Hyperspectral Image Classification Using Semi-supervised Random Forest
In this paper, a hyperspectral image classification technique is proposed using semi-supervised random forest (SSRF). Robust node splitting in the random forest requires enormous training data, which is scarce in remote sensing applications. In order to overcome this drawback, we propose utilizing unlabeled data in conjunction with labeled data to assist the splitting process. Moreover, in order to tackle the curse of dimensionality associated with a hyperspectral image, we explore nonnegative matrix factorization (NMF) to remove redundant information. Experimental results confirm the efficacy of the proposed method.
KeywordsHyperspectral imaging Semi-supervised learning Random forest Nonnegative matrix factorization
The authors thank Prof. Gamba for providing the University of Pavia dataset.
- 1.Mather PM, Koch M (2010) Computer processing of remotely-sensed images, 4th edn. Wiley, New YorkGoogle Scholar
- 8.Zhang T, Oles F (2000) A probability analysis on the value of unlabeled data for classification problems. In: 17th International conference on machine learning. ACM Press, California, pp 1191–1198Google Scholar
- 9.Lee DD, Seung S (2000) Algorithms for non-negative matrix factorization. In: Neural information processing systems. MIT Press, Denver, pp 556–562Google Scholar
- 11.Liu X, Song M, Tao D, Liu Z, Zhang L, Chen C, Bu J (2013) Semi-supervised node splitting for random forest construction. In: Conference on computer vision and pattern recognition. IEEE Press, Oregon, pp 492–499Google Scholar
- 13.Criminisi A, Shotton J, Konukoglu E (2011) Decision forests for classification, regression, density estimation, mainfold learning and semi-supervised learning. Technical report, Microsoft ResearchGoogle Scholar