Incomplete Multi-Source Feature Transfer for Hyperspectral Image Classification

Abstract

As a significant strategy to handle the small-sample problem in hyperspectral image (HSI) classification, cross-scene transfer learning generally assumes that the effective knowledge is transferred from one source scene to the target scene. However, in reality, there may be some situations where multiple source HSIs are needed, since one source may not cover the complete classes of the target image. In this paper, we propose an incomplete multi-source transfer learning method for HSI classification. In our framework, we first extract spectral-spatial features of source scenes and target scene. Then, we extend the maximum mean discrepancy (MMD) based cross-domain distribution distance to an incomplete multiple sources form, and build a multi-source locality preserved distribution alignment (MLPDA) model to learn transferable features for classification. Experimental results show that the proposed method can effectively transfer discriminative knowledge from several incomplete sources to improve the classification performance of target HSI.