We've created a machine learning model designed to detect the presence of cloud streets (horizontal convective rolls) in satellite images. We leveraged NASA's Global Imagery Browse Services (GIBS) to collect these high resolution images, and used the data set provided in this challenges resources tab to train the model.

We also successfully created a clone of NASA's WorldView on a local server (https://spaceboys.xyz), which we call Nimbus Worldview, and created a new layer there which shows which of the ~12,800 images our model has detected cloud streets in. Our "Nimbus Horizontal Convective Rolls" layer provides an accessible, user-friendly interface to view these results.

Google's Tensorflow platform was the perfect start for our ML (machine learning) system, since it supports Convolutional Neural Networks (ideal for machine learning with analysis of visual imagery) out of the box. After configuration, we found the performance quite impressive - on a moderate gaming desktop (using a 2060 Super graphics card), our model ran through the cloud streets training data to the point of 80% accuracy in just under 30 minutes! In addition, the model takes less than 100 microseconds to evaluate an image (which is 1/100th of a millisecond)!

NASA's Worldview application displays layers by first requesting satellite imagery from GIBS, displaying those image tiles, and then by adding layers from separate information also obtained from GIBS. In Worldview Nimbus, all the same layers are available, however Nimbus also calls upon our own server (spaceboys.xyz) to overlay red pixels where cloud streets have been detected.

In summary, our ML model classifies each image tile it evaluates as "yes" or "no" for the presence of cloud streets. It then takes "yes" tiles, translates the row/column values into a set of coordinates compatible with Nimbus Worldview, and then feeds that information to our spaceboys.xyz server. When using Nimbus Worldview, each image tile is rendered before any layers - and when you zoom in, more tiles are displayed for a given 'piece' of the Earth's surface. When the Nimbus layer is applied, upon increasing the magnification of the map, each newly displayed tile is rendered, and Nimbus calls upon the server to determine if cloud streets have been detected in each new tile by our ML model - if they have been, it will draw an overlay (in real time) on that tile. This means we are not pre-rendering images for the overlay, but requesting, drawing, and rendering the presence of each part of our layer with no buffering each time the magnification increases, or the map is moved!

While, in this project, our model has been trained to detect cloud streets, the model could be given other data in order to detect (or even predict) different natural phenomena.