Visible RGB cameras receive lights in red (564–580 nm), green (534–545 nm), and blue (420–440 nm) board spectral bands in the visible spectrum regions to emulate human vision. Due to the broad applicability of RGB cameras in all the fields, they are well-developed products by private sectors with a large selective range in price, sensor sizes, lens availability, data storage, and transmission. For characterizing crop traits, the RGB cameras generally have the best performance in providing high spatial-resolution images with the ability to be associated with various sensing environments, which makes them the most commonly adopted sensors in the areas of remote sensing.

Applications in potato trait characterization have been leveraging the advantages of RGB images in different ways. In potato, studies have been conducted on quantifying disease severity using the RGB images by detecting pixels of diseased leaves with that of healthy leaves and soil. It was shown that the diseased leaves could be clearly distinguished from the transformed color spaces (Sugiura et al., 2016) or spectral indices (Gibson-Poole et al., 2017; Siebring et al., 2019) from the RGB color space. Also due to the contrast between plant and soil background captured by the RGB image, promising applications have been seen in detecting potato emergence (Gibson-Poole et al., 2018; Li et al., 2019) and calculating ground vegetation coverage (Gibson-Poole et al., 2018). Other than the color-based features, high resolution images from the RGB cameras can be used to reconstruct three-dimensional (3D) dense point models using photogrammetry methods. Plant height is then obtained by subtracting the digital elevation model (DEM) which represents the absolute elevation of the bare ground from the digital surface model (DSM) which represents the absolute height of the plant canopies. Li B. et al. (2020) compared the performance in estimating potato plant height from the RGB 3D models with the plant height regressed from the hyperspectral reflectance and the results showed that the RGB 3D models lead to a higher correlation with the ground truth data [Coefficient of determination R² = 0.93, root mean square error (RMSE) = 6.39 cm] than the full spectra regression (R² = 0.85, RMSE = 7.24 cm).

Compared with other sensors, the RGB cameras provide an intuitive way to visualize and quantify crop traits. RGB cameras are typically low-cost, lightweight, and can be easily mounted on various platforms for data collection. However, three spectral bands and the broad bandwidth of the RGB cameras limit their ability to receive crop traits other than human perceptions and consequently their applications in sensing more complicated crop traits, such as yield and nutrient deficit. Besides, it requires extra efforts to remove background noises from the RGB images under certain situations when the plants have similar visual colors to the background or when plants are covered with shadows and other plants (Li et al., 2014).

Multi- and hyper-spectral cameras provide information with higher spectral resolutions than the RGB cameras and spectral bands additional to the visible range. The multispectral cameras usually consist of a small number (less than 10) of discrete wavebands in the visible-near infrared range from 400 to 1,000 nm. The spectral wavebands usually are narrower RGB bands and those with prior knowledge of their sensitivity in receiving important crop information (Hunt et al., 2013), such as red edge (700–730 nm) and NIR (760–900 nm). There are two types of multispectral cameras currently used in precision agriculture, that is, customized and commercial off-the-shelf cameras. As suggested by the name, the customized camera has the internal NIR filter removed from a visible RGB camera and is then replaced with an external filter to alter the camera spectral sensitivity (Zhou et al., 2016a), for example, from R-G-B to R-G-NIR. The commercial multispectral cameras are explicitly made with an interference filter to block or transmit specific lights (Xie and Yang, 2020). Compared with the customized camera, the commercial ones can provide more than three wavebands with other selectable spectral regions, but with lower spatial resolutions (Table 1).

By integrating both spectroscopic and imaging techniques into one system, hyperspectral cameras provide 3D hyperspectral cubes that contain both contiguous spectral information and two-dimensional spatial information. There are four types of hyperspectral cameras based on their image acquisition methods, namely point scanning, line scanning, area scanning, and snapshot cameras (Wu and Sun, 2013). The line scanning, also known as the pushbroom cameras, records a line of an image at each time and then forms the whole image by scanning along the other direction. The line scanning cameras are particularly suitable for being carried on imaging platforms that follow certain moving patterns (e.g., zigzag paths) to cover a large field and thus are the most popular ones in precision agricultural applications (Xie and Yang, 2020). Compared with the multispectral camera, hyperspectral cameras provide a real spectrum in every image pixel, allowing to calculate advanced vegetation indices (VIs) and estimate unperceived crop phenotypes. However, hyperspectral cameras have more restrictions on reflectance calibration procedures and moving patterns and they are more costly both in the devices and data storage (Li et al., 2014).

