Assimilation of PFISR Data Using Support Vector Regression and Ground Based Camera Constraints
Abstract
In order to best interpret the information gained from multipoint in situ measurements, a Support Vector Regression algorithm is being developed to interpret the data collected from the instruments in the context of ground observations (such as those from camera or radar array). The idea behind SVR is to construct the simplest function that models the data with the least squared error, subject to constraints given by the user. Constraints can be brought into the algorithm from other data sources or from models. As is often the case with data, a perfect solution to such a problem may be impossible, thus 'slack' may be introduced to control how closely the model adheres to the data. The algorithm employs kernels, and chooses radial basis functions as an appropriate kernel. The current SVR code can take input data as one to three dimensional scalars or vectors, and may also include time. External data can be incorporated and assimilated into a model of the environment. Regions of minimal and maximal values are allowed to relax to the sample average (or a user-supplied model) on size and time scales determined by user input, known as feature sizes. These feature sizes can vary for each degree of freedom if the user desires. The user may also select weights for each data point, if it is desirable to weight parts of the data differently. In order to test the algorithm, Poker Flat Incoherent Scatter Radar (PFISR) and MICA sounding rocket data are being used as sample data. The PFISR data consists of many beams, each with multiple ranges. In addition to analyzing the radar data as it stands, the algorithm is being used to simulate data from a localized ionospheric swarm of Cubesats using existing PFISR data. The sample points of the radar at one altitude slice can serve as surrogates for satellites in a cubeswarm. The number of beams of the PFISR radar can then be used to see what the algorithm would output for a swarm of similar size. By using PFISR data in the 15-beam to 45-beam modes, a cubeswarm size of 32 can be judged for auroral analysis purposes. The output of the algorithm is then compared to all sky camera data to determine the accuracy of the prediction. In addition, the camera data can be used to constrain the SVR fitting. The MICA data serve as additional test input for the algorithm, providing in situ data for comparison.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2013
- Bibcode:
- 2013AGUFMSA11A1915C
- Keywords:
-
- 1952 INFORMATICS Modeling;
- 1910 INFORMATICS Data assimilation;
- integration and fusion;
- 1906 INFORMATICS Computational models;
- algorithms;
- 2407 IONOSPHERE Auroral ionosphere