We used Google Mobility data combined with Sentinel-1 (SAR) and Planet imagery (optical) to monitor a worldwide decrease in mobility.
Google's Mobility data allows for a high-level overview shown by a decrease in the use of transit stations around the world; while using satellite imagery enables us to delve deeper into specific locations.
Having access to high resolution Planet imagery meant that we could develop a deep-learning based "labelling" tool to find the location of airplanes; thus opening the possibility of rapidly mapping the activity of an airport. Combining this type of insight with other types of data such as flight routes could enable having a deeper understanding of the impacts of the pandemic on an international level.
Sentinel-1 on the other hand is a great tool for time-series analysis and allows us to have a temporal understanding on how the organization and activity of airports changes over a longer period of time.
While it is only possible to have a limited understanding on these behaviours in a couple of days, we hope that this project can showcase to those unfamiliar with Earth Observation how it could be used in a world-changing context like this; but also how combining different types of data yields a better understanding of these complex phenomena.
Our inspiration came from our experience with EO data, and the situation that we are faced with - which has led to, in part, a loss in our ability or willingness to travel.
Both of us are familiar with Earth Observation data, and know the usefulness of the Copernicus Program from ESA. Sentinel-1 was used to get a temporal understanding of the impacts of the pandemic on airports (using Google Earth Engine, not requiring us to locally download images and do the analysis on our hardware). SAR imagery is particularly well fit for these time-series as it is not dependent on good weather, and allows to monitor an area on a 5 to 6-day revisit time, enabling change tracking over long periods of time.
Planet imagery, which is at a higher spatial resolution than Sentinel, allowed us to have a deeper view of how airports were reacting, like how they were storing their airplanes.
We used python libraries for deep learning but also EO like PyTorch, GeoPandas and Rasterio as well as Google Earth Engine for the analysis. Apart from Sentinel-1 images (done on Google Earth Engine), all the analysis was done locally on our personal laptops, in Python.
The lack of a good segmentation or object detection dataset of airplanes pushed us to think a bit differently and turn a 2-class image dataset (20x20 images either containing a plane or something else) into an "object detection" model by making a sliding window with different offsets (or strides). This obviously is far from perfect (white buildings were a very common false positive) but allows us to demonstrate the idea nonetheless.
Both of us are used to the capability but also the limitations of Earth Observation data and tools, we would like to emphasize that this was done in 2 days, and while some interesting insights were found during this time, there are a lot of short-comings. To name a simple example, Google Mobility Reports are a tricky source of data to work with as the collection methods are unknown and might not be comparable in all parts of the world. The seriousness of the pandemic pushes us to re-state that this is in no way a complete study. Unfortunately data science, epidemiology and even Earth Observation have been challenged a lot lately and we would like to state that this is closer to a side project (which it is) than to any serious study.
https://docs.google.com/presentation/d/1tY9_F4WjKq6-NNWKG3O4ctxRzj68QTvyHARdGzFWq9w/edit?usp=sharing