Automated analysis of overhead imagery for habitat segmentation

Abstract

The overall objective of our project is to be able to classify the evolution of land usage since the advent of aerial imagery. In practice our aim is to bring automatic habitat classification to the level achieved by a human expert performing a fine-scale classification of habitat at resolutions covering from hedges and lake to fields, pastures or forest. Relying on the recent progresses in machine learning algorithm and in particular convolutional neural networks trained using deep learning (e.g. SegNet, DeepLab), our approach trains a machine to segment an overhead imagery into a dozen of expert-specified land use classes.