What is a polar low?
Polar lows are medium sized mesocyclones that only occur near the poles and are usually intense and short lived. Due to their short lived nature, they are quite difficult to identify which can be problematic given that they can produce harsh weather near coastal areas. Polar lows can either be comma-shaped or spiral-shaped. In the video to the right, peer mentor Jack Larcome explains how polar lows are identified and how they differ from regular storms.
Our Data
Images of polar lows are collected manually by FIRE-CC students using NASA Worldview. Satellite images taken by the Aqua and Terra satellites are searched. When a polar low is found, the following are recorded: date, longitude, latitude, the satellite the image was taken by, and a screenshot is taken of the polar low. The snapshot and metadata are stored in our shared Google Drive.
In order to provide the most accurate results, the data must run through multiple quality checkers. This ensures our machine learning algorithms provide accurate and consistent results.
We wanted to visualize the geographic distribution of collected polar lows in our database. With this information, we would be able to better compare and contrast our data to literature.
A team from the 2020 cohort - Jeffrey Greiner, Mili Gupta, Julia Herrera, John Le used analytic methods in python to show the spatial distribution of the Fall 2019 and Spring 2020 collections of polar lows on a scatter plot.
Looking at the Antarctic region, the team concluded that there was very little correlation when it came to common locations of polar lows and there are not as many clustered locations. The results of these projections show a consistent and reoccurring development of polar lows in and around both the Arctic and Antarctic regions.
This effort was revisited after the 2021 cohort added more images to our database. Plotted here are the spatial distributions of polar lows for each hemisphere colored by confidence level. You can see there is no obvious dependence on confidence level in the spatial distribution. You can find the code that created these plots here.
In order to analyze when polar lows occur, and to see where there may be possible gaps in our dataset, a team from the 2020 cohort - Isabella Neel, Ryan Moore, Rohan Panickar, Rebecca Marchesi -- set out to display the temporal frequency distribution of our polar low dataset.
Using analytic methods and DataFrames in python, the team was able to display the temporal frequency to the left.
Polar Low Quality Assurance
The FIRE-CC stream has been collecting polar low images from NASA worldview since 2019 and collected nearly 1,000 snapshots! With a plethora of images, it is important to stringently enforce protocols so the images are of the highest accuracy and quality.
PRM Jack Larcome enforced the following quality checks on each image:
Polar low is poleward of 40 degrees north and 40 degrees south
The image is in a png format with 1000 x 1000 pixels and has a 1km resolution
Polar low entry has an image present
Each polar low entry that didn't meet one of these quality checks would be flagged and modified to meet them.
Motivation
An automated algorithm able to identify polar lows would be an incredibly useful tool. It could identify the dangerous and destructive storms before they make landfall, giving communitites and economies the information they need to protect life and property. To accomplish this goal, we attempted to use machine learning models to identify polar lows in satellite photographs.
There are two types of machine learning: supervised and unsupervised.
Supervised Learning
Supervised machine learning involves training a computer model. You give the model a set of training images and labels identifying what the images are of. Whenever the model is given an unlabelled image, it can select a label for it based on which of the training data the image is most similar to.
We trained the free supervised learning model Teachable Machine by Google to identify polar lows. In order to have a strong set of training data to teach our Teachable Machine, we took screenshots from NASA world view of satellite images that we knew were polar lows, and others that we knew weren't polar lows.
Classification Errors
The confusion matrix to the right demonstrates the accuracy of one of our Teachable Machines. To summarize, the Machine was very good at identifying true polar lows, correctly labeling 100% of the polar images it was given. However, it also labeled some images as containing polar lows even when they didn't.
A team of 2021 cohorts-- Cameron Willette, Sharath Kannan, Percy Olken-Hunt, and Kenny Cheung-- focused their research on identifying why this might be. They compared the training images and how the students collected them in order to find flaws. They found that many of the training images were misclassified due to human error (a non-polar low being incorrectly labeled as a polar low by a student) or due to technical error (distorted images). This means the Machine was trained incorrectly. Their research can be found here.
Unsupervised Learning
Unsupervised machine learning is similar to supervised machine learning, with one key difference: the computer model isn't given any training data. Instead, the model is given a large amount of unlabelled images all at once, which it sorts into categories on its own. The computer doesn't know what it should call these categories-- it isn't told that some images are polar lows and others are not-- but it still learns to differentiate between the two.
Finding similar images
2020 cohorts Aastha Senjalia, Mili Gupta, Jack Larcome, and Ryan Moore attempted to use an unsupervised learning model to identify polar lows. The team used VGG-16, an image identification model created by Oxford scientists that can be implemented in Python. The model sorted similar images into groups, a few of which can be seen left. Notice how the polar lows in group 3 all have a distinctive spiral shape.
Plotting principle components
A team from the 2022 cohort used VGG16 to extract features from images, then reduced the dimensionality of those feature vectors using PCA. Each dot on this plot represents an image. Red are labeled as non polar lows by observers, blue as low confidence polar lows, and black as high confidence polar lows. You can see that just among the first 2 principle components of the feature vector there is decent separation of polar lows (blue and black) and non polar lows). If you would like to see more of their research results you can find them here.
Further Reading