Classification of apple leaf conditions in hyper-spectral images for diagnosis of Marssonina blotch using mRMR and deep neural network

Abstract

In general, in order to reduce the number of bands in hyper-spectral images, transformed features such as principal component analysis (PCA) cannot directly provide information indicating which raw spectral bands are essential, even though they provide reduced dimensionality for the consecutive analysis. This paper proposes minimum redundancy and maximum relevance (mRMR) feature selection techniques to directly choose essential raw bands from hyper-spectral images. In addition a deep neural network is suggested to classify the hyper-spectral data with reduced dimension, which consists of a convolutional neural network (CNN) followed by a fully connected network (FCN). The CNN extracts meaningful features like PCA, but in a nonlinear, supervised manner, and the FCN classifies the six apple leaf conditions including normal, young, malnutrition, early and late stages of apple Marssonina blotch (AMB), and background. Experimentally, we found five essential spectral bands using the mRMR techniques (777.24 nm, 547.77 nm, 474.32 nm, 859.45 nm, and 735.85 nm), which proved to have better accuracy than RGB image for the classification using the deep neural network. The proposed scheme is applicable for band reduction to obtain the efficient multispectral sensor system in the sense that only several essential spectral bands are chosen, and more accurate diagnosis of plant diseases is possible by reducing the complexity involved in hyper-spectral images.