TWOPAC – A new approach for automated classification of satellite imagery

Kurzfassung

ABSTRACT: Land cover classification from satellite imagery provides base data for planning activities in several fields like integrated water resources management, land management, etc. The WISDOM project (www.wisdom.caf.dlr.de) aims for the implementation of a water-related information system to support planning activities within Vietnamese institutions. Reliable and reproducible land cover and land use maps are one of the main products which are provided through the WISDOM information system. Common image classification techniques often include high degree in manual operator interaction, for e.g. data preparation and sampling over a variety of software tools. Our approach aims to reduce these manual sequential processing steps. Increasing needs for automation of classification procedures result from the requirement of processing large amounts of data – either to cover large areas or to handle time series data. The introduced approach TWOPAC – Twinned object- and pixel-based automated classification chain – realizes pixel- and objectbased supervised classification of multi-sensor and multi-resolution satellite imagery. It basically supports management and processing of sample data as also the classification of earth observation data in either vector or raster form. The classification utilizes a large number of pixel- and object samples stored to a database allowing for multiple usages of those for training and validating of classifiers. With the C5.0, Maximum Likelihood Estimation, and Supported Vector Machines TWOPAC is currently supporting different modular classification methods. The software realizes OGC conform Web Processing Services which decreases the need for special commercial image classification software. The automated modular classification process chain is tested for several data sets from study areas in the Mekong Delta, and classifier stability and classification accuracy are analyzed. The method is considered to retrieve very good accuracy for stable and comparable classification results.