A recent project saw me dive quickly into convolutional neural networks and I implemented a proof of concept within a week.

Detecting buildings in satellite imagery with deep-learning

• Developed and trained Convolutional Neural Networks (CNNs) using Keras with TensorFlow backend to detect buildings in a highly unbalanced set of proprietary satellite images and generated an interactive visualization of the holdout image classification outcomes (~93% average recall & accuracy) using geospatial information embedded in the geotiffs;

• Further explored image segmentation using computer vision techniques to delineate building outlines in images;

• CNNs were trained on AWS GPU enabled instance and the interactive visualizations employed Geopandas, GDAL, Matplotlib and Folium Python modules

Python scripting files are found in docs