A Study of Spatial-Spectral Information Fusion Methods in the Artificial Neural Network Paradigm for Hyperspectral Data Analysis in Swarm Robotics Applications

Object classification using remote sensing data is a proponent for applications across varied domains such as public safety, environmental science, and agriculture. As remote sensing has evolved towards more complex and drone mountable sensors that are able to capture hyperspectral data including more than 100 spectral bands, equivalently techniques required to process and classify the data have grown more computationally complex as well. The need for extensive computational and memory requirements for data analysis limits the real-time applicability of hyperspectral data collected using limited-resource and portable platforms such as drones in surveillance and swarm robotic applications. As a consequence, the objective of this study is to explore methods to implement Artificial Neural Networks (ANN) in a parsimonious manner by using dimensionality reduction, different network architectures, and regularization techniques with respect to its practical applicability in the context of swarm robotics intelligence. The results indicate our proposed spatial-spectral information fusion ANN architecture comprised of only a depth of a few layers, but with many neurons per layer outperformed the models that did not use dimensionality reduction or used too many deep layers to achieve hyperspectral image classification.