Joint Classification of Hyperspectral and LiDAR Data Using Improved Local Contain Profile

Joint classification using multi-source remote sensing data has drawn increasing attention. For some complex surveyed scenes, relying on a single hyperspectral image (HSI) is not enough to meet the purpose of high-precision classification. Comparatively, light detection and ranging (LiDAR) data is rich in structure and elevation characteristics. To extract discriminative features, state-of-the-art extinction profile (EP) and local contain profile (LCP) have been investigated. However, EP is more sensitive to the external environment such as shadow occlusion, and the filtering strategy in LCP is prone to lose useful information when dealing with complex terrain scenes. Therefore, an improved LCP (ILCP) method is proposed to extract features from HSI and LiDAR data for joint classification. The proposed ILCP is more stable than EP, and the dimension of features extracted by ILCP is half of EP, which can avoid Hughes phenomenon. Compared to LCP, ILCP uses threshold-based filtering instead of extinction value, which retains more useful information for classification. Furthermore, feature-level fusion is applied to extracted features, and then the integrated features are input to the support vector machine (SVM) for final classification. In this paper, EP and LCP are also employed as comparison methods. Experimental results validated with one real multi-resource remote sensing data demonstrate that the proposed ILCP is superior to traditional methods.