquid or none) based on polarimetric radar variables.
One model was a long-shot model for the purpose of winning
the competition (it didn't). The second was a complex black box
model to represent the average learnability of the dataset.
The third was a simple, human-readable model that would perform
similarly to the above two models but possess the additional advantage
of being easy to implement and comprehend.
[ paper (PDF) online ].
[ slides (PDF) online ].
In (36), we developed a data-driven streamflow prediction model using observations of rainfall and runoff over the heavily instrumented Ft. Cobb basin in western Oklahoma. The statistical model was developed using five rainfall events and subsequent streamflow observations. Similarly, we calibrated two additional models using the same rainfall events that are based on a) a conceptual understanding of infiltration and runoff mechanisms and b) a spatially distributed, physical description of runoff production. Following the calibration period, each model was evaluated on independent events, including an event with a 100-year return period.
[ paper (PDF) online ].
[ slides (PDF) online ].
In (37), we describe a process of computing scalar features from gridded values falling within the geographic and temporal extent of clusters identified at different scales through contiguity-enhanced K-Means clustering of texture features. We demonstrate the utility of the extracted properties through a decision tree that is capable of identifying the storm type based on just the automatically extracted properties.
[ paper (PDF) online ].
[ poster (PPT) online ].
In (38), we describe the adaptations done to
the quality-control algorithm for WSR0-88D data to enable
to work on C-band non-Doppler radars with a different
scanning strategy.
[ paper (PDF) online ].
[ poster (PPT) online ].
In (39), we describe the dataset for automated
storm type classification. The dataset is the result
of a clustering program run on multi-radar reflectivity
images and properties of those clusters computed by a set of severe
weather algorithms. The labels on the dataset,
created by manual identification, identify each cluster
as being one of four categories: supercells, convective
lines, pulse storms or non-organized cells.
We then provide an overview of submitted techniques,
compare and contrast the varying approaches and describe
how the techniques were validated and ranked. The ranking
was carried out using the True Skill Statistic, and
significance of differences was estimated through
permutation tests
[ paper (PDF) online ].
[ presentation slides (ppt) online ].
In (40), we describe our progress in identifying and tracking storms at multiple scales from satellite infrared (11-micron Band 4)
and visible (Band 1) channels.
Storms are identified by clustering the pixels
in the input images using spatial-contiguity-enhanced K-means
clustering. Identified clusters are then processed morphologically
to yield self-consistent storms.
Identified storms (at all the scales) are tracked using
a hybrid scheme that minimizes mean absolute error between frames of
the input sequence of images and then smoothed temporally using
Kalman filtering. This yields a grid of motion vectors at each
pixel in the spatial domain.
The motion vector estimated from the sequence is used to nowcast
the images. Comparison of the nowcasts with the observed values
at the corresponding time gives a measure of skill of the nowcast.
Statistical properties are extracted for each cluster. The extracted
properties are used as inputs to an automated decision tree training
algorithm to identify regions of overshooting tops.
Results and measures of skill are demonstrated on a sequence of
images from Oct. 12-13, 2001.
[ paper (PDF) online ].
[ presentation slides (ppt) online ].
In (41), we describe a method of creating such a coherent view. In polar coordinates, this involves creating a rapidly updated "virtual volume scan". The virtual volume scan is created by treating each of the phased array radar range
gates as "intelligent agents" that place themselves in the resulting polar grid, know how to collaborate with other agents to create optimal estimates of the radar values at each range gate of the virtual volume and know when they have either been superseded or are too old. The resulting virtual volume, created in real-time, is used by the downstream applications. This enables the downstream applications to work with a regularly spaced grid that is created at periodic intervals.
[ read PDF online ].
[ presentation slides (ppt) online ].
An update of WDSS-II activities since the first workshop. The talk gave a brief synopsis of various projects over the last 2 years.
The presentation slides are here:
[ read (ppt) online ].
In (42), we describe our approach to addressing the problem of creating good probabilistic forecasts when the entity to be forecast can move and morph. We formulate the tornado prediction problem to be one of estimating the probability of an event at a particular spatial location within a given time window. The technique involves clustering Doppler radar-derived fields such as low-level shear and reflectivity to form candidate regions. Assuming stationarity, the spatial probability distribution of this region T minutes ahead is estimated and combined with the probability that the candidate region becomes tornadic T minutes later. Using these two probabilities and the variability of the motion estimates, a spatio-temporal probability field is derived.
The neural network training required to correctly estimate the probabilities has not yet been developed. Therefore, this paper illustrates the underlying idea using fuzzy logic, storm half-life and motion variability.
