A Multi-Dimensional Deep Hierarchical Approach Towards Aerial Hyperspectral Image Classification

Aerial Hyperspectral Image Classification has established itself as an integral part of remote sensing research by providing a viable approach for deriving accurate inferences on land compositions, similar region segregations and crop type identifications, based on the analysis of inherent spectral-spatial information present in the observed sensor data. Automated classifiers powered by deep learning algorithms have been extant but tend to consider only spectral or spatial information for the feature derivation process. The interdependencies between the spectral and spatial features play a crucial role in determining the effectiveness of the feature extraction process and hence, must not be ignored. The application of 3D CNN for the classification will solve the problem by providing joint representations of spectral spatial vectors. However the high computational complexities involved in the 3D convolutional architecture and its underperformance in environments with low texture variations are the primary challenges to its mainstream implementation. This paper proposes MDA-HSI, a dual model integrated voting network for hyperspectral image classification. The first model employs a 3D-2D hierarchical structure to provide the spectral spatial joint features along with the abstract level spatial representation of 2D CNNs. The second model is a 3D CNN where the performance increase is provided by incorporating overlapping patches on the original hyperspectral cube to create feature maps over multiple contiguous bands. The performance of the MDA-HSI was tested with benchmark datasets like University of Pavia, Salinas and Indian Pines to draw a comparative analysis with existing methods and was proved to be significantly more efficient.