A PRE-OPERATIONAL SYSTEM FOR OBJECT BASED IMAGE CLASSIFICATION FEEDING POSTGIS DATABASES AUTOMATICALLY
Object-based and geo-spatial analysis are efficient techniques in the field of object recognition in high resolution remote sensing images. By working at a higher level of abstraction, the first allows to capture structures and features that would not be handled properly by low-level pixel-based or neighbourhood-based methods. Because of the intrinsic geo-spatial nature of remote sensing data, it looks straightforward to benefit from the flexible tools provided by the latter, such as GIS systems and databases. Yet extracting pertinent information from raw products and exporting it to a GIS database is a challenging task, because of the variability and volume of the data to process. A single standard very high resolution product contains billions of pixels with features that can not be easily extracted and generalized. We propose a semi-automatic system to perform this task, by incrementally extracting information from the raw data using well-known performing techniques from the literature : pan-sharpening, segmentation, object-based features extraction, vectorization and SVM training with active learning, and final massive export to a GIS database in which geo-spatial queries can easily be performed. Our system is in a pre-operational state, since it is able to process full Quickbird products automatically apart from the learning step, which could be done only one for a representative set of scenes. As such, it could very well be plugged to the output of a ground segment and automatically produce value-added products. Finally, it is also highly flexible, because each step can be extended and tuned to specific user needs. The paper will focus on two points of view, one describing the whole approach in terms of image processing chain, the other highlighting the software architecture. Results from each step will be given, from segmentation to geo-spatial queries.
Julien Michel - CNES
Jordi Inglada - CNES/CESBIO
Julien Malik - CS-SI
Manuel Grizonnet - CNES