[ read PDF online ].
In (43),
we describe the WDSS-II architecture used to achieve high
resolution radar data in real-time over the entire continental United
States. In the CONUS-wide real-time system that we describe here,
the radar products are rapidly updated every 2 minutes from elevation
scans as they arrive from any radar in the country. The resulting products
are at a resolution of approximately 1km x 1km x 1km. These products can be
disseminated as GeoTiff, NetCDF and/or Grib2 files for easy incorporation into
other decision-support and visualization systems.
[ read PDF online ].
[ presentation slides online (PPT) ].
In (44),
we describe a multi-sensor application that uses iso-therm levels from
the Rapid Update Cycle (RUC2) model, radar reflectivity data,
and cloud-to-ground lightning
data from the National Lightning Detection Network (NLDN)
to predict the onset of cloud-to-ground lightning.
The application uses a radial basis function (RBF)
to form a relation between past observed reflectivity
at various isotherm levels
to current cloud-to-ground lightning activity.
The RBF relationship matrix is constantly updated in real-time,
and used to predict the onset of cloud-to-ground activity
in the future based on current observations of radar reflectivity
at various isotherm levels.
[ read (MSWord) online ].
[ presentation slides (ppt) online ].
In (45), we describe the use of a neural network
to identify which of the circulations detected by the MDA are
tornadic.
[ read (pdf) online ].
[ presentation slides (ppt) online ].
In (46), we describe how we processed 3D lightning
mapping array (LMA) data for display in WDSS-II, including creating
lightning density grids and vertical and horizontal slices.
[ read (pdf) online ].
[ poster (pdf) online ].
In (47),
we describe an enhancement to the radar-only neural network (described
in (48) and (24)) that uses satellite data
and surface temperature with some time and space correction
to improve the quality control. This helps resolve issues with
biological targets, chaff and terrain-induced ground clutter.
[ read (pdf) online ].
We describe using a neural network with features derived from three
radar moments to clean up anamalous propagation (AP), ground clutter
(GC) and clear-air return from radar reflectivity data (48).
[ read (pdf) online ].
We describe the use of statistically derived hierarchical cluster
s of weather data to derive movement estimates from pairs of frames
in a time sequence (49).
[ read (pdf) online ].
A number of papers at the 2002 Severe Local Storms Conference
in San Antonio, Texas described aspects of WDSS-II and described
algorithms under development in this system.
A white paper, too long for the conference, that may be useful
for background on the Warning Decision Support System - Integrated
Information (WDSS-II).
[ read (pdf) online ].
Description of the algorithm developer's API for developing
meteorological algorithms of the future. This (50) was
the lead paper in the session.
[ read (pdf) online ].
[presentation slides].
Description of the K-Means clustering technique, as it relates
to the concept of algorithms that run on different sources (51).
[ read (pdf) online ].
[presentation slides].
A new concept (52) in accessing and visualizing weather radar data.
[ read (pdf) online ].
Describes a way of merging (creating a mosaic or mosaicking) multiple
sources of related data (such as radar reflectivity from weather radars,
TDWR, etc.) on a common three-dimensional grid as the data arrives,
i.e. virtual volumes of data from each source (54,53).
[ read (pdf) online ].
[presentation slides].
We describe a method (15) of multiscale storm identification,
computing motion estimates and making short range forecasts
from radar and satellite images.
[ read (pdf) online ].
Linear prediction of valid radar reflectivity values, combined with
run-length encoding of missing data values produces a compression rate
that exceeds currently used methods (55).
[ read (pdf) online ].
We apply our texture segmentation technique to elevation scans of
radar reflectivity data in (56).
[ read (pdf) online ].
In (57),
a hybrid method is introduced wherein we use
a segmentation-based approach to identify storms in images and arrange
the identified storms heirarchially. A genetic-algorithm method of
matching these segmented regions across frames is also introduced.
We show how this method can easily handle splits and merges of
storm cells. Since it deals with all scales, this method should,
theoretically, overcome the
drawbacks inherent in the cross-correlation and cell-tracking methods.
[ read (pdf) online ].
The conference presentation is also available:
[ read (pdf) online ].
A meteorological paper on the updraft algorithm was presented at the radar
conference (58) in Austin, '97. It is a
description of what the algorithm does rather than how it does it: You can read the entire paper online at
http://www.nssl.noaa.gov/wrd/swat/radcon97.lak.html
Next: Talks/seminar slides
Up: Research Publications
Previous: Papers at Engineering Conferences
V. Lakshmanan : valliappa.lakshmanan@noaa.gov