The spectral cameras have been widely used to predict potato leave biomass (Li B. et al., 2020; Peng et al., 2021)/tube yield (Sankaran et al., 2015; Tanabe et al., 2019; Li B. et al., 2020; Sun et al., 2020), detect canopy damage/disease (Nebiker et al., 2016; Hunt and Rondon, 2017; Duarte-Carvajalino et al., 2018; Franceschini et al., 2019), and estimate potato biophysical properties, such as chlorophyll content (Elarab et al., 2015; Franceschini et al., 2017a; Li C. et al., 2020), nitrogen concentrations (Nigon et al., 2014, 2015; Hunt and Rondon, 2017; Hunt et al., 2018), and leaf area index (LAI, Franceschini et al., 2017a). Among them, disease detection is the common application of visible RGB and spectral cameras. Unlike separating the healthy and unhealthy leaf areas in RGB images, spectral cameras are used to investigate the relationship between spectral measurements and disease severity at plot/field level. For example, the normalized difference vegetation index (NDVI) combined with spectral reflectance values in NIR and R was able to highlight sites infested by blight disease in a large potato field (Nebiker et al., 2016). Atherton et al. (2015) also reported that heavily diseased plants had significantly different spectral profiles from healthy plants.

Radiation emitted by objects in the longwave infrared range (8–1.3 μm) is proportional to their surface temperature, which can be exploited for remote measurements of plant canopy temperature using TIR cameras (Kuenzer and Dech, 2013; Williams et al., 2018). Plant canopy temperature has been used as an indicator of plant stomatal traits (e.g., stomatal conductance, stomatal aperture, or leaf porosity) because leaves surfaces are cooled by evaporation, and stomatal opening or higher transpiration rates lead to high evaporation rates (i.e., decreased temperatures; Deery et al., 2019). In plant science and agronomy, measuring plant leaf temperature using thermal imaging has long been used to study the mechanisms behind plant water relations (Chaerle and Van Der Straeten, 2000), plant physiology processes (Jones et al., 2002; Aldea et al., 2006a; Grant et al., 2010), plant responses to pathogens (Chaerle et al., 1999) or environmental stressors (Jones et al., 2009), and plant photosynthetic rate (Aldea et al., 2006b; Vítek et al., 2020). In the area of precision agriculture, substantial interests have been witnessed in irrigation scheduling (Li et al., 2022), yield prediction (Bellis et al., 2022), evapotranspiration estimation (Lu et al., 2022; Wolff et al., 2022), pre-symptomatic diagnosis (Cohen et al., 2022), and stress tolerance (Zhou et al., 2020; Degani et al., 2022; Rippa et al., 2022).

In potato applications, TIR cameras are primarily used to evaluate potato water status. Plant water stress indicators, for example the Crop water stress index (CWSI), were calculated using TIR image-derived temperature and validated by comparing with biophysical measurements of plant water status. Related studies showed that the TIR image-based indices highly agreed with the ground measurements and were able to reflect the effects of different irrigation treatments (Rud et al., 2014; Gerhards et al., 2016; Cucho-Padin et al., 2020; Elsayed et al., 2021a). In some other applications, the TIR cameras were used as an aid tool for spectral cameras for detecting pest stress (Théau et al., 2020) or assessing chlorophyll content (Elarab et al., 2015).

The TIR cameras are the most effective tool in remotely sensing plant leaf/canopy temperature and water-related status. In practical, pre-calibration and post-correction procedures are required for receiving reliable temperature readings from the TIR cameras especially under changing ambient environmental factors such as air temperature, humidity, and wind speed (Jones et al., 2009; Jin et al., 2021). Additionally, the low spatial resolutions of the TIR cameras (Table 1) make them easily affected by image background temperature. It is also reported by Théau et al. (2020) that spectral indices (including visible ones) would be preferred over the TIR indices when both performed similarly in detecting various plant characteristics.

