20
Region-based Querying of Solar Data Using Descriptor Signatures Authors: Juan M. Banda 1 , Chang Liu 1 and Rafal A. Angryk 2 1 Montana State University, Bozeman, MT 2 Georgia State University, Atlanta, GA 1

Region-based Querying of Solar Data Using Descriptor Signatures

  • Upload
    cael

  • View
    18

  • Download
    0

Embed Size (px)

DESCRIPTION

Region-based Querying of Solar Data Using Descriptor Signatures. Authors: Juan M. Banda 1 , Chang Liu 1 and Rafal A. Angryk 2 1 Montana State University, Bozeman, MT 2 Georgia State University, Atlanta, GA. Introduction. - PowerPoint PPT Presentation

Citation preview

Page 1: Region-based Querying of Solar Data Using Descriptor Signatures

Region-based Querying of Solar Data UsingDescriptor Signatures

Authors: Juan M. Banda1, Chang Liu1 and Rafal A. Angryk2

1Montana State University, Bozeman, MT 2Georgia State University, Atlanta, GA

1

Page 2: Region-based Querying of Solar Data Using Descriptor Signatures

IntroductionIn this work, we present our first results of our experiments to prove region-based querying capabilities to our existing Solar Dynamics Observatory (SDO) content-based image-retrieval (CBIR) system

2

Page 3: Region-based Querying of Solar Data Using Descriptor Signatures

Motivation• Current SDO CBIR (http://cbsir.cs.montana.edu/sdocbir)

system returns images based on full-image similarity

• We have too many similar images returned (one image every 10 seconds produces plenty of close matches)

• In a practical scenario a researcher is only interested in a particular region of the image only

• Query SDO images from sources of multiple repositories

3

Page 4: Region-based Querying of Solar Data Using Descriptor Signatures

Our Approach• Use our existing image parameters extracted from the

solar images to create a 10-bin descriptor signature per region of interest.

• Create a testing and training dataset based on HEK (Heliophysics Event Knowledge base) reported events

• Generate the nearest-neighbor list of each event using different metrics based on the signature descriptors

• Validate retrieval results with classification algorithms

4

Page 5: Region-based Querying of Solar Data Using Descriptor Signatures

Image Parameter Extraction

Image data becomes 64x64 cells of ten image parameters each

5

Page 6: Region-based Querying of Solar Data Using Descriptor Signatures

Solar Event DatasetBased on the dataset introduced by Schuh et al. [1], we selected the following solar events

[1] M.A. Schuh, R.A. Angryk, K.G. Pillai, J.M. Banda, and P.C.H. Martens. A large-scale solar image dataset with labeled event regions. In 20th IEEE Int. Conf. on Image Processing (ICIP), pages 4349–4353, 2013.

6

Page 7: Region-based Querying of Solar Data Using Descriptor Signatures

Descriptor signature generation• Find intersecting cells with

region of interest.

• Average all image parameter values individually per region.

• Descriptor is a 10-bin histogram-like structure with each bin representing the average value of each parameter in the region of interest

7

Page 8: Region-based Querying of Solar Data Using Descriptor Signatures

Distance measures and classifiers used

8

Classifiers: Naïve Bayes, C 4.5 and Support Vector Machines

Page 9: Region-based Querying of Solar Data Using Descriptor Signatures

Experiments

• S1 - 1-to-all comparison within the same type of events / wavelength combination

• S2 - 1-to-all comparison within the same type of events

• S3 - 1-to-all comparison of all signatures against each other (all events and wavelengths)

9

Page 10: Region-based Querying of Solar Data Using Descriptor Signatures

Experiment - S1

• In this experiment we expect to identify which wavelength / distance measure combinations are ideal to retrieve certain events, with our signature-based approach, and how this matches with what the FFT modules have determined over the past

• The main aspect of this scenario is observing how we can improve the separation of one event / wavelength combination by using different distance measures

10

Page 11: Region-based Querying of Solar Data Using Descriptor Signatures

Results – S1 (1)

11

Note: On the X and Y axis we have each event arranged by their timestamp from earliest to latest. The diagonal of this plot should always be blue since the distance from each event to itself is zero. Blue indicates more similar and red least similar

Dissimilarity plots for wavelength 171 - flare (FL) event with: a) Euclidean distance, b) Cosine distance

Page 12: Region-based Querying of Solar Data Using Descriptor Signatures

Results – S1 (2)

12

Note: On the X and Y axis we have each event arranged by their timestamp from earliest to latest. The diagonal of this plot should always be blue since the distance from each event to itself is zero. Blue indicates more similar and red least similar

Dissimilarity plots for wavelengths (a) 131 and (b) 171 for the flare (FL) event. Both plots represent KLD measure

Page 13: Region-based Querying of Solar Data Using Descriptor Signatures

Experiment - S2• Instead of using just one wavelength to find events (as

on S1), we now try to identify them in a mixed wavelength scenario

• Here we want to try to validate our visual results, the FFT module results, or find new and interesting combinations

13

Page 14: Region-based Querying of Solar Data Using Descriptor Signatures

Results – S2

14

Page 15: Region-based Querying of Solar Data Using Descriptor Signatures

Experiment – S3• Simulating a more realistic retrieval scenario, we

combined all wavelengths and events

• We will not rely on any FFT information for this experiment and then see how our results match or can be used with the previous experiments S1 and S2

15

Page 16: Region-based Querying of Solar Data Using Descriptor Signatures

Results – S3

16

Page 17: Region-based Querying of Solar Data Using Descriptor Signatures

Independent Evaluation Results based on Classification

17

Page 18: Region-based Querying of Solar Data Using Descriptor Signatures

Overall conclusions• By following the traditional wavelength / event pairings

we get solid retrieval precision (75% for all events on average)

• We still need to rely on some of our expert knowledge (FFT modules) – particularly for sigmoids

• With our approach the retrieval calculations can now be done in real time due to the compactness of descriptor signatures

18

Page 19: Region-based Querying of Solar Data Using Descriptor Signatures

Future/Current Work• Add more different types of events

• Better class balancing using a larger dataset

• Deploy functionality on our CBIR system for researchers to use

• Test the descriptor signature approach in other domains (astrophysics) and with other parameters

19

Page 20: Region-based Querying of Solar Data Using Descriptor Signatures

Questions?

Thank you for your time!

20