Abstract
Image classification and interpretation are greatly aided through the use of image segmentation. Within the field of environmental remote sensing, image segmentation aims to identify regions of unique or dominant ground cover from their attributes such as spectral signature, texture and context. However, many approaches are not scalable for national mapping programmes due to limits in the size of images that can be processed. Therefore, we present a scalable segmentation algorithm, which is seeded using k-means and provides support for a minimum mapping unit through an innovative iterative elimination process. The algorithm has also been demonstrated for the segmentation of time series datasets capturing both the intra-image variation and change regions. The quality of the segmentation results was assessed by comparison with reference segments along with statistics on the inter- and intra-segment spectral variation. The technique is computationally scalable and is being actively used within the national land cover mapping programme for New Zealand. Additionally, 30-m continental mosaics of Landsat and ALOS-PALSAR have been segmented for Australia in support of national forest height and cover mapping. The algorithm has also been made freely available within the open source Remote Sensing and GIS software Library (RSGISLib).
Highlights
The classification and analysis of remotely-sensed optical data has become a key technology for the mapping and monitoring of land cover [1,2]
We segmented a range of optical imagery, and the resulting number of segments, their size and their spectral variation were assessed to ascertain whether the segmentation results were fit for the purpose of land cover classification, including change
It is readily scalable to large areas, such as nations or regions (e.g., [55]), which is desirable for preventing hard boundaries on tiles. It uses a small number of parameters, which may be consistently used across a large range of geographic areas and data types
Summary
The classification and analysis of remotely-sensed optical data has become a key technology for the mapping and monitoring of land cover [1,2] Such analysis has been performed on a per pixel basis, but over the last 20 years, there has been a significant movement to embrace context and segment-based classifications [3] due to observed improvements in classification accuracy [3,4,5,6,7]. High-resolution remotely-sensed data such as RapidEye, Sentinel 1 and 2, SPOT (4–5) and Landsat (TM, ETM+, OLI), provide the majority of data. These data typically have a resolution of 5–30 m, thereby providing sufficient detail for monitoring land cover, such as forestry, agriculture and grasslands. Where errors occur within the segmentation result, a small amount of over-segmentation is preferable to retain classification accuracy [9]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
