A Transfer Learning based CNN approach for Classification of Horticulture plantations using Hyperspectral Images

Abstract

Hyperspectral images (HSIs) are satellite images that provide spectral and spatial detail of a given region. This makes them uniquely suitable to classify objects in the scene. Classification of Hyperspectral images can be efficiently performed using the Convolutional Neural Network (CNN) in Machine Learning. In this research, a framework is proposed that leverages Transfer Learning and CNN to classify crop distributions of Horticulture Plantations. The Hyperspectral dataset consists of images and known labels, also known as groundtruth. However, some of the HSIs are unlabelled due to the lack of groundtruth available for the same. Hence, the proposed method adopts the Transfer Learning technique to overcome this. The model was trained on a publicly available and labelled hyperspectral dataset. This was then tested on the field samples of Chikkaballapur district of Karnataka, India which was provided by the Indian Space Research Organisation (ISRO). The CNN built leverages both the spectral and spatial correlations of the HSIs. Due to the amount of detail in HSIs, they are fed in as patches into the convolutional layers of the network. The diverse information provided by these images is exploited by deploying a three-dimensional kernel. This joint representation of both spectral and spatial information provides higher discriminating power, thus allowing a more accurate classification of the crop distributions in the field. The experimental results of this method prove that feeding images as patches trains the CNN better and applying Transfer Learning has a more generic and wider scope.