Remote Sensing Monitoring of Rice and Wheat Canopy Nitrogen: A Review

Abstract

Nitrogen(N) is one of the most important elements for crop growth and yield formation. Insufficient or excessive application of N fertilizers can limit crop yield and quality, especially as excessive N fertilizers can damage the environment and proper fertilizer application is essential for agricultural production. Efficient monitoring of crop N content is the basis of precise fertilizer management, and therefore to increase crop yields and improve crop quality. Remote sensing has gradually replaced traditional destructive methods such as field surveys and laboratory testing for crop N diagnosis. With the rapid advancement of remote sensing, a review on crop N monitoring is badly in need of better summary and discussion. The purpose of this study was to identify current research trends and key issues related to N monitoring. It begins with a comprehensive statistical analysis of the literature on remote sensing monitoring of N in rice and wheat over the past 20 years. The study then elucidates the physiological mechanisms and spectral response characteristics of remote sensing monitoring of canopy N. The following section summarizes the techniques and methods applied in remote sensing monitoring of canopy N from three aspects: remote sensing platforms for N monitoring; correlation between remotely sensed data and N status; and the retrieval methods of N status. The influential factors of N retrieval were then discussed with detailed classification. However, there remain challenges and problems that need to be addressed in the future studies, including the fusion of multisource data from different platforms, and the uncertainty of canopy N inversion in the presence of background factors. The newly developed hybrid model integrates the flexibility of machine learning with the mechanism of physical models. It could be problem solving, which has the advantages of processing multi-source data and reducing the interference of confounding factors. It could be the future development direction of crop N inversion with both high precision and universality.

1. Introduction

The effective guarantee of national food security is a key objective for China. Therefore, it has become a need of green agriculture to reduce the amount and increase the efficiency of chemical fertilizers, which could improve the effective supply of agriculture [1,2,3]. Rice and wheat are the main crops in the world, with a wide distribution and highly suitability. How to achieve high quality and yield is currently a major challenge for agricultural production [4,5,6]. N made up more than 40% of the mineral elements needed for the growth of rice and wheat [7]. Its content would impact the physiological traits, photosynthesis, and enzyme activities, leading to variations in protein content and grain production [8,9,10]. Healthy plants have a total N content that ranges from 0.3% to 5% of their dry matter, which directly affects crop production. N deficiency can hinder chlorophyll (Chl) synthesis and reduce the effective number of spikes, thus reducing yields. A surplus of N prevents photosynthetic products from reaching the seeds, delaying maturation [7,11]. Therefore, N content during crop growth has become one of the most important indicators in agricultural production management. Currently it is common to use base fertilizer, or top dressing through experience. It will result in too much or not enough N, which not only inhibits crop growth but also pollutes the environment. How to monitor rice and wheat canopy N timely and accurately, so as to guide variable rate fertilization has become a core research issue [12,13,14,15].

Currently the experts in plant protection and agronomy test N manually by observing crop symptoms, plant growth and leaf color. They also use chemical diagnostics, such as plant total N diagnostics and rapid nitrate diagnostics, to detect the N content of each organ. The former has variances in results due to subjective judgments. The latter has good accuracy but destroys the plant and has a time lag, making large-scale application difficult [16]. Therefore, traditional N assessment methods do not facilitate variable rate fertilization, because the information on the timing and extent of crop N abundance and deficiency is not efficiently provided. Remote sensing has enabled the rapid development of multiple scales of application, including satellite, unmanned aerial vehicles (UAVs) and ground. It is the current technology for rapidly acquiring spatial and temporal continuum information on a large range. The spectrum is sensitive to the N response in the Visible–Near Infrared (VIS–NIR). This spectral information can show subtle changes in N content, allowing for more accurate N retrieval [17,18]. Remote sensing monitoring of N in rice and wheat can be non-destructive, effective, and real-time for large-scale studies. It offers significant potential for crop nutrient diagnosis and as a basis of subsequent guidance on fertilizer application [19,20,21].

Both rice and wheat belong to the gramineous cereal crop in the botanical classification, and both are C3 crops with similar photosynthetic systems [22]. Thus, the absorption of colored light by Chl in both canopies is consistent and the response to the spectrum is similar. N is mainly found in the Chl of the photosynthetic systems and the process of N accumulation in rice and wheat has a high degree of similarity [23,24]. They are both transformed into nutrient bodies at the vegetative growth stages and into reproductive organs at the reproductive growth stages, and there are anisotropic changes in N accumulation in each organ at the growth stage [23,25]. Although the cropping patterns differ, one being dryland and the other paddy, this effect can be attenuated when pre-processing the remote sensing data [26]. Therefore, in remote sensing-based studies, the N transformation processes in rice and wheat are highly consistent, making their remote sensing monitoring systems relatively similar [14,27,28]. In this study, the remote sensing monitoring techniques for both crops are explored in an integrated manner.

