Semi-supervised hyperspectral manifold learning for regression

Regression based on hyperspectral remote sensing data contains two-fold complications, i.e., lack of labeled data and difficulty in collecting quantitative ground-truth. In this paper, we propose semi-supervised subspace learning methods for regression based on a generalized eigenvalue problem. The methods exploit abundant unlabeled data for low-dimensional subspace learning. Quantitative target values are replaced by ordinal values that can be easily acquired in comparison with accurate quantitative ground-truth. The subspace learning methods are further expanded into nonlinear manifold learning methods by the kernel trick. The methods are applied to estimation problems of growth-state-related properties of rice based on hyprspectral remote sensing data.