A review on the combination of deep learning techniques with proximal hyperspectral images in agriculture

Abstract

Hyperspectral images can capture the spectral characteristics of surfaces and objects, providing a 2-D spacial component to the spectral profiles found in a given scene. There are several techniques that can be used to extract information from hyperspectral images, with deep learning becoming the preferred choice in the last decade due to its ability to implicitly extract features from images. The combination of hyperspectral images with deep learning has been extensively explored in the context of remote sensing (data collected by drones and satellites), which has generated an extensive literature that has been thoroughly analyzed in a series of reviews and surveys. The application of deep learning to proximal hyperspectral images is more recent, but there are already many articles dedicated to this objective, especially in the areas of agriculture and food science. Although significant progress has been made in just a few years, there are still many aspects that are not well understood and many problems that have not yet been overcome. This review aims at characterizing the current state of the art of deep learning applied to hyperspectral images captured at close range, focusing on the main challenges and research gaps that still need to be properly addressed. Some possible solutions and potential directions for future research are suggested, as an effort to bring the techniques developed in the academy closer to meeting the requirements found in practice. Only applications related to vegetable production and products were considered in this review, because applications related to animal farming and products of animal origin have their own particularities and the literature on the subject is not as extensive.