To analyze and summarize the current research hotspots and trends in the field of remote sensing of canopy N in rice and wheat, this paper traces the related literature in the Web of Science ™ Core Collection Database. The literature is retrieved and filtered by the topics “nitrogen concentration” or “nitrogen content”, “rice” or “wheat” and “remote sensing”. An initial collection of 572 apparently relevant records covered the period 2003–2021. An initial screening progress is conducted to exclude literature that is not relevant to the review, such as conference proceedings, patents, etc. However, it still contains some irrelevant literature, and a further screening is necessary. This is followed by a more detailed screening on titles and abstracts to exclude the following: (1) non-targeted research topics (e.g., corn, cotton, grassland, etc.); (2) not directly estimated N (e.g., Chl, protein, etc.); (3) estimated other indicators (e.g., leaf area, plant height, yield, etc.); (4) measured soil N or other trace elements in agricultural fields; and (5) review articles. By carefully reading the titles and abstracts, off-topic papers are obviously manually excluded. Eventually, a total of 174 articles were identified and analyzed in depth. Figure 1 shows the number of studies retrieved from 2003 to 2021 that used remotely sensed data to assess the canopy N status of rice and wheat, reflecting the general trend in research on the application of remote sensing monitoring of canopy N.

Remote sensing technology has shown great potential for crop N monitoring, providing research ideas from various perspectives. This paper based on the mechanism of remote sensing monitoring of canopy N, respectively summarizes the current techniques and methods from three aspects: remote sensing platforms for canopy N monitoring; correlation between remotely sensed data and N status; and the retrieval methods of N status. Then it discusses the factors affecting the accuracy in remote sensing of canopy N and sketches future areas for research.

2. Mechanisms for Remote Sensing Monitoring of Canopy N

2.1. Physiological Mechanisms of Crop N

Focusing on crop physiological mechanisms, N is closely linked to Chl. N is an important component in the formation of chloroplasts and Rubisco enzymes. Increasing leaf N content will significantly increase leaf chlorophyll content (LCC) and Rubisco enzymes, ultimately leading to a significant increase in crop photosynthetic rate and consequent changes in the external morphology and internal structure of crop leaves [9,29]. Chl can respond to N uptake by crops, and the strength of the relationship directly affects the accuracy of N estimation. However, N is redistributed and reused in the crop during growth. Early in crop development, N is initially concentrated in nutrient bodies such as leaves, and as nutritional and reproductive growth coexist, N starts to be distributed to nutrient bodies and reproductive organs; at the crop maturity stage, N in nutrient bodies is transferred to reproductive organs. This resulted in varying degrees of N and Chl content reduction in the canopy leaves, which changed the correlations between crop N and Chl at various growth stages [25]. In addition to Chl, other mineral deficits, diseases, frost damage, and water stress may also produce leaf yellowing, and using Chl content as a proxy for N is misleading, which limits the ability to estimate N directly from Chl [28]. It has been recommended to utilize leaf protein as a substitute for leaf N content in several studies [30,31], because in contrast to Chl, protein is also a major nitrogenous component in crops and contributes to the varied distribution of N in crop plants. The mechanism of protein–N interactions is under investigation.

2.2. Spectral Response Properties of Canopy N

N affects the spectral reflectance of crops by influencing the Chl content of green crops (Figure 2). Healthy crops’ VIS reflectance spectrum is determined by the Chl’s absorption effect, which forms a prominent reflectance peak near 550 nm. Multiple reflections in the NIR combine to generate a red edge region of reflectance in the range of 700–780 nm, and the rising slope of the curve reflects the Chl content per unit area to some extent. NIR (780–1350 nm) is closely related to leaf structure and is instructive for exploring whether N is influenced by leaf structure. Under N stress, both the canopy spectral reflectance and the vertical distribution of N will alter.

During N deficiency, the VIS reflectance of the crop canopy spectrum increased, while the NIR reflectance and red-edge position (REP) decreased; in excess of N, the VIS reflectance decreased, while the NIR reflectance and REP increased (Figure 2) [14,32,33,34]. As the growth stages proceed, the response of canopy spectral reflectance to crop N status reduced, and the VIS regions also displayed “red shift” and “blue shift” with the development [33,35,36]. It can be found that the analysis of crop N abundance and deficiency using spectral techniques can be specific to a band interval [7,37]. Hyperspectral techniques can even be precise to a specific band [38,39], which provides the possibility of effective identification of crop N deficiency.

