| Abstrakt: |
Hyperspectral remote sensing images (HSRSI) comprise hundreds of adjacent channels with rich spectral-spatial signatures, making it possible to discriminate earth objects. Thus, it has contributed to its wide use in urban mapping, environmental management, and crop analysis. To completely take advantage of the abovementioned uses, it often requires the identification of the class of each pixel, which at times faces scarcity or limited availability of labeled training samples. Hence, it is an open challenge to achieve higher interpretation accuracies when processing such increased spectral-spatial resolution imagery. To this end, we, therefore, propose a 3D inception spectral-spatial network model (3D-ISSN). First, we adopt the principal component analysis (PCA) for dimensional reduction of spectral channels that are very much correlated while preserving the desirable information, second, we fuse the hierarchical spectral-spatial related features extracted for CNN1-D and CNN2-D, with our model for learning and achieving the correct classification with a softmax regression classifier. We verified our model through experiments carried out on two HSRSI data sets, namely Botswana (BT) and Kennedy Space Center (KSC), and compared our classification accuracies with the state-of-the-art (SOTA) methods. [ABSTRACT FROM AUTHOR] |
