Investigation of RGB to HSI Conversion Methods for Early Plant Disease Detection Using Hierarchical Synthesis Convolutional Neural Networks

Abstract

An early detection of disease can save the plant. One of the ways is by using eye-observation, which is time-consuming. Having a machine learning technology that can automate early detection would benefit modern and conventional farming. This study emphasizes the review of Hyperspectral Image (HSI) reconstruction using the Hierarchical Synthesis Convolutional Neural Networks (HSCNN) based method in early plant disease detection. Capturing hundreds of spectral bands during image acquisition enables the HSI capturing devices to provide more detailed information. Detection of disease with Red Green Blue (RGB) images needs to be done when it shows a notable spot or sign. However, the disease can be spotted with the correct range of spectral bands on HSI before a notable spot or sign is shown. The usage of HSI image is significantly important as it is rich in information and properties needed for image detection. Although HSI device is significantly important in early plant disease detection, the devices are expensive and require specialized hardware and expertise. Thus, reconstructing the Reg Green Blue (RGB) image to HSI is required. This research implemented two types of HSCNN-based methods, Densed network (HSCNN-D) and Rectified Linear Unit network (HSCNN-R), for HSI reconstructions. The results show that HSCNN-D outperformed the HSCNN-R with less Mean Relative Absolute Error (MRAE) of 2.15%.