Back To School With Planet: Week 6 | See Change
This is the sixth installment of our Back to School with Planet series, a weekly update for K-8 students interested in learning more about the science we do with Earth and Space. Learn more about this series here.
You can find this entry in German here / Hier findet ihr den Blogpost auf Deutsch
You can find this entry in Spanish here. / Puedes encontrar este texto en español aquí.
The Earth is always changing. There are daily cycles, yearly cycles and some cycles that change the Earth forever. Cameras in space allow people on the ground to see that change in action, and use those changes to make decisions, plan where to make more changes and even protect the planet.
Daily Tides
If you’ve been to a beach, maybe you’ve seen that sometimes there’s a deep shoreline. Sometimes the water is far away from buildings, while other times the water comes close to the buildings. When you make a sand castle, you make it where the waves can’t reach. But if you stick around and watch the waves, you’ll realize later in the day that they’ll come up and take it! This is the daily dance where the moon pulls water to different tide levels through the day.
These two images were taken in the same place on the same day, and we can see changes in the water levels near the Great Barrier Reef.
Capturing a changing landscape
It’s hard to get a good look at an Artic ice shelf in person. Places like the Milne Ice Sheet are in a frozen sea, off the coast of places that are about as far north as you can get by land.
Understanding how ice changes is important as people study the relationship between climate, ocean levels and the creatures that feed in these icy places. A Planet Dove snapped an image of this crack in the ice, and past photos made it possible to compare and measure how the ice is changing.
Finding out when things change
Daily satellite imaging makes it possible to see before-and-after imagery of any place on Earth. But it’s too many images to compare every day! New techniques with coding make it possible for computers to inspect new images and tell us when change has happened.
Career Spotlight: Machine Learning
Staff Engineer, Machine Learning
Machine learning specialists like Planeteer Brad Neuberg build software that looks for changes. As the name describes, machine learning software examines real-life observations that you know and understand. Software learns patterns and trends about how to classify images based on that “training data” and uses it as a reference to identify what’s in brand-new images.
What kinds of changes can machine learning detect?
Potentially anything! This could include seeing changes in the natural environment or urban environments like roads and buildings.
How do you design a good change detector?
Much of the work in machine learning involves gathering your data, known as data curation. It’s similar to cooking a recipe: you want to make sure you have the right mix of ingredients and spices. For machine learning data, you want to make sure that your training dataset really represents what you care about for your change problem (your signal) and captures different sources of noise that might be seen (seasonal changes, different satellite angles, etc.). You revisit and update this dataset as you train your model, seeing what it does well and where it fails, adding “spices” of different training examples to help it learn better.
What are some of the limitations of modern-day machine learning?
Many machine learning systems need a large amount of training data, perhaps thousands or tens of thousands of labelled training examples, so we are limited to detecting changes with machine learning where we can collect enough data.
Machine learning models use this training data to understand what to focus on and what to ignore. Usually this training data is also labelled, which means that humans have pointed out where something has changed or where it has not changed. This can be a time-consuming process. Using labelled training data to train machine learning systems is called Supervised Machine Learning.
It is sometimes possible to skip the labeling and train machine learning systems without having labelled data. This is called Semi-Supervised Machine Learning. However, these methods are still being researched and don’t always perform as well as Supervised methods.
We are also limited by what satellites can actually see in terms of resolution, quality and revisit rate. For example, if we wanted to detect whales surfacing on the ocean and the change over time of whale populations, we would be limited by whether our satellites can actually “see” smaller whales surfacing. We would also be limited regarding weather (like clouds that block the view) and whether we image the ocean enough to confidently detect whale population changes.
What kinds of changes should the software ignore?
A machine learning model will produce relevant results known as “signal” and irrelevant observations called “noise.” The software should learn to separate noise from signal. These changes could be imaging noise from the atmosphere or noise from the sensor itself. Noise can also mean changes that we don’t care about, such as seasonal land changes going from fall to winter, or when it appears an object on the ground has moved because of differences in the satellites.
What kinds of skills and education are required to produce machine learning models?
Machine learning is a wonderful intersection of software engineering, applied math and a scientific, experimental mindset. Python is the most common computer language for machine learning and a great general purpose language to learn. Much of machine learning is a mix of some calculus, linear algebra and statistics.
What excites you most about the capabilities of change detection?
Planet has created an incredible fleet of satellites that can image the entire Earth daily. Unfortunately, this flood of data is too much for a human to process. How can we use machine learning to process all of this data, helping us monitor and take care of our changing Earth? For example, can we automatically detect illegal logging roads in the Amazon so we can prevent rainforest logging? These are the kinds of things that automatic change detection can provide.
What are some of the limitations of change detection?
You’re always limited by the actual quality and resolution of your satellite images, including how often image capture happens. Sometimes machine learning models can produce false signals. You need to make sure that doesn’t happen too often.
Activity: See Change Near You
This week, take a closer look at the world around you, share your changing world with others, and think about how you can see it change. Choose an object in your local world, and document its change over the week using whatever method you like best. For example, you might decide to take a photograph of a flower in your yard every day, or to take a careful look at the clouds outside at the same time every day and write a description of what you see.
Share your thoughts and pictures on social media to inspire others to see change too!
Send your change observations to us at backtoschool@planet.com for a chance to be featured in next week’s #backtoschoolwithplanet update!”
Where on Planet Earth?
Based on these clues, can you determine what this is a picture of?
I grow, and I grow, every day from fresh snow.
I am an ancient, icy river that shaped Argentina
I bring water to mountain lakes, but one disappeared!
Who am I? (Answer)
Further Inspiration
Reading (Grades K-3): Yara’s Tawari Tree by Yossi Lapid
Reading (Grades 4–8): Malala: My Story of Standing Up for Girls’ Rights by Malala Yousafzai
Movie to Watch (Rated PG): Hidden Figures
