Novel Deep-Learning-Based Spatial-Spectral Feature Extraction For Hyperspectral Remote Sensing Applications

Hyperspectral remote sensing presents a unique Big Data research paradigm through its rich data collected as hundreds of spectral bands which embodies vital spatial and spectral information about the underlying terrains. Typical hyperspectral data analysis methods are often based on spectral information. Although there has been prior efforts in literature for incorporation of spatial, spectral, contextual and other forms of information to improve the classification performance of hyperspectral data analysis, this additional information extraction and knowledge discovery process comes at the expense of increased computation and memory requirements. Therefore, the caveats of large scale data analysis such as increased computation, transmission and memory requirements presents a major impediment to efficient automation and classification performance of hyperspectral data analysis methods. In this respect, this paper presents a novel deep learning-based hyperspectral data analysis model, which provides an efficient means for automation and extraction of the spatial and spectral information present in the hyperspectral data compared to conventional spatial-or spectral information-only based methods. In this work, the concept of Gabor filtering is used for spatial feature extraction along with sparse random projections for computationally efficient spectral feature extraction and dimensionality reduction purposes. A convolutional neural network based supervised classification is then performed to validate the performance of the proposed method with respect to conventional spatial-spectral information extraction methods. Experimental results reveal that the proposed hyperspectral data analysis model outperforms the conventional spectral-spatial feature extraction techniques compared.