Light detection and ranging (or laser scanner) sensors are used to obtain depth information of targeted objects. A LiDAR device has a single-beam laser (narrowband in 600–1,000 nm) that emits a light pulse and a receiver measures the propagation time of the light pulse from emission to reflection. The time is then converted into a distance given the known speed of light (ten Harkel et al., 2020). LiDAR devices are lightweight with compact sizes and thus can be easily integrated on different platforms. Unmanned aerial vehicle (UAV)-based LiDAR systems have been applied to map the 3D structure of the plant canopy and topography of the landscape (Lin et al., 2019; ten Harkel et al., 2020). UAV-based LiDAR devices were reported to deliver accurate estimations of plant height with the RMSE of 0.02, 0.07, and 0.12 m compared with the manual measurements on snap bean crops (Zhang et al., 2021), sugar beet, and potato (ten Harkel et al., 2020), respectively. However, LiDAR devices can be expensive and require longer imaging times, compared to other stereo vision methods from images for 3D mapping.

Satellite imagery, as one of the remote sensing means, began more than half a century ago and has been used efficiently in multiple agricultural applications (Ma et al., 2021; Wang Y. et al., 2021). Its major benefits include non-destructive capture of the earth surface at wide geographical regions (hundreds of square kilometers in a single image), selectable revisit time (1 day–2 weeks), and sensors with high spatial and spectral resolutions (Zhang et al., 2020). In recent years, a number of studies have investigated the potential of using satellite-based remote sensing platforms for precision potato management activities. The specifications for eleven representative platforms and their corresponding applications in precision potato management have been summarized in Table 2.

Among all the selected satellite missions, the Landsat program involves a series of satellites that have been successfully used in agricultural monitoring because of the availability of historical data of the Earth’s surface for over 40 years. The most recent Landsat-8 was launched in 2013 with two imagers (Operational Land Imager and Thermal Infrared Sensor) that provide spectral observations in nine spectral bands at a 30-m spatial resolution. Several studies thus adapted the high-resolution observations from Landsat-8 for potato yield prediction (Al-Gaadi et al., 2016; Newton et al., 2018; Awad, 2019).

Besides the Landsat missions, Sentinel-2 is another popular platform for agriculture monitoring. For the potato management, Sentinel-2 imagery has been used for LAI and chlorophyll content estimation (Herrmann et al., 2011; Clevers and Kooistra, 2012; Clevers and Gitelson, 2013; Clevers et al., 2017), radiation utilization assessment (Peng et al., 2021), yield prediction (Al-Gaadi et al., 2016; Gómez et al., 2019; Abou Ali et al., 2020), and crop mapping (Ashourloo et al., 2020). Launched in 2015 by Copernicus Programme, Sentinel-2 systematically acquires optical imagery with global coverage. It is equipped with a push-broom multi-spectral instrument (MSI) which provides optical observations at 13 spectral bands ranging from visible (VIS) to short-wave infrared (SWIR) at high spatial resolutions (10–60 m/pixel). Like Landsat-8, Sentinel-2 also provides a global coverage but with a much higher temporal resolution (Table 1), which makes Sentinel-2 the most widely used satellite platform in potato precision management.

Other satellite platforms with unique specifications in one of the resolutions have been used to address specific purposes in potato management. For example, Terra Moderate Resolution Imaging Spectroradiometer has the highest spectral resolution of 36 bands in the range between 400 and 1,440 nm and thus the benefits for detecting the abnormal reflectance from leaves in the narrow bands for potato disease assessment (Dutta et al., 2014). Planet Scope provides images with the highest spatial resolution (3 m/pixel) every day through a constellation of about 130 low-orbit satellites. Though costly (∼$ 218 per 100 km²) compared to the publicly launched satellites, its imagery has been used as a good source for the decision-making at field-level and plot-level potato management on a daily basis (Abou Ali et al., 2020).

Over the decades, UAV-based platforms have been a popular approach in precision agriculture. UAVs refer to devices with flying capacity without humans onboard (Eisenbeiss, 2004; Sankaran et al., 2015). Other than the UAV and its accessories (e.g., batteries and remote controller), UAV-based remote sensing platforms for agricultural purposes generally include onboard sensors for collecting data, a gimbal for motion correction and stabilization, and a flight planning App installed in a smart mobile device for automated flight control. To deliver precisely geo-referenced data, UAV platforms are typically coupled with ground control points (GCPs) with known global navigation satellite system positions.

The UAV platforms are available in several types with distinguishing features. The most widely used UAVs in precision agriculture are multi-rotor copters or fixed-wings (Li et al., 2021b). Multi-rotor copters can be quadcopters, hexacopters, or octocopters depending on the number of propellers they have. Multi-rotor copters are driven by the pairs of propellers to create thrust downward and directed by changing the thrust of one or more propellers. Thus, they are able to hover, take off and land vertically and have no lower limits on flight altitude. Fixed-wing UAVs, like airplanes, have a pair of rigid wings and a thruster at back. The lift of fixed-wing UAVs comes from the air pressure differences between the upper and lower sides that are generated passively as its wings cut through the air at a specific angle when being pushed forward. Fixed-wing UAVs are fast and energy-efficient, and thus allow much larger ground coverage than the multi-rotor copters. On the other hand, they do not have the capability to hover and fly close to the ground. A list of popular UAV models and their applications in potato applications are summarized in Table 3.

Selection of UAVs is conditional on specific tasks, including field area, spatial resolution, flexibility for carrying various cameras, and costs. Among all UAV types, quadcopters are the most popular as they are light weight and easy to operate. The DJI Phantom series with high-resolution motion RGB cameras have been used in applications for obtaining plant structural traits (Li B. et al., 2020) and recently they have an updated multispectral version for those who favor spectral measurements. Hexacopters generally are not sold with onboard sensors but offer interfaces to power and control third-party cameras. Thus the hexacopters are often used to carry multi- or hyper-spectral cameras for perceiving complicated crop traits such as yield responses (Feng et al., 2020; Zhou et al., 2021b) or stress (Zhou et al., 2020, 2021a). Similar to the hexacopter, octocopters are flexible in carrying multiple sensors and other devices for durable flight. The most widely accepted octocopter is the Altura AT8 (Aerialtronics, South Holland, Netherlands) which was used to carry a hyperspectral imaging system (2.0 kg in total) as the original version of the Hyperspectral Mapping System (HYMSY, Suomalainen et al., 2014). The HYMSY system costs around $12,769 and is largely used in potato studies conducted in European. On the opposite to multi-rotor UAVs, the fixed-wings offer a solution to the need of covering large fields because of their nature of high flight speed and altitude. Cameras on fixed-wings should have high frame rates and short exposure time to avoid blurred images caused by the high flight speed (Li et al., 2021b). As products evolve, fixed-wings vendors now offer a variety of versions with dual or triple cameras for agricultural purposes.

Achieving the maximum crop yield at the lowest investment is an ultimate goal for farmers (Al-Gaadi et al., 2016). Early detection of problems associated with crop yield can greatly help with the decision-making for field management and reduce profit loss. In addition, yield mapping pre- or post-harvest can be used as spatial references for implementing variable rate technologies to achieve precise applications of field-level inputs and thus optimize production across entire fields (Al-Gaadi et al., 2016).

Potato tuber yield is jointly affected by genetic variation, environmental conditions (soil and weather), seed quality, and crop management practices (Li et al., 2021a). Classical potato yield prediction models are often used to estimate yields within the growing season by considering nitrogen fertilizer, temperature, and daylight or the incidence of solar radiation (Wang X. et al., 2021). Recent studies showed higher prediction accuracies by using machine learning models with remote sensing data at various levels, providing target-specific insights and recommendations for precision agriculture practices.

Salvador et al. (2020) predicted potato tuber yield at a municipal level in Mexica using machine learning models with satellite remote sensing imagery. The models were trained with NDVI values calculated from the TERRA satellite images combined with meteorological data on a monthly basis. The best testing performance was from the Random Forest model for the winter cycle (R² = 0.76 and RMSE% = 18.9) and the polynomial Support Vector Machine for the summer cycle (R² = 0.86 and RMSE% = 14.9). There are a few similar studies using satellite imagery and machine learning models for predicting potato yield at large scales in Spain (Gómez et al., 2019), Saudi Arabia (Al-Gaadi et al., 2016), and Bangladesh (Bala and Islam, 2009).

