An overview of integrating deep learning methods with close-range hyperspectral imaging for agriculture

Abstract

Hyperspectral imaging combines spectroscopy with imaging, thus capturing both spectral and spatial features. This makes it a useful technology in several application areas such as remote sensing and smart agriculture. Extracting spatial-spectral information of objects-of-interest from hyperspectral images requires sophisticated computational methods. The last decade saw the rapid advancement of deep learning methods due to their superior automatic feature extraction capability from images, and hence it is no surprise that these methods have been adapted and used for hyperspectral image analysis. Yet, while deep learning methods have achieved some success for hyperspectral remote sensing, it has been less explored in close range (or proximal) hyperspectral imaging, which is likely because at this range, it is more akin to spectroscopy with spatial information, rather than the case of remote sensing, which is more akin to imaging with higher spectral resolution. Close-range HSI allows for fine-scale analysis of plant health, nutrient levels, disease detection, and crop quality, which is very important in precision agriculture. In light of the new computational methods in deep learning, this review article provides an in-depth analysis and comparisons of such methods when applied to proximal hyperspectral imagery, with a particular emphasis on unsolved challenges (e.g., limited availability of annotated datasets, the need for robust models under real-world conditions, and the integration of spatial and spectral information) and potential future research directions for agricultural applications. The review emphasizes the importance of further explorations and has provided recommended directions for future research that could elevate close-range hyperspectral imaging technology from research to industry use for smart agriculture applications.