Smart fertilizer management: the progress of imaging technologies and possible implementation of plant biomarkers in agriculture

ABSTRACT

Precise use of fertilizers has been a focal point in world agriculture for years because it increases crop production and reduces the negative impact of over-fertilization. Smart fertilizer management using information/data, sensors, and smart tools allows correct fertilization in precision agriculture, smart agriculture, and integrated nutrient management. Recent progress in prediction methods, such as hyperspectral imaging with machine learning, supports accurate N fertilization. In contrast, damage caused by the deficiency of several nutrients, such as Fe, K, and N, can be accurately identified by red, green, and blue (RGB) images. Current imaging technologies cannot cover all nutrients, but it has been proposed that nutrient biomarkers could be able to quantitatively predict the status of a particular nutrient with high specificity. This review presents an overview of the current approaches to plant phenotyping by imaging/sensor- and plant biomarker-technologies, and to soil analysis by imaging/sensor technologies for smart fertilizer management.

KEY WORDS:

• Fertilizer management
• imaging
• integrated nutrient management
• plant biomarkers
• precision agriculture
• soil analysis

1. Introduction

High-input agricultural production by high-yielding crops requires a drastic increase in fertilizer use (Roberts and Mattoo 2018; Timsina 2018). In fact, the current world agricultural production includes the use of large quantities of fertilizers in most commercial crops, mainly nitrogen (N), phosphorus (P), and potassium (K). A study entitled ‘World fertilizer trends and outlook to 2020’ (FAO 2017) reported that the global demand for fertilizer nutrients by the end of 2020 (201.66 million tons) was about 10% higher than that in 2016. Reduction/optimization of fertilizer input would be an essential step in the agricultural sector which would allow the crop producers to manage fertilizer use more precisely and to improve productivity and sustainability.

In this context, smart fertilization, utilizing the Internet of Things (IoT) and other information and sensing tools is an important concept in current agricultural production (Sun et al. 2020). It is a part of a wider concept called ‘smart agriculture’ which encourages the use of information technologies and robotics (https://www.maff.go.jp/e/policies/tech_res/smaagri/robot.html) or sensors and IoT tools in agricultural practice (Walter et al. 2017; Lavanya, Rani, and Ganeshkumar 2019). In fact, precision agriculture (PA), which has been adopted by 40% of farms in the USA (Hellerstein et al. 2019), introduces optimized fertilization with smart tools in order to reduce production costs and increase plant yield. In turn, it reduces the environmental cost of fertilizer use because it saves energy and raw materials required for processing fertilizers, and it protects the environment from pollution by over-fertilization (Bacenetti et al. 2020). In addition, optimized fertilizer management can improve the quality of crops, which can decrease as a consequence of over-fertilization. The quality of olive oil was found to decrease because of N over-fertilization (Erel et al. 2013), and that of tomato fruits decreased because of P over-fertilization (Nowaki et al. 2017).

Smart N fertilization has been adopted on farms that use PA, where farmers try to supply fertilizers by implementing the correct treatment in the right place at the right time (Pierce and Nowak 1999; Robert 2002). Recently site-specific N management in PA is supported by the ‘newly-developed technologies’, such as global positioning systems (GPS) and geographic information systems (GIS) developed in the 1990s (Stafford 2000), which enabled customized fertilization management for small areas within fields instead of managing whole fields as a single unit (Rogovska et al. 2019) in order to increase nitrogen use efficiency. In fact, smart N fertilization is an important target in PA (Hellerstein et al. 2019). PA has also employed remote sensing, sensors, and cameras to evaluate the nutrient status in soils and plants for predicting optimal fertilizer levels. This was achieved by continuous development of sensing/imaging tools (e.g., hyperspectral cameras) and methods of image analysis (e.g., machine learning), which further allowed more precise fertilization. PA originally involves fertilizer management for the reduction of production costs and for the increase in plant yield (Bacenetti et al. 2020).

The development of reliable procedures for smart N fertilization is related to several factors. The N status of crop plants correlates with their capacity for photosynthesis, which is well correlated with their growth/yield (Makino 2003). In contrast, as a major component of photosynthesis, the content of chlorophyll can be predicted by imaging (Wang et al. 2019). To date, such relationships between nutrient status and some detectable indexes have been studied for other nutrients. In this review, we summarized the current progress of fertilizer management in PA, mostly related to N fertilizers, and we discussed various technologies that monitor the status of other nutrients in plants and soil.

2. Smart N fertilizer management by remote sensing

Currently, for the decision-making process, PA utilizes GIS data and plant/soil parameters obtained through remote sensing and other analyses. For example, several developed countries, such as Sweden, the USA, Australia, and Canada, have implemented national projects to map the soil properties, including organic matter content and other chemical soil properties, in different regions using GIS (Say et al. 2018), for the development of soil management protocols (Söderström et al. 2016). On the other hand, vegetation indices (VIs), which are generated from multi-spectral data, allow the evaluation of crop growth and N status (Osco et al. 2019a). The combination of these information allows the control of a variable-rate applicator (VRA) equipped with a GPS-positioned tractor, which accurately conducts N fertilization at the right positions (Figure 1). This can be achieved because of the increased resolution of GPS and satellite images, while it is also related to the development of cameras, sensors, and imaging methods (Maes and Steppe 2019).

Figure 1. Example of normalized difference vegetation index (NDVI)-based site specific fertilizer management. NDVI-map can be generated using the planet satellite imagery for the field growing crops (https://www.planet.com/) (A). Blue color indicates high NDVI, while red color indicates low NDVI value. According to the NDVI-value, fertilizer level was decided at 280 (yellow), 350 (green) and 420 (blue) Kg/ha (B). Images were taken from wheat field (Kaizu, Gifu, Japan) for designing for variable-rate fertilizer application

At the early stage of satellite imaging, natural plant habitation was monitored by VIs, which were transformed from the spectral reflectance of red and near-infrared (NIR) (Lloyd 1990). The red/infrared ratio is based on the fact that healthy vegetation reflects more NIR than red light (Herrero et al. 2020), which is partially explained by the difference in the amount of chlorophyll between healthy vegetation and stressed vegetation (Fillella and Penuelas 1994; Gitelson, Gritz, and Merzlyak 2003). Several VIs (using different formulas and different wavelengths) have been proposed for modeling plant growth capacity, yield, and N status. Among the various VIs, the normalized difference vegetation index (NDVI) is the most popular VI which is extensively used to determine the N status of crops in PA (Padilla et al. 2015; Stefanini et al. 2019). NDVI is calculated as NDVI = (NIR-Red)/(NIR + Red), and it can well describe the variation in chlorophyll content and thus in the crop N status (Vizzari, Santaga, and Benincasa 2019). Chlorophyll fluorescence [spectral emission ranges approximately 650–800 nm in intact leaves with two maxima (under normal ambient temperatures): one in the red spectral region around 685–690 nm (F₆₈₅) and the other in the far-red spectral region around 730–740 nm (F₇₄₀); Mohammed et al. 2019] driven by solar or laser radiation, which allows the evaluation of chlorophyll contents more accurately than VIs, has recently been used for the determination of crop N status (Yang et al. 2018; Mohammed et al. 2019). Continuous development of imaging technologies, which include unmanned aerial vehicles (UAVs; Liu et al. 2018; Osco et al. 2019b; Yang et al. 2020), hyperspectral data (camera/sensor; see next section for details on hyperspectral imaging, HSI), and data analysis (e.g., machine learning, ML; Tantalaki, Souravlas, and Roumeliotis 2019), allow more precise N fertilizer management. Several new VIs have been proposed by the data mining of HSI, which allow the prediction of the N content in crops other than well-studied grain crops, for example in spinach, (Corti et al. 2017), citrus species, (Liu et al. 2015) and cucumber (Ji-Yong et al. 2012) (Table 1).

Table 1. Evaluation of nutrients status and content in crops by imaging techniques

Imaging Technique	Detection method/Algorithm	Nutrients	Analyzed plants	Experimental setup	Combination of additional data	References
	linear-plateau model (Yield-model using VIs)	N-status	Maize (canopy)	Field (using airplane)	SPAD data	Á, Lloveras, and Martínez-Casasnovas 2018
	SAM (prediction model)	N-content	Orange(leaf)	Field (using UAV)	Spectroradiometer data (Ground sensor)	Osco et al. 2019b
	Stepwise-MLR (prediction model)	N-content	Rice (canopy)	Field (using UAV)	Hyperspectral data (Ground sensor)	Zheng et al. 2018
	LRM (prediction model)	N-use efficiency for variety	Winter Wheat (canopy)	Field (Using UAV)	No	Yang et al. 2020
MSI	LRM (prediction model)	N-content	Sugarcane(leaf)	Field (Using UAV)	No	Shendryk et al. 2020
	MLR (prediction model with VIs)	N-status	Oilseed rape (canopy)	Field (using UAV)	Hyperspectral data (Ground Sensor)	Liu et al. 2018
	LRM (evaluate the performance of the VIs)	N-status	Winter Wheat (canopy)	Field (using UAV)	No	Lu et al. 2019
	LRM (estimation model)	N-content	Winter Wheat (canopy)	Filed (using UAV)	No	Geipel et al. 2016
	ELR (estimation model)	N-management	Soybean (canopy)	Filed (using UAV)	RGB image &Thermal image	Maimaitijiang et al. 2017
	MLR (prediction model)	N-deficiency	Cucumber (leaf)	Greenhouse & Lab	No	Ji-Yong et al. 2012
	RF (wavelength selection), PLSR (prediction model)	N-content	Pepper (leaf-stem-root)	Greenhouse & Lab	No	Yu et al. 2014
	SG-PLS (prediction model for N), PLS(prediction model for P)	N & P-contents	Citrus (leaf)	Greenhouse & Lab	No	Liu et al. 2015
	LNC-indices (prediction model)	N-content	Rice (canopy)	Field & Lab	No	Moharana and Dutta 2016
	PLSR (prediction model)	N-content (high accuracy)P,K,S-contents (less accuracy)	Maize (leaf), Soybean (leaf)	Lab	No	Pandey et al. 2017
HSI	PLSR (prediction model)	N-content	Spinach (canopy)	Greenhouse & Lab	No	Corti et al. 2017
	SAE-FNN (Predicted model)	N-content	Oilseed rape-leaf	Field & Lab	No	Yu, Lu, and Liu 2018
	PLSR (prediction model)	N-status	Wheat (leaf)	Lab	No	Bruning et al. 2019
	Normalized difference chlorophyll index	N-use efficiency for variety	Wheat (leaf)	Lab	RGB	Banerjee et al. 2020
	MLR (prediction model)	N-content	Apple (leaf & canopy)	Lab and field	No	Ye, Abe, and Zhang 2020
	PCA(Image feature evaluation)	N-status	Barley (leaf)	Greenhouse & Lab	SPAD data	Pagola et al. 2009
	LRM (RGB component analysis)	N-status	Wheat (canopy)	Field	GreenSeeker data	Li et al. 2010
	LRM (RGB component analysis)	N-status	Tomato (leaf)	Greenhouse & Lab	SAPD data	Mercado-Luna et al. 2010
	Correlation modeling	N-deficiency	Rubber Tree (leaf)	Lab	No	Yao and Luo 2012
	MLR (prediction model)	N-status	Rice (canopy)	Field	No	Lee and Lee 2013
RGB	Regression models (Image feature evaluation)	N-content	Rice(leaf)	Filed	SPAD data	Tewari et al. 2013
	Exponential functions (prediction model)	N-status	Cotton (canopy)	Field	No	Jia et al. 2014
	SVM (Image feature evaluation)	N-status	Rice (leaf)	Greenhouse & Lab	No	Lisu, Yuanyuan, and Ke 2017
	PLSR(Image feature evaluation)	N-status	Wheat (canopy)	Field	Reflectance sensor data	Elsayed, Barmeier, and Schmidhalter 2018
	CNN (Damage recognition)	K & Fe-deficiency	Soybean (leaf)	Field	No	Ghosal et al. 2018
	Linear regression (Yield model)	N-management	Maize (canopy)	Field	No	Zhang et al. 2020
	BPNN (Damage recognition)	N,P,K,Mg, Fe,Zn-deficiency	Rice (canopy)	Field	No	Anami, Malvade, and Palaiah 2020

HSI, Hyperspectral imaging; MSI, Multispectral imaging; RGB, Red-green-blue; MLR, Multiple linear regressions; RF, Random frog; PLSR, Partial least squares regression; SG, Savitzky–Golay; LNC, Leaf nitrogen concentration; SAE, Stacked auto-encoders; FNN, Fully connected neural network; SAM, Spectral angle mapper; LRM, Linear regression model; VIs, Vegetation indices; ELR, Extreme learning machine based regression; PCA, Principal component analysis; SVM, Support vector machine; CNN, Convolutional neural network; BPNN, Back propagation neural network; UAV, Unmanned aerial vehicle; SPAD, Soil plant analysis development.

The N status of crops (i.e., endogenous N level that supports physiological processes such as photosynthesis), which is associated with plant growth and yield, is more useful than tissue N content for efficient N management. In order to improve the accuracy of N status prediction, multispectral imaging (MSI) is combined with data obtained by other methods, such as HSI, visible red, green, and blue (RGB) images, and data obtained by sensors (Figure 2). Portable sensors that estimate the amount of chlorophyll and N status are currently used at farming sites (Cardim Ferreira Lima et al. 2020). Two nondestructive sensors have been widely adopted for measuring the plant N status, namely soil plant analysis development (SPAD; chlorophyll absorbance assessed by transmission of red light at 650 nm; Xiong et al. 2015) and GreenSeeker (red and near-infrared reflectance; Cao et al. 2015). These sensors can easily obtain proximal data at farming sites and can be combined with remote sensing data. This integration sometimes improves the effectiveness of the N fertilization model (Á, Lloveras, and Martínez-Casasnovas 2018). For phenotyping major cereal crops in the field, it is important to develop airborne images using UAVs (e.g., drones) with various prediction models (Maes and Steppe 2019). Recently, Maimaitijiang et al. (2017) integrated different types of cameras and sensors on the same drone to obtain the most accurate and efficient combination of data for accurate estimation of plant traits with the goal of stress assessment. These types of studies can promote the development of models for combined estimation of grain yields and economic returns (Á, Lloveras, and Martínez-Casasnovas 2018).

Figure 2. Diagrammatic representation of smart fertilizer management. Smart fertilizer management utilizes imaging, sensors, and biomarkers to phenotype the nutrient status in plants and soils. HSI, hyperspectral imaging; MSI, multispectral imaging; RGB, red-green-blue; SPAD, soil plant analysis development

Recently, RGB imaging has been used to predict the N status of crops in field conditions, including barley (Pagola et al. 2009), rice (Lee and Lee 2013), wheat (Li et al. 2010), and cotton (Jia et al. 2014) (Table 1). Some of these studies utilized on-ground sensors (e.g., reflectance sensors) with air-born images to improve the accuracy of the prediction model (Elsayed, Barmeier, and Schmidhalter 2018).

3. Phenotyping of other nutrient status and content by imaging and sensor technologies

Numerous studies have been carried out to investigate how imaging and sensor technology can be used to improve the management of fertilizers other than N (Mulla 2013). This section focuses on the combination of imaging and sensor technology for the analysis of crop and soil nutrient status with the goal of smart fertilizer management.

3.1. Plants nutrients phenotyping by imaging and sensor technologies

MSI, which includes two or three wavelengths (e.g., red and NIR), is quite useful for evaluating the N status in plants (Geipel et al. 2016; Zheng et al. 2018; Lu et al. 2019; Shendryk et al. 2020) (Table 1), but is not useful for evaluating the status of other nutrients. Recent progress in HSI and imaging analysis methods has made it possible to predict the N status along with other nutrients. HSI data, which are obtained by spectral ranging with narrow intervals (e.g., 1 or 10 nm), can be transformed into radiometric quantities, such as reflectance, absorbance, and transmittance, which can be further correlated to physiological variables such as the content of water, fat, protein, and other carbon-hydrogen-oxygen bonded constituents (Mahesh et al. 2015).

HSI technology usually uses the light reflected from objects to detect their constituents (Kamruzzaman and Sun 2016). The intensities of certain precise spectral bands can be utilized if they are associated with nutrient content (e.g., Yu et al. 2014). Recent progress in data analysis methods allows the utilization of HSI data for predicting the status of various nutrients. Different algorithms, including statistical methods such as multiple linear regression (MLR; Ji-Yong et al. 2012; Bruning et al. 2019; Ye, Abe, and Zhang 2020) and machine learning methods such as neural networks (Wiegmann et al. 2019; Patrício and Rieder 2018) and random forest (Bruning et al. 2019), have been used for identifying certain wavelengths used for the prediction of the content of each nutrient. HSI has been intensively used for the prediction of various nutrients content along with N content and status in various crops (e.g., Moharana and Dutta 2016; Corti et al. 2017; Pandey et al. 2017; Yu, Lu, and Liu 2018; Banerjee et al. 2020) (Table 1), which was not well studied using MSI.

Using HSI, Borhan et al. (2017) predicted the N content in potato, and Liu et al. (2015) predicted P contents along with N in citrus leaves. The contents of several nutrients can be predicted by HSI data mining, but this prediction would be less accurate than that of N. Pandey et al. (2017) predicted N, P, and K contents in soybean and maize using HSI. The model developed for predicting N contents was accurate (R² by PLSR > 0.95), but those developed for predicting P and K contents were less accurate (R² by PLSR <0.7). The authors concluded that further improvement of the prediction of P and K content, such as integration of other sensing methods, would be necessary for optimal P and K management.

Visible RGB images obtained by regular cameras have been utilized to predict the nutrients status of crops, including their N status and deficiencies of several nutrients (Pagola et al. 2009; Li et al. 2010; Mercado-Luna et al. 2010; Yao and Luo 2012) (Table 1). Digital cameras are a simple source of RGB images, which are widely used for plant phenotyping (Jia et al. 2014). The progress of phenotyping by RGB is supported by the improvement of camera and image analysis methods, which can predict the N status (e.g., using linear regression, Tewari et al. 2013, or using the support vector machine, Lisu, Yuanyuan, and Ke 2017). In addition, RGB imaging using regular cameras (e.g., smart phones) along with IoT could be useful in smart fertilizer management (Patil, Adamuthe, and Umbarkar 2018). Several studies have shown that RGB imaging can be useful for identifying plants damage caused by several nutrient deficiencies. For example, Ghosal et al. (2018) provided an explanatory deep-learning method that was able to accurately classify soybean leaf damage using RGB imaging, enabling Fe and K deficiency to be differentiated from other tested damages. This method would be useful for phenotyping nutrient deficiencies that occur in low-input agriculture systems. In the past several decades, the Indian government introduced integrated nutrient management (INM) to increase agricultural production, which can enhance production in low-input agriculture systems (http://agricoop.nic.in/divisiontype/integrated-nutrient-management). INM utilizes chemical, organic, and free-living microbes, such as plant growth-promoting rhizobacteria or arbuscular mycorrhizal fungi, to obtain synergetic effects of multiple fertilizers (such as P and N) (Kurle and Pfleger 1994; Roesti et al. 2006; Khan and Zaidi 2007; Agrahari et al. 2020a). The avoidance of nutrient deficiency is an important technology for increasing the benefit of INM in low-input agriculture. Anami, Malvade, and Palaiah (2020) recently succeeded in recognizing N, P, K, Mg, Fe, and Zn deficient spots in Indian rice paddy fields using RGB image analysis. This approach is important for establishing smart fertilization, particularly at the farms operating INM. Recently, studies have also successfully identified the cause of stress in plants based on plant morphological indicators (Olas, Fichtner, and Apelt 2020). For example, Zhang et al. (2020) used a robust method (U-Net model for RGB image analysis) for purple rapeseed leaf segmentation, providing a new method for crop N stress monitoring. Recognition and discrimination between different stress factors that damage crops would help to rearrange the fertilization strategies for the subsequent harvest.

3.2. Soil nutrients phenotyping by imaging and sensor technologies

Quantitative evaluation of the N status of plants has almost been made available by imaging-based remote sensing and by several portable sensors (e.g., SPAD and GreenSeeker). In addition, high correlation between the N status and crop growth and yield compare to the other nutrients can help to develop a model to predict crops yield, therefore smart N fertilization has been more successful than smart fertilization with other nutrients.

The problem of less effective management of other nutrients could be solved by developing reliable and economical procedures that could estimate their status in soil and plants. Integrated management of digitized data of soil fertility and plant nutrient status is one approach for achieving this, but it has a high cost for setup nationwide digitized soil data. This is only available in certain countries, such as Sweden (Söderström et al. 2016) and the USA (Owens et al. 2020). In other countries, different procedures are required to evaluate soil properties, e.g., sensors (which can assess soil organic substances, moisture, color, nutrients, pH; Povh et al. 2014; Murray 2018), diagnostic kits (which can assess soil pH; Stevens, Dunn, and Phipps 2001), or imaging technologies (which can assess soil organic carbon; Hobley et al. 2018).

Sensor technology is based on electrical, electromagnetic, optical, radiometric, hydraulic, acoustic, pneumatic, and electrochemical principals and measurements, using which the sensors are capable real-time measurements of soil nutrients and pH, thereby supporting fertilization management (Adamchuk et al. 2018). Most recently, Erler et al. (2020) developed a sensor that utilizes laser-induced breakdown spectroscopy with multivariate regression methods (i.e., least absolute shrinkage and selection operator regression), which enables in-field determination of the concentration of several soil nutrients (such as Ca, K, Mg, N, P, Mn, and Fe).

On the other hand, already developed HSI technology in the NIR range (NIR-HSI) provides the information that can predict soil nutrients accurately (Manley 2014). NIR-HSI utilizes electromagnetic radiation between 750 nm and 2500 nm, which is mainly absorbed by molecular bonds with highly harmonic vibrational modes, such as C-H, O-H, and N-H (Pasquini 2018) which are characteristic of soil organic matters (Araújo, Demattê, and Vicente 2014). Recently Pudełko et al. (2020) compared the performance of NIR spectroscopy and NIR-HSI in predicting the organic carbon and total N contents in mine soils. They developed predictive models based on partial least square regression or artificial neural networks (ANN). They found that NIR-HSI based models had distinctly higher accuracy of organic carbon and total N content prediction than NIR spectroscopy. Swetha et al. (2020) developed an RGB imaging method that can predict soil texture using a deep learning algorithm [convolutional neural network (CNN) model].The development of RGB-based soil imaging is in the beginning stage, but in the future, it could help to further improve soil nutrient status management using HSI and sensor-based technologies.

4. Biomarkers for smart fertilizer management

In smart fertilizer management, it is very important to evaluate the plant response to nutrients at the early stage of plant development. This can be done by developing sensitive procedures, such as biomarker assessment, which can evaluate plant nutrient status using physiological and molecular biological markers. The term ‘biomarker’ originates from the field of pre-clinical research and clinical diagnosis (Biomarkers Definitions Working Group 2001). For example, chlorophyll is a kind of biomarker that can predict the N status in plant leaves, suggesting that other biological molecules could be utilized as biomarkers in agricultural practice. Recently, several laboratory and field studies have been conducted to identify biomarkers for potential use in agriculture.

4.1. Transcriptional biomarkers

Transcriptional or RNA biomarkers, which could be defined as specific changes in gene expression toward particular factors (Sewall et al. 1995), are correlated with various nutrients. Although transcriptional biomarkers have some limitations in terms of cost and handling RNA samples, they have high specificity and robustness among a wide range of plant species (in terms of gene ontology; Ito et al. 2019; Agrahari et al. 2020b), as well as high reliability for diagnosing various plant conditions other than nutrient deficiencies. Hermans et al. (2010) reported specific gene expression in the mature leaves of Arabidopsis thaliana under Mg deprivation, which was not affected by other mineral deficiencies (i.e., N, P, and K deficiencies) and environmental stresses (salt and drought stress) (Table 2). Recently, plant microRNA profiling revealed various miRNAs related to P deficiency and their specific expression under P-deficient and P-sufficient conditions (Zhao et al. 2013; Hackenberg et al. 2013; Bej and Basak 2014) (Table 2). These markers were not tested in field or pot soil experimental conditions, but several studies have identified certain biomarkers in field conditions.

Table 2. Biomarkers used for plant nutrients status analysis

Approach	Biomarkers	Response	Plants	Reference
Transcriptional	112 genes set	N-status	Maize	Yang et al. 2011 (see Table II)
	miR399, miR827	P-deficiency	Barley	Hackenberg et al. 2013
	14 genes set19 genes set	P-statusN-status	Rice	Takehisa and Sato 2019 (See Table 1)
	DDF1,TSPO,RRTF1	Mg-deficiency	Arabidopsis	Hermans et al. 2010
	ALS3	Al-toxicity level in acid soil	Arabidopsis, Soybean	Sawaki et al. 2016; Agrahari et al. 2020b
Biochemical	NR, GDH	N-deficiency	Tomato	Urbanczyk-Wochniak and Fernie 2005
	Anthocyanin	P-deficiency	Various plant species	Schachtman and Shin 2007
	Putrescine	K-deficiency	Various plant species	Cui et al. 2020 (See Table 1)

ALS3, Aluminum sensitive 3; DDF1, Dwarf and delayed flowering 1; TSPO, Tryptophan-rich sensory protein related; RRTF1, Redox responsive transcription factor 1; miR, microRNA; NR, Nitrate reductase; GDH, Glutamate dehydrogenase.

Yang et al. (2011) identified N-responsive transcriptional biomarkers using various maize genotypes grown under different N conditions, and they found several biomarker genes with N specificity and quantitative differences among various maize genotypes. N-biomarkers are robust among genotypes, and they could potentially serve as early N deficiency indicators for field-grown plants (Table 2). Takehisa and Sato (2019) identified three gene sets for monitoring the N and P status of rice throughout its growth under field conditions. These gene sets were selected by large-scale co-expression gene network analysis from previous transcriptome data of rice plants under N, P, and K deficiencies in laboratory conditions (Takehisa et al. 2013, 2015).

Recently, based on model plant studies (Sawaki et al. 2016); another type of transcriptional biomarker which can quantitatively monitor aluminum toxicity in acid soil was established in soybean (Agrahari et al. 2020b). An ortholog of the Arabidopsis Al-specific gene ALUMINUM SENSITIVE 3 (ALS3; Sawaki et al. 2016) in soybean (GmALS3) was inducible in the shoots of plants grown in acid soil with insufficient CaCO₃. The Al-inducible expression of GmALS3 was found to be robust among genotypes and highly specific to Al stress compared to other stresses. Along with soybean, other legumes showed Al-specific inducible expression of ALS3 orthologs in the shoots. This suggested that ALS3 orthologs can be utilized as potential transcriptional biomarkers for crops to determine the liming level (Al toxicity) in acidic soil.

4.2. Biochemical biomarkers

Nutrient deficiency in plants leads to changes that affect their metabolic pathways, resulting in major changes in particular metabolites (e.g., amino acids, organic acids, lipids, and sugars). Such metabolites act as biomarkers for nutrient deficiencies (Steinfath et al. 2010; Sung et al. 2015). Several signature metabolites have been identified as biochemical biomarkers for plant nutrient responses.

In plants, K deficiency strongly influences various metabolites, such as the accumulation of free putrescine, one of the polyamines that play a role in the replacement of K⁺ as an organic cation in K-deficient plants (Tachimoto et al. 1992). Putrescine in leaves is considered to be specific to severe K deficiency symptoms (Cui et al. 2020). Conversely, excess K causes a decrease in putrescine by converting it to higher molecular weight polyamines (i.e., spermine and spermidine; Aurisano et al. 1993; Reggiani et al. 1993) or it causes putrescine extrusion efflux from the cytoplasm of the roots into the external medium (Tamai et al. 2000). Thus, putrescine level is sensitive to external K and appears to be a good biomarker for K deficiency (Table 2). In contrast, the accumulation of free spermidine and spermine in plant leaves can be a biomarker for excess K (Geny et al. 1997).

Anthocyanin production is stimulated by prolonged P starvation in plants (Jiang et al. 2007; Xu, Weng, and Yang 2007). It protects the photosynthetic apparatus against photodamage and increases photosynthetic efficiency (Steyn et al. 2002; Henry et al. 2012). On the other hand, P starvation induces arrested primary root elongation and lateral root development. These changes in root architecture include the formation of starch granules in the columella cells of primary roots (Sack 1997; Sato and Miura 2011; Péret et al. 2011; Waidmann et al. 2019). The sensitivity to anthocyanin accumulation in the leaves and starch granules in the primary roots may act as a biomarker for P starvation in plants (Schachtman and Shin 2007; Wang et al. 2015) (Table 2).

Plants reduce their chlorophyll content in response to N deficiency (Coleman, Rosen, and Schwartzbach 1988), which, in combination with SPAD (nondestructive sensors), is considered as a biochemical biomarker (Table 2) for detection of plant N status in plants in agricultural fields (Yue et al. 2020; Cardim Ferreira Lima et al. 2020). Jouili, Bouazizi, and El Ferjani (2011) found that under heavy metal stress conditions, the changes in quantitative and qualitative profiles of peroxidases, along with enzyme activity, are considered to be biomarkers of heavy metal stress. As in the conditions of heavy metal stress, plants also reduce the activity of nitrate reductase (NR) and glutamate dehydrogenase (GDH) during N starvation (Coleman, Rosen, and Schwartzbach 1988). It is possible that these enzymes are also affected by other nutrient or environmental stresses, but in the statistical sense (order of magnitude; Jouili, Bouazizi, and El Ferjani 2011), they can be considered as biochemical biomarkers for detecting the N response of plants to N deficiency (Scheible et al. 1997; Urbanczyk-Wochniak and Fernie 2005; Srivastava and Singh 2006).

On the other hand, it was found that S deficiency reduces glucosinolate accumulation in plants (Falk, Tokuhisa, and Gershenzon 2007). After the integration of transcriptomics and metabolomics in order to understand the global response of nutrient deficiency in Arabidopsis, Hirai et al. (2004) found that genes or transcription factors that encode the metabolites involved in glucosinolate metabolism belong to a particular gene–metabolite network which is coordinately regulated under S deficiency. However, glucosinolate accumulation was not investigated as a biomarker for S response under field conditions.

As described in this review, the recent development of high-throughput technologies, particularly in transcriptomics and metabolomics, has facilitated the identification of plant biomarkers that can monitor plant responses to nutrients at the early stage of plant development in real time. The development of such approaches can be helpful for further development of smart fertilization (Figure 2).

5. Conclusion and future prospects

Smart fertilizer management has greatly impacted N fertilization by expanding the limits of advanced phenotyping of nutrient status in plants and soils (Á, Lloveras, and Martínez-Casasnovas 2018; Pudełko et al. 2020). Although most of these techniques have not been tested at the field level to date, advanced ideas/technologies which show better performance than other technologies applied in current agriculture are being developed in laboratories. Further studies are necessary to examine the efficient sampling scheme in field-level surveys, but several investigated biomarkers can evaluate plant nutrient status more accurately than other available methods. If the primary diagnostic criteria derived from current imaging technologies and biomarkers are used as auxiliary information, a smaller sample size for biomarker assessment might be sufficient for rational diagnosis. Therefore, biomarkers can be utilized for standardizing and developing the diagnostic tools for managing plant nutrients in real time, in a more precise and specific manner. On the other hand, the advancement of imaging-technology tested in different research fields (e.g., nutrient content in fruits in post-harvest qualification, Torkashvand, Ahmadi, and Nikravesh 2017; Maes and Steppe 2019) would increase the prediction accuracy for other minerals.

The following challenges in smart fertilizer management include the development of automatic, fast, and real-time management, which can be made possible by integration of phenotyping with IoT and robotics. In addition, advanced genotyping, such as genome-wide association studies and genomic selection, are promising prospects for genetic improvement of crop genotypes with improved nutrient efficiency, which can help to improve fertilizer management (Zhang et al. 2014; Lyra et al. 2017; Jaiswal et al. 2019; Nakano et al. 2020).

Acknowledgments

The author RKA is thankful to the Japanese Government (Monbukagakusho: MEXT) for the scholarship. The authors RKA and HK are thankful to the Japan Society for the Promotion of Science KAKENHI Grant.

Disclosure statement

The authors declare no conflicts of interest.

Correction Statement

This article has been republished with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

This work was supported by the Japan Society for the Promotion of Science KAKENHI (grant number 18H02113).