Unmanned aerial vehicle-based imagery is largely used for predicting potato yield in site-specific studies. Multi- (Tanabe et al., 2019; Li et al., 2021a) and hyper-spectral (Li B. et al., 2020; Sun et al., 2020) have been used to establish and validate a variety of machine learning models for predicting potato yield. The prediction performance of images collected at different growth stages was compared. It was found that images at earlier growth stages, i.e., the tuber initiation stage (Li et al., 2021a) and vegetative growth stage (Tanabe et al., 2019), tuber bulking stage (Li B. et al., 2020) had better prediction performances than those at the later stage when close to the end of seasons. Even though, Sun et al. (2020) reported a decent prediction accuracy with R² = 0.63, RMSE = 3.03 ton/ha, and RMSE% = 5.8 when using hyperspectral image features at tuber maturation stage and machine learning models to predict potato tuber yield under four irrigation levels.

Above-ground biomass is a basic agronomic trait in field investigations and is often used to indicate crop growth status, the effectiveness of agricultural management measures, and the carbon sequestration ability of crops (Li B. et al., 2020). Particularly for potatoes, AGB is one of the most important indicators for calculating the nitrogen nutrition index and diagnosing the potato nitrogen status for precise nitrogen fertilizer management (Jin et al., 2021; Yang et al., 2021). Manual measurement of AGB requires cutting culms at the ground level for a defined portion of the experimental plot, and then weighing before (AGB fresh weight) and after drying (AGB dry weight) in an oven until constant weight (Pask et al., 2012), which is laborious and subject to experimental errors (Jimenez-Berni et al., 2018).

Estimations of potato AGB using remote sensing methods have been explored in a variety of ways, by integrating spectral-based VIs, plant height (manual- or sensor-based) with machine learning, classical and empirical models. Ground-based data acquired on potato canopies by spectroradiometers (Pei et al., 2019; Elsayed et al., 2021b; Yang et al., 2021), RGB, and thermal cameras (Elsayed et al., 2021a) were used as predictors in machine learning models and the studies showed that the agreements between the estimated and measured biomass were up to R² of 0.7–0.8. Pei et al. (2019) also compared the estimation performance using the ground spectral data at different stages and the best result was given by the spectral features at the tube growth stage with R² = 0.71 and RMSE% = 19.3.

In addition to the spectral features, plant height was considered as an important contributor to potato biomass. ten Harkel et al. (2020) derived the potato plant height from a UAV-LiDAR device and imported it to an empirical model originated from (Jimenez-Berni et al., 2018) for the estimation of biomass from LiDAR. The biomass solely based on plant height had an R² of 0.24 with the RMSE of 22.1 g/m². In another case, the UAV image-based plant height (from RGB images using the Structure from Motion method) was integrated with UAV hyperspectral VIs into a Random Forest model. The model accuracy for the biomass fresh weight was up to R² = 0.90 with RMSE% = 20.6, and for the biomass dry weight up to R² = 0.92 with RMSE% = 17.4 (Li B. et al., 2020).

An interesting comparison in the estimation performance of potato biomass using ground, UAV, and satellite data was conducted by Peng et al. (2021). Data from each source were independently used as variables in a classical model (the Carnegie–Ames–Stanford approach). The results showed that the ground spectral data reached the highest accuracy with R² = 0.76 and RMSE% = 22.2, while the UAV and the Satellite multispectral image features had the similar performance with R² = 0.70 (RMSE% = 22.4) and R² = 0.72 (RMSE% = 24.4), respectively.

Potato plants are sensitive to water stress due to their shallow root systems. Approximately 70% of the total water needed for potato growth comes from the upper 30 cm of the soil (Ahmadi et al., 2011). Potato commonly grows in coarse-textured soil with low water holding capacity (Rud et al., 2014) or in arid and semiarid countries (Elsayed et al., 2021a). Therefore, inappropriate water management practices or water deficit stress in potato plants could cause significant reductions in tuber yield (Karam et al., 2014) and degraded tuber quality (Djaman et al., 2021).

The plant water deficit is conventionally detected by measuring plant water statuses, such as leaf water potential and stomatal conductance, or by measuring soil gravimetric water content. All of the measurements are direct, destructive, labor-intensive, and unable for large-scale cultivation or tracking intra-field variability (Rud et al., 2014). CWSI, accounting for the differences between canopy and air temperatures at a fixed air vapor pressure deficit, has been extensively validated to represent plant water status, after it was designed by Idso et al. (1981). The use of leaf or canopy temperature to detect the water-related stresses is based on the principle that a plant’s stomatal closure takes place during water stress, which results in a decrease of energy dissipation and an increase in plant temperature (Idso et al., 1981; Elsayed et al., 2021a). The CWSI has been proposed as a good indicator of plant water stress (Cucho-Padin et al., 2020) and proved relating to stomatal conductance (Rud et al., 2014) and yield (Ghazouani et al., 2017).

