Classification of Citrus Crops Using Satellite Multispectral Imagery and Deep Neural Network

This paper presents a Deep Neural Network classifier for use in the identification of citrus fruit crops using satellite multispectral images. It aims at classifying pixels and regions in 4-band multispectral images acquired by the remote sensing satellite platform GeoSat-2, identifying each pixel or region as belonging either to bare soil, orange, mandarin or grapefruit trees or crops. The system relies on the reflectance responses coming from the blue, green, red, and near-infrared bands of previously labelled images. These reflectance responses together with derived data from them, such as the computation of the Difference Vegetation Index (DVI), the Normalised Difference Vegetation Index (NDVI), the Excess Green Index (ExGI), and statistic metrics of the 5-Nearest Neighbours to a testing pixel are used as features to train a multilayer deep neural network classifier. The trained classifier is then tested on multispectral images of other citrus crops, providing a pixel-wise classification accuracy of 90.92%, increasing this metric to 98.08% when further applying a voting-based conditional discriminator to classify crop regions.