A composed statistical pattern recognition and geosciences analysis approach for segmentation-based remotely sensed imagery classification

Statistical Pattern Recognition is one of the basic techniques for land use/land cover classification of remote sensing image. Focusing on establishing a method to integrate statistical pattern recognition with geosciences' analysis, this paper proposed a segmentation-based remotely sensed imagery classification method, called ISODATA-Geosciences Imagery Classification (ISOGIC). The result from case application of TM image in Yanchi County shows that ISOGIC can distinguish large land types via ISODATA clustering, while geosciences knowledge including topography, soil type and vegetation type, etc. can significantly improve the accuracy of cultivated land, different types of forests, different coverage grassland, sandy land and saline and alkaline land. The average classification accuracy is up to 87.9%. Compared to Pixel-based maximum likelihood method, the segmentation-based remotely sensed imagery classification method can effectively resolve problems such as same object with different spectrums and different object with same spectrums.