作者: Huang, X (Huang, Xin); Han, XP (Han, Xiaopeng); Zhang, LP (Zhang, Liangpei); Gong, JY (Gong, Jianya); Liao, WZ (Liao, Wenzhi); Benediktsson, JA (Benediktsson, Jon Atli)
|
摘要: Differential morphological profiles (DMPs) are widely used for the spatial/structural feature extraction and classification of remote sensing images. They can be regarded as the shape spectrum, depicting the response of the image structures related to different scales and sizes of the structural elements (SEs). DMPs are defined as the difference of morphological profiles (MPs) between consecutive scales. However, traditional DMPs can ignore discriminative information for features that are across the scales in the profiles. To solve this problem, we propose scale-span differential profiles, i.e., generalized DMPs (GDMPs), to obtain the entire differential profiles. GDMPs can describe the complete shape spectrum and measure the difference between arbitrary scales, which is more appropriate for representing the multiscale characteristics and complex landscapes of remote sensing image scenes. Subsequently, the random forest (RF) classifier is applied to interpret GDMPs considering its robustness for high-dimensional data and ability of evaluating the importance of variables. Meanwhile, the RF "out-of-bag" error can be used to quantify the importance of each channel of GDMPs and select the most discriminative information in the entire profiles. Experiments conducted on three well-known hyperspectral data sets as well as an additional World View-2 data are used to validate the effectiveness of GDMPs compared to the traditional DMPs. The results are promising as GDMPs can significantly outperform the traditional one, as it is capable of adequately exploring the multiscale morphological information.
|
