A Python and Flask based REST API that serves a Keras/TensorFlow Convolutional Neural Network (CNN) model trained to classify satellite image tiles into 17 different possible labels. This API currently interfaces with a React and Leaflet front-end.
When the user clicks on an area selected on the map, a POST request, containing the center latitude/longitude coordinate for the area selected, is sent to the REST API. The API then searches a PostgreSQL database for the file location of a satellite image tile containing the selected area. This image is then processed, in real time, through a Keras/TensorFlow ResNet50 model. This model makes a multilabel classification over 17 different labels returning a score between 0 and 1 for each label. The resultant scores are filtered based on a cutoff value, and then returned as JSON to the front-end.
The model was trained using the public dataset from Planet that was part of their Kaggle competition in 2017. This dataset consisted of ~42,000 image tiles of the amazon rainforest, all labeled. The main labels that appear in the current implementation are defined as the following:
| Label | Description |
|---|---|
| No Clouds | No clouds in the image |
| Primary | A segment of dense tree cover |
| Habitation | Any human homes or buildings |
| Agriculture | Any area of agriculture |
| Road | Any road within the image |
| Water | River or Lake |
There are many more labels which can be found here.
A ResNet architecture was chosen for the CNN due to it's fast inference time, good accuracy and smaller model size. See this paper for a comparison on all of these traits for the most common CNN architectures.