Identifying Urban Canopy Coverage from Satellite Imagery Using Convolutional Neural Networks

The availability of high resolution satellite imagery offers a compelling opportunity for the utilisation of state-of-the-art deep learning techniques in the applications of remote sensing. This research investigates the application of different Convolution Neural Network (CNN) architectures for pixel-level segmentation of canopy coverage in urban areas. The performance of two established patch-based CNN architectures (LeNet and a pre-trained VGG16) and two encoder-decoder architectures (a simple 4-layer convolutional encoder-decoder and Unet) was compared using two datasets (a large set of images of the Geerman town of Vaihingen and smaller set of the US city of Denver). Results show that the patch-based methods outperform the encoder-decoder methods. It is also shown that pre-training is only effective with the smaller dataset.