Oarfish

Oarfish is an experiment to see how well deep learning works at classifying degree resolution all sky images from the Orville Wideband Imager running at the LWA stations.

Approach

Oarfish uses a two stage classifier to assign a quality score and label to an all sky image. The first stage is a binary classifier that is used to distinguish "good" data (data with low to moderate RFI that is free of instrumental problems) from "bad". The second stage takes this a step farther and tries to distinguish between:

• good images,
• medium RFI images - images where the brightest is comparable to that of "A team" soures,
• high RFI images - where the RFI is brigher than anything else in the sky,
• corrupted images that are likely caused by instrumental problems,
• images where the Sun is flaring, and
• images where Jupiter is bursting.

The first stage is good for separating images that can generally be used for science from those that cannot but it cannot be used to determine what the RFI environment is like beyond good vs. bad. The second stage is good for understanding why an image is bad but it also has trouble with separating no RFI from low RFI from medium RFI.

To overcome this oarfish introduces a "quality score" which combines the two classifier results into a single number. It take the binary classification confidence (downweighted by 25%) and combines it with the top two (by confidence) results from the multi-class stage to create a single number that ranges from 0 (bad) to 1 (good). This number is then mapped to a text label that is one of:

• good
• low RFI
• medium RFI
• high RFI
• bad

The division of quality score into bad to good is not uniform and the thresholds between the levels have been adjusted to match what a person might label a variety of images.

Training

The two models included with the library are trained using a collection of images pulled from the LWA1 PASI archive and the LWA-SV Orville archive¹. These images span the full range of time and frequency that Orville has been operating over at Sevilleta. For the binary classifier a training set of roughly 900 good and 900 bad images was used. The multi-class model was trained on:

• roughly 2200 RFI-free images,
• about 1600 low to medium RFI images,
• ~1000 high RFI images,
• a little under 1000 corrupted images,
• almost 260 images where the Sun was flaring, and
• almost 50 images where Jupiter was bursting.

The validation sets for both followed a 80/20 split for training/validation.

For processing the The Stokes I and |V| images were normalized to 0 to the 99.75-th percentile of the Stokes I image and resampled to a uniform size 256 by 256 pixels for pattern recognition. In addition, key features were extracted from the images that traced the A team, the Sun/Jupiter, and characterized the horizon and sky contrast.

Using

The easiest way to use oarfish is through the oarfish.data.MultiChannelDataset and oarfish.predict.DualModelPredictor classes. MultiChannelDataSet is a PyToach DataSet sub-class that prepares a collection of NumPy array for pattern recognition and astronomical feature extraction. DualModelPredictor wraps the model loading a prediction into a single object. It loads in the binary and multi-class models, runs the prediction on a DataSet, and returns a full set of metrics.

If you are working with Orville .oims files there is also a ZeroMQ-based server and client included with the scripts.

Footnotes

1. There were a handful of LWA-NA Orville images used but not enough to dramatically alter the training. ↩