Toward the Use of Deep Learning for Topographic Feature Extraction from High Resolution Optical Satellite Imagery

This paper introduces the exploitation of a convolutional neural network for the extraction of topographic features from high-resolution optical satellite imagery. A UNET based model was trained for seven feature classes of roads, buildings, waterbodies using two 3-band (RGB) images for a study site in Kingston (Canada). The trained model's accuracy was evaluated on eight tiles of 8000×8000 pixels using a confusion matrix, the overall accuracy and kappa. The results show overall accuracy varying between 90 % and 99 % and kappa varying between 0.48 and 0.98, with five of the eight tiles being over 0.85. The model generally produced accurate predictions, except for commercial and industrial buildings and for unpaved roads, which were under represented in the training data. The project provided perspective for the development of a training database for topographic feature extraction using deep learning and for expansion to the national scale.