Geospatial knowledge-based verification and improvement of GlobeLand30

0
IMAGE

Credit: ©Science China Press

Global land cover (GLC) data with fine spatial resolution and high quality are essential for global environment changes research, earth system modeling, management of resources and sustainable development planning. Assuring data product quality has been one of the major challenges for all of the operational GLC mapping projects.

Many large area land-cover mapping projects cannot deliver high quality data with single automated routines, although significant progress has been achieved in the area of automated remote-sensed image classification during the last twenty years. In particular, automated classification may cause significant classification errors when applied to 30-m land-cover mapping at a global scale. Identifying potential errors in the preliminary automated classification results through a suitable process of verifying and improving the results with a post-classification verification strategy has become a critical step for improving the quality of land-cover mapping results. Still, efficient tools are lacking for identifying and removing classification errors using geospatial knowledge.

In a recent study, a geospatial knowledge-based verification and improvement approach is developed and used for assuring the data quality of GlobeLand30. It consists of a set of geospatial knowledge-based verification rules and a group of web-based supporting tools.

Natural conditions, human activities and ecological environments affect the geospatial distribution and temporal transformation of land cover. Geospatial knowledge about land cover and its change are summarized by a combination of three different aspects: natural, cultural and temporal constraints. The verification rules are formulated to represent geospatial knowledge. Verification rules are formulated and represented using the so-called production-rule method or decision tree approach.

The web system is used to integrate heterogeneous and dispersed data resources (including primary Landsat-like images, various ancillary datasets and preliminary classification results) and external web services (such as Google Earth and Map World) (Figure 1). It provides a number of interactive tools to facilitate data-sharing and manipulation, such as geo-browsing (zoom in/out and pan), synchronized visualization (maps in two split windows for contrast), annotation (annotating sample, paper, photo, etc.), publication (publishing data service), etc.

The verification and improvement of GlobeLand30 is a collaborative process in which a group of project managers, quality inspectors and data operators work together is designed (Figure 2). With the support of the web system, the detection of potential classification errors and their modifications are accomplished in a collaborative manner. First, the project managers integrate the ancillary data in the web-based system. Then, they allocate the verification tasks and supervise the whole verification process. Second, quality inspectors verify the intermediate classification results to discover the potentially misclassified regions where spatial or temporal inconsistency may occur with the help of knowledge-based rules and ancillary data. They may use high-resolution images from integrated external services to identify and annotate the classification errors. These messages are published and sent to data operators for further modification.

The implementation of this knowledge-based approach has greatly improved the data quality of GlobeLand30 by identifying and removing classification errors. According to a third-party assessment, the overall accuracy of GlobeLand30-2010 reached 83.50% with a kappa coefficient of 0.78. In Italy, the overall accuracy of GlobeLand30 is better than 80% and the accuracy of water bodies of GlobeLand30 in Thessaly, Greece is 91.9%. All these indicate that the geospatial knowledge-based verification and improvement approach is feasible and reliable. It can also be used for other large scale land-cover mapping.

###

This work was funded by the National Science Foundation of China (Project #41231172), et al. The relevant paper is published in Science China Earth Sciences.

See the article: Zhang Wei Wei, Chen Jun, Liao An Ping, et al. Geospatial knowledge-based verification and improvement of GlobeLand30[J]. Science China Earth Sciences, 2016, 59(9): 1709-1719.

This article was published online (http://engine.scichina.com/doi/10.1007/s11430-016-5318-4)

Media Contact

Zhang Weiwei
[email protected]

http://zh.scichina.com/english/

Leave a Reply

%d bloggers like this: