Remote Sensing Image Classification Based On Object Information

Compared with the low or middle resolution image, the high resolution Remote Sensing image has richer structure and texture information. The result of traditional statistical classification technology based on pixels'spectrum containing lots of "Pepper and Salt" noises has a very low accuracy. And the phenomena of "the same thing with different spectrums" and "different things with the same spectrum" cannot be distinguished. Those are all the inevitable limitations of the traditional statistical classification algorithm based on pixels'spectrum.The idea of object-oriented is introduced into the information extraction technology from the high resolution image in this thesis. This technology produces homogeneous image objects through segmentation technology, and provides a way to analyze object's features. At last it carries out the information extraction by fuzzy classification. An algorithm of image segmentation, based on color and spatial information, is proposed in this thesis. The initial regions are hierarchically merged based on the region distance defined by color and spatial information. And a criterion is proposed to decide the termination of the merging process. Finally, the accuracy of the segmentation result is as high as the result made by eCogniton,"the No.1 Enterprise Image Intelligence Processing Software". When extracting the object's features in this thesis, we not only consider the spectral difference between the regions, and shape measures, texture measures, position measures as well. Once using these added object's measures, we could distinguish different covers on the ground with same spectral.Taking a typical region of Tongzhou District, Beijing City as the test area, a case study on SPOT-5 image classification with object-oriented approach is carried out at the end of this thesis. In order to verify the accuracy of object-oriented classification, a comparison between this approach and classical classification approaches has been carried out. The case study shows that the application of object-oriented approach on high resolution satellite image classification not only can improve the overall classification accuracy, but also has more semantic information, reduces the"Pepper and Salt"noises, and overcomes such phenomenon as different things with the same spectrum.