paper
(2018_TGRS)___Hyperspectral Image Classification With Imbalanced Data Based on Orthogonal Complement Subspace Projection
There are mainly two steps to deal with imbalanced data, including increasing minority and reducing majority.
Step 1. For minority classes, i.e., small-size classes, some samples are originally and randomly generated by gradient-constraint to maintain same variation tendency. Specifically, is the mean sample vector of S and the gradient vector can be calculated by , where . The sampled samples should has the same variation tendency to the gradient. In other words, adjacent bands should vary with the variation in same tending, increasing or decreasing. The sampled samples are denoted by R.
Then these samples are slected by OCSP, where an orthogonal complementary subspace is constructed as , where the last S# is the pseudoinverse of S. Project R to the orthogonal complementary subspace by . Then those samples yeilding small projections will be selected.
Step 2. For majority classes, some distinctive samples are selected to reduce the number of samples and represent with-in class variations at the same time. OCSP is also employed here and samples are selected with largest projection on orthogonal subspace.