The computation of CWSI requires the knowledge of plant canopy temperature which can be acquired through remote sensing technologies. Handheld infrared cameras have been used to obtain temperature values of plant canopy and soil to calculate the CSWI values for potato plants under irrigation treatments. Cucho-Padin et al. (2020) found that the CWSI values derived from a TIR camera had a high correlation (the Pearson’s correlation coefficient r = 0.84) with those from ground proximal sensors. The results also showed that no significant reduction in tuber yield was observed by using CWSI < 0.4 as a threshold for irrigation in potato crops, compared to the irrigation amount with CWSI > 0.4, which agrees with other works (Ramírez et al., 2016; Rinza et al., 2019) and provides a reference for saving irrigation water in arid and semiarid regions (Cucho-Padin et al., 2020).

The effectiveness of aerial TIR images in representing potato water status was evaluated in Rud et al. (2014) by comparing the image-derived CWSI with the biophysical measurements of leaf stomatal conductance values. In this study, RGB images were acquired simultaneously with the TIR images and used to create masks for the TIR images to distinguish between vegetation and soil. The CWSI was calculated based on the difference between leaf and air temperatures normalized to the variation in environmental meteorological conditions and showed high correlations (0.64 ≤ R² ≤ 0.99) with the stomatal conductance measurements.

A few studies have tried to combine spectral indices with the CWSI and canopy temperature to indicate potato water stress. Elsayed et al. (2021a) reported a good alignment (R² = 0.89) between the estimated and measured water content of the potato AGB by using the spectral indices derived from RGB and thermal images combined with machine learning models. Similarly, Gerhards et al. (2016) compared the performance of detecting water content in potato plants between spectral and thermal-based indices. The results showed that the CWSI was most suitable for water stress detection among all the thermal-based indices. Additionally, it was found that the NIR/SWIR reflectance-based indices and the spectral emissivity (non-imaging devices) were equally sensitive to plant water content and responded quicker than the indices from imaging devices (such as the NDVI from a hyperspectral TIR camera; Gerhards et al., 2016).

Potato growth depends largely on the nitrogen availability in the soil (Nigon et al., 2014). However, the shallow-root system coupled with their common cultivation in coarse-textured soils leads to poor nitrogen use efficiency in potato plants (Nigon et al., 2014). Excessive applications of nitrogen fertilizers, on one hand, might have negative impacts on potato productivity by inducing a greater growth of aboveground biomass but reducing the root growth. On the other hand, it raises the health risks associated with nitrate-contaminated groundwater caused by the high rates of nitrate leaching in irrigated potato fields (Zebarth and Rosen, 2007; Lynch et al., 2012). Therefore, precise spatial and temporal nitrogen management is important to optimize production and minimize environmental nitrogen losses (Nigon et al., 2014).

One of the typical guidance on the timing and amount of nitrogen applications on potato is nitrogen nutrition index, defined as the ratio of the actual nitrogen concentration (N_c) over the minimum crop nitrogen concentration allowing for maximum biomass production (N_cri). While N_cri is an estimated value depending on the potato AGB, N_c needs to be determined along with the potato growth and traditionally by wet tests in laboratory. However, N_c values from the sampled potato plants might not accurately represent nitrogen status for an entire field and potato nitrogen status might change considerably by the time of analysis (Zhou et al., 2018). Thus, fast and accurate estimation of potato N_c values is urgently needed to assist the precision nitrogen management in potato productions.

Canopy spectral reflectance has been used extensively in agriculture to measure the actual nitrogen status of crop growth in order to guide nitrogen fertilization (Raun et al., 2002; Van Evert et al., 2012; Zhou et al., 2016c). The principle that the canopy spectral reflectance can indicate the canopy N_c values is based on the fact that the light reflected in certain wavebands is sensitive to leaf chlorophyll concentration, which in turn closely depends on leaf nitrogen concentration (Lawlor et al., 1989; Baret and Fourty, 1997; Hansen and Schjoerring, 2003). Studies have been conducted to compare the performances in estimating the potato N_c among canopy spectral reflectance obtained at different scales (handheld/UAV/Satellite) with narrow or wide spectral bandwidths and various ranges in spectral wavelength. The commonly used spectral features and their details have been summarized in Table 3.

