Geo-Semantic Segmentation with TerraibNet, Atrous Convolution and Graphical Model for Human-level Landscape Understanding

Authors: Xiran Zhou*, Arizona State University, Wenwen Li, Arizona State University, Jun Liu, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Topics: Remote Sensing, Geographic Information Science and Systems, Landscape
Keywords: Geo-semantic segmentation, deep geomorphometric autoencoder, deep convolutional neural network, atrous convolution, conditional random fields
Session Type: Paper
Day: 4/13/2018
Start / End Time: 1:20 PM / 3:00 PM
Room: Grand Ballroom A, Astor, 2nd Floor
Presentation File: No File Uploaded

Semantic segmentation refers to a key task of computer vision which aims to identifying and understanding image content at pixel level with end-to-end learning of deep convolutional neural network (DCNN). Currently, few approaches have been created to enable semantic segmentation on landscape with DCNN from remote sensing data. Unlike the objects in digital photos, the characteristics and boundary of landscape always include the visible features of places, and the elevation variation of natural landform, which makes the in-depth use of visible features and elevation-related indices as the crucial concerns for implementing semantic segmentation on landscape with deep learning techniques. In this paper presentation, we report our investigation on exploiting DCNN to create geo-semantic segmentation for identifying landscape, based on spatial and global constraints over the multi-dimensional space jointing satellite image and DEM. We propose a deep geomorphometric autoencoder for automatedly detecting geomorphometric features from DEM. To avoid the reduction of spatial resolution by max-pooling, average-pooling and downsampling, we use Atrous Spatial Pyramid Pooling (ASPP) to filter satellite image over multiple scales. Fully-connected conditional random fields (CRFs) is used for constructing the details of fine edges, which has proved to be highly more effective than skip-layer and “hyper-column” features in addressing invariance of spatial transform in classification. We hope our work can pave the way forwards for the advance of semantic segmentation on landscape and human-level landscape understanding by remote sensing data.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login