Evaluation of transfer learning techniques with convolutional neural networks (CNNs) to detect the existence of roads in high-resolution aerial imagery

Abstract

Infrastructure detection and monitoring traditionally required manual identification of geospatial objects in aerial imagery but advances in deep learning and computer vision enabled the researchers in the field of remote sensing to successfully apply transfer learning from pretrained models on large-scale datasets for the task of geospatial object detection. However, they mostly focused on objects with clearly defined boundaries that are independent of the background (e.g. airports, airplanes, buildings, ships, etc.). What happens when we have to deal with more complicated, continuous objects like roads? In this paper we will review four of the best-known CNN architectures (VGGNet, Inception-V3, Xception, Inception-ResNet) and apply feature extraction and fine-tuning techniques to detect the existence of roads in aerial orthoimages divided in tiles of 256 × 256 pixels in size. We will evaluate each model´s performance on unseen test data using the accuracy metric and compare the results with those obtained by a CNN especially built for this purpose.