Zoltan Koppanyi Post Doctoral Researcher, The Ohio State University
Charles Toth Research Professor, The Ohio State University
Dorota Iwaszczuk Visiting Scholar, The Ohio State University; Post Doctoral Researcher, Technical University of Munich
Bing Zha Ph.D. Student, The Ohio State University
Alper Yilmaz Professor, The Ohio State University

14E

By now, modern photogrammetric workflows are fully automatized from image orientation to dense point cloud, mesh and DSM generation. However, extracting the semantic information of the images is still a challenge. Segmentation and segment classification are labor intensive, and therefore, the automation of this step of the photogrammetric workflow is of interest for researchers and practitioners. Recently, deep learning, in particular, convolutional neural networks (CNN) have become one of the promising approaches to tackle this problem. CNN has already achieved high accuracy for various image classification, segmentation and labeling tasks, however, current efforts are typically focused on segmentation of close-range, mobile or indoor imagery. The goal of this paper is to investigate the performance of an encoder-decoder convolutional neural network, called SegNet, on aerial images for image segmentation and classification. The encoder-decoder network structure allows for pixel-wise segmentation of an image, where all pixels are annotated with a label by the network. The performance analysis is conducted on the ISPRS’s aerial benchmark dataset, and includes the investigation of various hyper parameters, such as learning rate, input image size, the effect of training set size and pretrained weights on the segmentation and classification accuracy.

14:15 Investigation on the SegNet Deep Neural Network for Aerial Image Segmentation and Classification, Zoltan Koppanyi