Multi-Scale Change Detection Research of Remotely Sensed Big Data in CyberGIS

Abstract:

Big remotely sensed data, the heterogeneity of satellite platforms and file formats along with increasing volumes and velocities, offers new types of analyses. This makes big remotely sensed data a good candidate for CyberGIS, the aim of which is to enable knowledge discovery of big data in the cloud. We apply CyberGIS to feature-based multi-scale land use/cover change (LUCC) detection. There have been attempts to do multi-scale LUCC. However, studies were done with small data and could not consider the mismatch between multi-scale analysis and computational scale. They have yet to consider the possibilities for scalar research across numerous temporal and spatial scales afforded by big data, especially if we want to advance beyond pixel-based analysis and also reduce preprocessing requirements.

We create a geospatial cyberinfrastructure (GCI) to handle multi-spatio-temporal scale change detection. We first clarify different meanings of scale in CyberGIS and LUCC to derive a feature scope layer in the GCI based on Stommel modelling. Our analysis layer contains a multi-scale segmentation–based method based on normalized cut image segmentation and wavelet-based image scaling algorithms. Our computer resource utilization layer uses Wang and Armstrong’s (2009) method for mainly for memory, I/O and CPU time. Our case is urban-rural change detection in the Greater Montreal Area (5 time periods, 2006-2012, 100 virtual machines), 36,000km² and varying from 0.6m to 38m resolution. We present a ground truthed accuracy assessment of a change matrix that is composed of 6 feature classes at 12 different spatio-temporal scales, and the performance of the change detection GCI for multi-scale LUCC study. The GCI allows us to extract and coordinate different types of changes by varying spatio-temporal scales from the big imagery datasets.