Optimal Spatial-Spectral Input For Real-Time Hyperspectral Image Classification

Abstract

Heavy metal contamination in soil is of great danger for the environment as well as for the human health. This is owed to their hazardous toxicity alongside with their ability to be easily captured by crops and plants. Detection of soil contamination is a major step that aids in the soil rehabilitation process. Our goal in this project is to explore the potential of using hyperspectral imaging and deep Learning techniques to identify contamination in real time. The large size of the hyperspectral image and the relatively limited resources in onboard processors make the real-time classi- fication a challenging task. In this work, we propose a novel approach for real-time hyperspectral image classification using optimal spatial-spectral input. The optimal input consisted of the main pixel with two of its spatial neighbors. Two novel deep learning models based on the optimized three pixel method were developed. The first is a Deep Neural Network (DNN) model focused on fast online classification and the second is a Recurrent Neural Network (RNN) model focused on offline classification with enhanced accuracy. These models were evaluated using four datasets, three agricultural datasets, and the Sydney Bridge dataset. The DNN model achieved a maximum accuracy of 97% with an inference speed of 83,665 pixels per second while the RNN model achieved a maximum accuracy of 99% with an inference speed of 25,000 pixel per second. When compared, the DNN model is more suited for quick real-time applications while the RNN model is more suited for applications where accuracy is critical. Furthermore, in terms of latency, our approach maximized the preprocessing speed compared to other spatial-spectral methods. The attained fast speed in preprocessing and inference is applicable for Real-time applications.