Automatic object­-oriented land cover classification of optical remote sensing data

Satellite images are an important source of environmental data and have proven to be highly effective in many applications. With an increasing number of remote sensing data users are witnessing a strong need for simple products that are delivered in near real time and without human intervention. Land cover maps provide an important insight into the state of environment and are of great importance in Earth sciences. In Slovenia, they are mainly produced by visual interpretation of aerial photographs, while in other countries satellite images are used as a primary data source.

We have identified three main research challenges in land cover production: under exploitation of satellite imagery, non­optimal use of segmentation algorithms, and non­optimal use of classification algorithms. Although the potential for the use of satellite imagery is well demonstrated, their actual exploitation has not increased in recent years. The main reasons for this, other than the price, are time­consuming and complex pre­processing, and a complicated classification procedure. Due to increase in spatial and spectral resolution of satellite imagery pixel­based classification has been replaced almost entirely by object­based classification.

The main aim of the project Automatic object-oriented land cover classification of optical remote sensing data was to develop an automatic procedure for geometric and radiometric correction of satellite images, and production of a land cover map (vector layer) that can be used directly by end­users in their analyses using geographic information systems (GIS).

In the project we developed a robust automatic geometric correction procedure that is based on rational polynomial coefficients (RPCs) (WP1). The study of segmentation was devoted to existing algorithms and analysing the influence of different input data characteristics. We also introduced texture into segmentation (WP2). In analysing different classification methods, we looked for significant attributes for optimal within class separation and training samples determination on a multi­level scale (WP3). In the final phase we integrated procedures developed in previous phases into a workflow (WP4). The processing is semi-automatic and exploits a graphical user interface (GUI). Classification tool is still in a prototype phase and new functionalities are being added.

The project J2-6777 Automatic object-oriented land cover classification of optical remote sensing data  was financially supported by the Slovenian Research Agency.