A Hybrid Object-based/Pixel-based Classification Approach to Detect Geophysical Phenomena

Geophysical phenomena are observable events with spatiotemporal characteristics. These phenomena can have spatial extent and shape but the intensities are generally not homogeneous within this spatial extent. These phenomena also evolve, grow and perish over time. Therefore, developing a robust detection algorithm for geophysical phenomena is difficult and challenging. This paper presents a hybrid object-based/pixelbased classification methodology for shape-based geophysical phenomena detection from image data. The target phenomena for this study are the frontal systems in the atmospheric modelgenerated wind field data. The methodology is comprised of two levels of data mining. At the pixel-level, the image is soft classified pixel by pixel to calculate the probability that it is the component of the geophysical phenomena. The K-means clustering algorithm is used to create this soft classification. At the object-level mining, a hierarchical thresholding technique is coupled with a Gaussian Bayes classifier to optimally segment the individual regions of interest. At each hierarchical level, the probability image is segmented and shape factors are calculated for the segmented regions. The optimal threshold for a region is the one that produces the maximum likelihood from a Gaussian Bayes classifier based on the shape factors of the region. The object level mining is followed by post processing to filter false signatures. Experimental results show that this methodology effectively detects the frontal systems in the model data and the detected regions are in good agreement with the ones identified by the domain experts. This hybrid methodology can be applied to detect other geophysical phenomena in science datasets.