Potato production is threatened by several diseases resulting in considerable yield losses, causing decreases in the quality and increases in the price of potatoes (Taylor et al., 2008). In-season early detection is particularly important for disease management decisions to control disease severity and prevent the spread of diseases (Oppenheim and Shani, 2017). Besides, detecting the overall disease grade of potato fields allows growers to make decisions regarding the best way to store, sell, and manage the fields for future generations (Gibson-Poole et al., 2018). Traditional potato disease inspection requires visual identification of the signs of disease. However, the disease plants can be difficult to identify, especially when symptoms are not yet developed or at the early stages. The remote sensing technique provides an opportunity to overcome the limitations, and they have already been widely used to detect diseases of potatoes for many years.

Among all the diseases that occurred during potato growth, Potato Virus Y (Potyviridae, PVY) is the most common reason for the rejection of seed lots from certification and one of the most common potato diseases (Griffel et al., 2018). The foliar symptoms of PVY include mosaic, leaf drop, and interveinal necrosis, and this virus can also cause necrotic rings on potato tubers (Gray et al., 2010). However, the recently prevalent PVY strains exhibit transient and milder foliar symptoms, making it less effective the visual inspection of seed lots with high PVY incidence (Frost et al., 2013). Spectral reflections in the NIR and SWIR regions have provided differentiable profiles in the underlying responses to the PVY stress. Couture et al. (2018) assessed the efficacy of hyperspectral information (a handheld device with narrow bands between 350 to 2,500 nm) in detecting the presence of PVY in potato leaves prior to the onset of the visual symptoms and observed a significant difference in the whole spectral profile between the PVY-negative and PVY-positive plants, especially in the NIR and SWIR spectral regions. A partial-least square discriminate analysis model trained with normalized spectral indices derived from the hyperspectral bands correctly identified the positive PVY infection status in vivo with a mean cross-validation accuracy of Kappa = 0.73. Similarly, high accuracy of 89.8% was reported by Griffel et al. (2018) in classifying the PVY-infected and non-infected potato plants using the UAV-based spectral information in the NIR and SWIR regions. Another example was given by Polder et al. (2019) that the healthy potato plants were distinguished from PVY infected ones (the classification Precision = 0.78) by a deep learning model directly using the hyperspectral images (400–1,000 nm).

Potato late blight (Phytophthora infestans) is another serious disease affecting potato production worldwide (Hwang et al., 2014). The disease rapidly destroys leaves, and it consequently leads to yield losses and tuber quality deterioration. Symptoms of late blight on potato plants mostly appear on the leaves, developing from small, dark green, irregular-shaped water-soaked spots to large, dark brown or black lesions surrounded by a yellow chlorotic halo. Compared to the PVY, late blight affections are more easily detectable from potato canopies using remote sensing techniques. Various sensing combinations have been evaluated with promising potential to estimate the late blight severity, from ground-based spectrometer (Ray et al., 2011; Franceschini et al., 2017a; Gold et al., 2020) to UAV-based RGB (Sugiura et al., 2016), multi-/hyper-spectral (Nebiker et al., 2016; Franceschini et al., 2017b; Duarte-Carvajalino et al., 2018) cameras and to satellite-based MSI (Dutta et al., 2014). Despite using considerably different prediction models, the general idea of quantifying the late blight severity is first to differentiate the infected and the healthy leaves parts and then obtain as the estimations the percentage of the damaged part over the healthy part (Dutta et al., 2014; Sugiura et al., 2016; Duarte-Carvajalino et al., 2018).

Another potato disease that has been detected using remote sensing techniques is early blight (Alternaria solani). Atherton et al. (2015) analyzed the spectral profile of potato plants with the early blight and noticed a significant difference in the spectral reflectance between heavily diseased plants and healthy plants. Van De Vijver et al. (2020) developed a ground-based proximal sensing platform-Hypercart for acquiring hyperspectral images in the potato field. The machine learning models highly agreed (R² = 0.92) with the visual assessments. Additionally, significant differences between healthy leaf and injected leaf tissues were observed in the green (550 nm), red (680 nm), and NIR (720–750 nm) regions and the NIR was recognized as the most helpful region for detecting the early blight lesions.