THE PRACTICAL PRINCIPLES OF QUALITY EVALUATION OF NEURAL CLASSIFIERS OBJECT RECOGNITION PRODUCT ON MULTI- SPECTRAL HIGH RESOLUTION SATELLITE IMAGES USING GEOINFORMATION TECHNOLOGIES Y.Gambarova Institute for Aerospace Informatics National Aerospace Agency, Baku, Azerbaijan

The principles of evaluation of neural-classifiers object recognition results are presented and proved using geo- information data processing tools. The job was done for solving of real task on defining distribution area of rare vegetation community existing in the IKONOS satellite high resolution multi-spectral image. Its proved that using of geo- information technologies visualization and coincide data analysis capabilities is not only useful tool for classifiers product quality evaluation but it allows us organize and implement the whole cycle of this product verification and reception procedures as well. The visual analysis of classifiers products presenting in thematic raster images, is shown in details.

Fragment of IKONOS satellite Orthorectified Imagery

Initially 12 types of rare vegetation communities and soil were defined that - according to ecologists opinion – are indicator of antropogeneous impact on environment in the region being studied.

The Initial classification scheme - 12 vegetation communities and soil types. Class numberFull name of vegetation communities and soil types Class 1Chal meadow / reedbed wetland Class 2Chal meadow / Tamarix scrub (Tamarix) Class 3Coastal zone semi desert Class 4Phragmites australis reedbed wetland (Phragmaties australis) Class 5Salsola ericoides Class 6Salsola nodulosa Class 7Salsola nodulosa / Artemesia lerchiana Class 8Salsola Nodulosa /Grasses Class 9Semi desert vegetation, Kalidium caspicum (Kalidium caspicum ) Class 10Semi desert scrub alhagi dominated (Alhagi pseudalhagi) Class 11Bare ground Class 12Salsola nodulosa / Bare ground

VISUAL ANALYSIS OF THE RESULTS OF THE CLASSIFICATION a) On the Optimized classification schemeb) On the Initial classification scheme Classification results: non-classified pixels are represented in the black color