The top leaves of the plant under N stress will use the N transferred from the bottom leaves, causing the bottom leaves to yellow and decline prematurely, while the top leaves color changes are not obvious because of being less stressed by N [40,41]. As influenced by the level of soil N supply, the spectral reflectance of leaves at different leaf positions differed erratically in VIS and SWIR, while showing a clear gradient in NIR [42]. Duan et al. [43] suggested that N concentration at different leaf positions decreases from top to bottom at the jointing stage, flowering stages, and filling stage, while the flag leaf stage shows an increasing and then decreasing trend. The vertical distribution of N in the plant is not constant and varies between N conditions, planting densities and growth stages [41,43,44]. Exploring the spectral response properties of the different leaf positions can serve as a foundation for precise N quantification.

3. Techniques for Remote Sensing Monitoring of Canopy N

The continuous development of remote sensing technology on ground-based, UAV-based, and satellite-based platforms provides a wide range of modal options for N status monitoring studies. There is the technical focus in the field of N monitoring, including exploring the correlations between remotely sensed information extracted from multi-source data and N status, while achieving accurate modeling inversion of N status. To summarize the relevant technologies, three main aspects are discussed: remote sensing monitoring platforms; the correlation between remotely sensed data and N status; and the retrieval methods of canopy N status.

3.1. Remote Sensing Platforms for Canopy N Monitoring

3.1.1. Ground-Based Platform

Spectral data collected by ground-based sensors can be divided into non-imaging spectral data and imaging spectral data, with a spectral range primarily in the VIS–NIR, and in several studies involving the SWIR [18,45]. This close-range spectral information has ultra-high spectral resolution and can respond to subtle changes in N. It has been widely used in N monitoring for both leaf and canopy scale. Non-imaging spectral data are point spectral data, and lack spatial information for N estimation at regional scale [46]. Imaging spectral data, on the other hand, combine spatial and spectral features and allow the estimation of canopy parameters from faceted data. However, due to being restricted by data volume and acquisition method, it is generally used for ground study and rarely used directly for diagnosis and applications of large area. Current ground-based spectrometers used commonly include ASD FieldSpec (Analytical Spectral Devices, Boulder, CO, USA), RS-5400 (Spectral Evolution, Haverhill, MA, USA), HR-1024i (Spectra Vista Corporation, Poughkeepsie, NY, USA), SOC710 (Surface Optics Co. Ltd., San Diego, CA, USA), and FISS (Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing, China) [47,48,49,50], etc. These instruments provide a stable and high spectral resolution, but some of them are heavy and are generally measured in a backpack or mounted on a tripod in research, with a single angle and way of acquiring data. The advent of handheld instruments such as the RapidSCAN (Holland Scientific Inc., Lincoln, NE, USA), 4300 Handheld FTIR (Agilent Technologies Inc., Santa Clara, CA, USA), and Crop Sense (Beijing Academy of Agriculture and Forestry Sciences, China) [51,52] symbolizes the development of spectral sensors towards lightweight and flexibility. The portable spectrometer can acquire data on a wide range of measurement scales flexibly and efficiently, and it is easy to install on a variety of platforms such as lab benches, black boxes, and rocker arms. In addition, the multi-angle spectral acquisition device [53,54,55] consists of several moving parts to adjust the observation position and direction, where a goniometer is often used to control the observation of zenith angle changes. The sensor is placed on the goniometer to obtain spectral data in the viewing zenith angles (VZA) ranged from −60° to 60° [53,54]. It is a convenient platform for obtaining multi-angle crop spectral data, which has many applications in the study of the vertical distribution of N in the plant canopy.

3.1.2. UAV-Based Platform

With the development of lighter and smaller sensors and the increased carrying capacity of UAVs, the UAVs carrying sensors for data acquisition have become mainstream platforms in crop N monitoring. It is possible to rapidly acquire ground data with high spatial, temporal, and spectral resolution, facilitating the research at small and medium scales. Compared to airborne platforms (operating at kilometers of altitude), UAVs have the benefit of low cost, low operating altitude, and greater flexibility in terms of data collection arrangements. The sensors currently on UAVs specifically include digital cameras, multispectral/hyperspectral sensors, infrared thermal imagers, chlorophyll fluorescence sensors and LIDAR sensors [56,57,58,59,60,61]. The hyperspectral imaging spectrometer perfectly combines the advantages of spectroscopic and imaging technology for use in a large area. The payload of UAVs has made lightweight, low-cost sensors a research focus, and airborne hyperspectral imaging spectrometers such as the UHD185-Firefly and Cubert S185 (Cubert GmbH, Ulm, Baden-Württemberg, Germany), and Micro-Hyperspec (Headwall Photonics Inc., Boston, MA, USA) [62,63,64] are already being used for N monitoring. The PIS112 hyperspectral imaging spectrometer (Beijing Academy of Agriculture and Forestry Sciences, China), GaiaSky-mini hyperspectral imaging camera (Sichuan Dualix Spectral Imaging Technology Co., Ltd., Chengdu, China) [65] and other sensors, as well as the eight-rotor unmanned aircraft system based on RGB and 25-band small multispectral cameras (Zhejiang University, China) [66,67] developed by multiple teams in China, have also been used for agricultural monitoring with good results. Among the image data acquired by the UAV platform, hyperspectral can show subtle changes in crop spectral reflectance features due to its narrow bandwidth and wide continuous spectral range, which is conducive to the fine monitoring of crop N. Most of the multispectral data are small in volume and the spectral range encompasses the N response sensitive VIS–NIR bands. However, the low spectral resolution tends to result in “missing” spectral information while overcoming the high redundancy of hyperspectral information. Studies using multispectral data sources must consider whether the ‘missing’ spectral information contains sensitive bands and how it can be modeled. UAV remote sensing is not affected by the external environment, such as the atmosphere, and provides better access to high-quality spectral information, making it a common data source for crop N research.

3.1.3. Satellite-Based Platform

Images acquired by satellite platform sensors are characterized by a large range and multiple time phases, making large-scale spatial monitoring possible. Satellite data commonly used for agricultural monitoring include Landsat, Sentinel-2, etc. [68,69,70,71]. The small canopy area of rice and wheat requires high temporal and spatial resolution of data in seasonal N management. RapidEye, WorldView-2, etc. have better results in terms of temporal and spatial resolution, and the indices constructed in this way have achieved better results in N monitoring [72,73,74]. It is of practical significance to monitor crop N in large areas by satellite remote sensing with high spatial and temporal resolution. However, the use of field canopy spectral reflectance to simulate the spectral bands of satellite multispectral sensors, and then transfer the N response models from field experiments to satellite images, inherently results in mistakes due to the low spectral resolution of the data [75,76]. Since the 20th century, China has launched a series of resources satellites, HJ-1A/1B satellites and GaoFen (GF) satellites, etc. [77,78,79,80,81]. The types of sensors and spatial resolution they carry have reached international advanced levels, and their applications in agriculture are gradually spreading. Among them, GaoFen-6 (GF-6) is the first GF satellite for precise agricultural monitoring, for the first time adding green vegetation-sensitive red edge spectral bands and operating in a network with GaoFen-1 (GF-1) to significantly improve the monitoring capability of agricultural resources. GF-6 has been shown to be effective in improving the accuracy of image classification [78,79,80,81], and research into the quantitative inversion of crop parameters is still in its infancy. Satellite imagery has a high point and a wide field of view, but the quality has uncontrollable factors. Acquisition is influenced by radiation, aerosols, and weather, increasing the image processing process; analysis is influenced by complex geography and the presence of mixed image elements, increasing the difficulty of N estimation. The effectiveness of crop N estimation models based on satellite imagery still needs to be improved. In addition, as some important reflectance features associated with N can only be measured by hyperspectral sensors, future research could be explored based on satellite hyperspectral imagery.

Remote sensing technologies from ground-based, UAV-based, and satellite-based platforms are maturing and have been widely used in agriculture (Figure 3). The different platforms have their own advantages in remote sensing monitoring, with their mounted sensors having a high degree of overlap in the wavebands sensitive to N (Figure 4). Crop N monitoring based on ground-based and UAV-based platforms has been effective. Because these platforms provide highly accurate data, more valid information can be mined from spectroscopy, graphics, etc. Satellite data, although limited by data volume and accuracy, has become an effective complement to them in expanding the scale of research. Research on multi-sensor and multi-platform joint observations is gradually being developed.

3.2. Correlations between Remotely Sensed Data and N Status

Crop N status can be distinguished between two measurement perspectives: one is area-based measure “nitrogen content” (N_area, per unit area); another is mass-based measure “nitrogen concentration” (N%, per unit dry matter) [28]. N% can be converted to N_area by plant leaf (or other plant organs) dry biomass. In research, N content generally includes leaf/plant nitrogen accumulation per unit soil area (LNA/PNA), etc.; N concentration generally includes LNC, plant nitrogen concentration on a leaf dry weight basis (PNC), canopy N concentration (CNC), canopy N density (CND), etc. [28,82]. On this basis, Nitrogen Nutrition Index (NNI) is defined as the ratio of the actual crop N concentration to the critical N concentration. It is a direct indicator of whether the crop N concentration is at an optimum level [83,84]. In existing studies, the accuracy of remote sensing estimates of N content is significantly higher than that of N concentration, due to the fact that N concentration is more difficult to extract from remote sensing information compared to N content [64,85,86]. However, N concentration is not disturbed by density and is more accurate in responding to crop N status [82]. Throughout the whole growth stage of rice and wheat, N concentration has a narrower range of variation with a decreasing trend, and the change rate decreased before it increased; N content is a product of the combination of N concentration and plant dry biomass, so N content has a wider range of variation with an increasing trend [87,88]. With the support of remote sensing technology, the study directly links N indicators and spectral reflectance.

3.2.1. