Vang Identifying Poster CWWEWI Trondheim mai 2011

Vang Identifying Poster CWWEWI Trondheim mai 2011
Identifying false alarms and bird tracks in a full scale
radar tracks database using clustering algorithms and
SQL Server 2008 Analysis Services
Roald Vang, Roel May & Frank Hanssen
Norwegian Institute for Nature Research (NINA), NO-7047 Trondheim, Norway
All data from the MERLIN Avian Radar System from April 2008 until March
2011 has been processed and automatically stored in a SQL Server 2008
database. This database, however, contains both bird tracks and false
alarms. By studying each track drawn out on a map and studying its
signature, it is possible to make an educated guess whether a track is a
bird or false alarm. However, due to the huge number of tracks (April
2008 – March 2011; horizontal database: 130 million track points ~ 45
GB), it is impossible to do this job manually, and a too time-consuming
task to work with the entire database in tools like Excel, SPSS, R, etc. We
therefore wanted to take advantage of the powerful quad-core processor
server which the database resides on, and develop an automated method
for filtering the radar-data.
Since the start of the project, about 2.000 bird track segments have been
ground-truthed manually (visually confirmed) within the wind-power
plant. These tracks have been used to verify the accuracy of the models
created during our work.
4. We defined the splits between these reclassified clusters based on their
signature using a decision tree model. This resulted in a misclassification
error rate of 4 % (14,959 of 385,116) for birds, and 2 % (19 of 888)
(excluding vehicles and other non-bird tracks; track length > 4) for the
ground-truthed tracks.
Data mining tools used
Microsoft SQL Server 2008
SQL Server 2008 Analysis Services
Microsoft Clustering Algorithm
Statistical programme R version 2.10.1
Data mining steps
1. Create flattened dataset of the entire database to use as basis. A
subset of 484,088 tracks was used for testing purposes. For each track
a spatial column (Latitude/Longitude) was generated and proportions
near roads (<20m) and within radial clutter areas were calculated.
2. For each track average, variance and delta values of parameters
deemed biological or radar-technological relevant were calculated
(speed, heading, turning angle, track length, reflectivity, target area
and shape parameters, etc.). Inclusion of wind speed and direction,
and precipitation.
3. The tracks in the test-dataset were clustered using the Microsoft
Clustering Algorithm. The Microsoft Clustering algorithm is a
segmentation algorithm that uses iterative techniques to group cases
in a dataset into clusters that contain similar characteristics. The
algorithm first identifies relationships in a dataset and generates a
series of clusters based on those relationships.
The figure shows the
most significant
parameters and their
threshold values.
The chosen model was implemented on the entire tracks database and each
track was classified as bird, vehicle (i.e. targets following a road) or false
The various clusters in the
figure are natural
groupings of the tracks
based on the chosen
Figure showing all bird
tracks (blue) and false
alarms (red) per month
from April 2008 until
March 2011.
4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3
4. Each cluster was classified as signifying either birds or false alarms
(i.e. precipitation, interference, outliers) in a semi-quantitative manner
by comparing each cluster’s signature (e.g. long straight tracks are
likely birds; highly varying track clusters may signify false alarms). Our
ground-truthed tracks helped us classify the tracks in the different data
mining clusters.
Cluster 9
Cluster 8
Cluster 7
Cluster 10
Cluster 11
Cluster 1
14 %
Cluster 2
21 %
Cluster 6
22 %
Ground-truthed data divided in
Cluster 5
13 %
Cluster 4
Cluster 3
10 %
Monthly changes in the
number of bird tracks by
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF