|
Automation of image classification is a challenge to the image interpretation community. One of the most time consuming task is certainly the collection of training data. The introduction especially of low cost Global Positioning System (GPS) receivers and the higher accuracy of the GPS signal after turning-off the Selective Availability has enhanced the ease and versatility of spatial data acquisition. Has also made the approaches by which it is integrated with GIS and remote sensing data more flexible.
The emphasis of this paper is to present a method for improving the training data collection for classification purposes using remotely sensed imagery as well as various data sources combined in a GIS.
Firstly, a methodology for defining and describing training areas is demonstrated. The training data were stored in a vector data base (shape files) by using the geometry of the land parcels of the test site. Secondly, in addition to a conventional field mapping approach, an advanced GPS based field mapping methodology was used to collect new training data. Within this new approach single point information of the target ground truth class were collected along the roads in this test area. In this step, the following attributes were recorded: ID, left or right of the street and biotope class.
The goal for this approach is that one single person should handle the field mapping while driving in a car. The implementation of this approach is performed in ArcPad 6.03 and Application Builder from ESRI. The standard version of ArcPad was modified so that a one hand collection of training data is possible. After the field survey, the results were used within Erdas Imagine (version 8.7). In our approach all "left" points were moved to the adjacent left field - "orthogonal" to the street. All "right" points were shifted to the adjacent right field. Now those moved points were used as a seed pixel in a Euclidian distance algorithm to automatically derive new training sets. In conclusion it can be stated that the proposed training collecting method and its implementation have been proved to be a very valuable and reliable method for image classification purposes.
|
