Author(s): , , ,
Institution(s): 1. National Astronomical Observatories of China , 2. Wuhan University
The massive photometric data collected from multiple large-scale sky surveys offers significant opportunities for measuring distances of many celestial objects by photometric redshifts zphot in a wide coverage of the sky. However, catastrophic failure, an unsolved problem for a long time, exists in the current photometric redshift estimation approaches (such as k-nearest-neighbor). In this paper, we propose a novel two-stage approach by integration of k-nearest-neighbor (KNN) and support vector machine (SVM) methods together. In the first stage, we apply KNN algorithm on photometric data and estimate their corresponding zphot. By analysis, we observe two dense regions with catastrophic failure, one in the range of zphot [0.1,1.1], the other in the range of zphot [1.5,2.5]. In the second stage, we map the photometric multiband input pattern of points falling into the two ranges from original attribute space into high dimensional feature space by Gaussian kernel function in SVM. In the high dimensional feature space, many bad estimation points resulted from catastrophic failure by using simple Euclidean distance computation in KNN can be identified by classification hyperplane SVM and further be applied correction. Experimental results based on SDSS data for quasars showed that the two-stage fusion approach can significantly mitigate catastrophic failure and improve the estimation accuracy of photometric redshift.