Close Range Spectral Imaging for Disease Detection in Plants Using Autonomous Platforms: a Review on Recent Studies

A short introduction to the spectral imaging (SI) of plants along with a comprehensive overview of the recent research works related to disease detection in plants using autonomous phenotyping platforms is provided. Key benefits and challenges of SI for plant disease detection on robotic platforms are highlighted. SI is becoming a potential tool for autonomous platforms for non-destructive plant assessment. This is because it can provide information on the plant pigments such as chlorophylls, anthocyanins and carotenoids and supports quantification of biochemical parameters such as sugars, proteins, different nutrients, water and fat content. A plant suffering from diseases will exhibit different physicochemical parameters compared with a healthy plant, allowing the SI to capture those differences as a function of reflected or absorbed light. Potential of SI to non-destructively capture physicochemical parameters in plants makes it a key technique to support disease detection on autonomous platforms. SI can be broadly used for crop disease detection by quantification of physicochemical changes in the plants.


Introduction
Worldwide, plant diseases lead to major yield and quality losses in agriculture production systems [1]. Diseases are mainly caused by pathogenic microorganisms such as fungi, viruses, bacteria, mycoplasmas and protozoa [2]. Typically, diseases start with the introduction of a vector or infested plant material at a location, which in time spreads to neighbouring plants leading to significant damage, even before it is visible to humans. For detection of pathogens, different molecular and serological methods, such as enzyme-linked immunosorbent assay (ELISA) and polymerase chain reaction (PCR), can be used to target specific proteins and deoxyribose nucleic acid (DNA) sequences, respectively.
These molecular and serological methods provide a highly specific detection of pathogens; however, they are destructive and limited to the laboratory and require highly skilled manpower. Despite such detection being of high interest to scientific research towards understanding the plant-pathogen interactions, these methods cannot be used in situ for real-time detection of diseases in plants before the symptoms are visible. Furthermore, these destructive methods cannot be used to follow the progress of the disease or the effect of crop protection compounds that are being used to control the disease spread. Early detection of disease or disease symptoms using nondestructive sensors is of high interest as it can directly benefit the agricultural systems by reducing associated losses [3][4][5].
Early non-destructive disease detection in plants is a key emerging topic [6]. To accomplish this task, autonomous platforms are increasingly being implemented [7,8]. An autonomous agricultural platform in field conditions can be understood as a robotic vehicle with a range of sensors integrated to What Is Spectral Imaging? SI combines traditional imaging with spectroscopy. Unlike colour red, green and blue (RGB) imaging SI refers to a group of techniques which captures more than just the red, green and blue wavelength bands. SI can be explored in different parts of electromagnetic radiation (EMR) ranging from the highenergy X-rays to the very low-energy microwaves [11]. A typical SI setup is composed of three main components: a light source, a spectrograph or filter system, and a detector for image generation. Based on the spectral range to be explored, the components are selected to meet the selectivity and sensitivity. The images generated by SI are three dimensional (3D) with two spatial (x × y) and one spectral dimension (z). Particularly, in the spectral range of 400-2500 nm, the data is highly collinear and requires advanced multi-variate statistical tools to extract and model patterns. A better understanding of SI and data analysis can be found in recent reviews [12••, 13].
What Does Spectral Imaging Detect in Diseased Plants?
Plants, unlike animals, lack an adaptive immune system limiting their ability to adapt to diseases based on the memories from past infections [14]. However, due to a highly evolved genome, plants possess a wide range of resistance activities which can counteract the disease-causing pathogens. The resistance activities are activated as soon as the plant receives a signal about coming in contact with a potential vector. Plant defence includes morphological and structural changes at the cell wall, with a massive and directed reorganisation within the host cells; changes in gene expression and general metabolism [15,16].
At first, the symptoms can be identified locally, where the vector interacted with the host tissue: early physiological and structural changes reflect a deviation in photosynthetic activity. Local symptoms might include changes in concentrations of photosynthetic pigments such as chlorophylls, changes in water content and an increase in plant defence-response metabolites. Changes in cell structure at the site of infection can also be noted utilising microscopic techniques. SI in the spectral range of 400-2500 nm is of high interest, as it can provide access to plant biochemical parameters. Specifically, the visible (VIS) part (400-700 nm) provides information on the plant pigments, such as chlorophylls, anthocyanins and carotenoids, and the near-infrared (NIR) (700-2500 nm) supports quantification of biochemical parameters, such as sugars, proteins, water and fat contents, and cell structure of plant tissues [12••, 17, 18]. SI captures this information by measuring the percent of light being reflected/absorbed at specific wavelengths. For example, chlorophyll absorption peaks at 465 and 665 nm, while water absorbs at approximately 1400 and 1900 nm. With a change in plant physiological status, the presence of these specific compounds will change, influencing the absorption in relevant spectral bands. Diseased tissue exhibits clear differences in physicochemical properties compared with a healthy tissue, which allows the SI to capture those differences in the function of reflected or absorbed light (400-2500 nm). Therefore, the information captured by SI can be extracted to develop models for disease detection and severity quantification.

Recent Applications
A large number of applications can be found that utilise SI for disease detection (Lowe et al., 2017) [19]. However, there exist a limited number of applications, where SI has been integrated with automated agricultural platforms. A summary of recent applications of SI for disease detection (past 5 years) is provided in Table 1. A separate column is presented to highlight if the study utilised an autonomous platform. A study excluding an autonomous platform can be understood as a future direction for the implementation of the diseasedetection technique in autonomous platforms. Cucumber 946 nm to 2016 nm PLS-DA, least square S-SVM [29] Recent studies show that automated SI for plant disease detection is used either in a field or in a greenhouse setting. In the field, there are three main automation approaches being implemented: first is a fixed rail-guided system, where the mobile platform travels over the plant lines and continuously captures the images. Such a system was successfully used by Gerrit  for the tulip break virus (TBV) detection in the standing plants. The second type of field system includes a platform mounted on autonomous tractors. A tractormountable system was successfully implemented by [10•] for potato Y virus detection utilising a measurement box mounted in front of the tractor, carrying a spectral imager, protection from ambient lighting and an embedded PC. The third type of field systems is autonomous vehicles with integrated sensors, scouting the fields. Such a 4-wheel system with mountable spectral sensors was utilised by [20••] to detect anthracnose in strawberry plants. However, the application utilised a point spectrometer rather than SI. In the greenhouse scenario, the automation involves either automated movements of plants to the sensor compartments (via tracks or rails) or the movement of sensors to the plants for data acquisition. In recent work, [21••] utilised an automated greenhouse for detecting powdery mildew in Barley, where an automated sensor positioning system travelled amongst the plant pots for imaging.

Discussion and Outlook
A range of applications related to the use of SI in the autonomous platform can be found; however, it can be understood from Table 1 that for the imaging applications, the spectral range used was mainly in the VNIR range (~400-1000 nm). There are two reasons for this. First, spectral cameras in the complete spectral range (400-2500 nm) are not available in the market. Sensing in the VNIR (400-1000 nm) and the SWIR (1000-2500 nm) range requires different electronic detectors, meaning multiple cameras need to be purchased and applied. Second, the cost of SWIR is almost 5 times compared with the VNIR sensors, making the VNIR the preferred choice for integration with the autonomous setups. Furthermore, VNIR range has the benefit that it captures Vegetation indexes and spherical k-means [31] the signal of photosynthetic pigments, which is present only in the spectral range of about 400-700 nm, as well as the 3rd overtones of proteins, fats and moisture present in the range of 700-1000 nm. Due to the possibility of SI to capture a wide range of the physicochemical parameter, SI can be used for untargeted plants disease detection, thus not looking for specific pathogens but rather capturing the overall symptoms or changes. This has a major benefit over destructive chemical methods which are focussed on identifying specific pathogens responsible for the damage. The main benefit of SI is its ability to non-destructively capture the spatially distributed spectral properties of plants which can be used to localise the spots of diseases in initial stages which might be missed by destructive analysis. The other major benefit of SI is the ability to perform a real-time detection during acquisition. Real-time detection can be performed by pre-training a model on a small calibration set and then deploying the model onboard an autonomous platform.
A major difficulty with the implementation of SI is the illumination effects caused by the interaction of light with the complex geometry of plants and their surroundings [12••]. Such an interaction of light causes scattering effects which may mask the real spectral responses. These effects should be corrected for before further data processing. Another challenge is the integration of white reference with the autonomous setup. White reference is required to calculate plant reflectance. Proper positioning of white reference with respect to the plants can reduce major differences in the illumination. However, there is no single solution and the positioning needs to be adapted for different types of plants and their growth stages.
Until now, most of the SI utilised line-scan cameras; however, new snapshot SI systems are emerging in the market which opens possibilities for easy implementation of SI to autonomous setups. In a technical perspective, the future direction should be towards an integration of multiple sensors with autonomous setups, such as the two spectral cameras (VNIR and SWIR) and the 3D-shaped sensors. Another step is their information fusion to deal with the illumination effects. Another major trend is the application of advance concepts from the deep learning domain, such as reinforcement learning and recurrent neural networks (RNNs). As one of the major benefits of integrating SI with autonomous platforms, we foresee the early identification of diseases via precision agriculture, which could lead to early application of plant protection products for sustainable farming practices. As presented in the work, currently, there is little literature on disease detection using spectral imaging on autonomous platforms. Given the fast developments in both affordable SI sensors and autonomous platforms, a huge increase in application can be foreseen.

Compliance with Ethical Standards
Conflict of Interest The authors declare that they have no conflict of interest.
Human and Animal Rights and Informed Consent This article does not contain any studies with human or animal subjects performed by any of the authors.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.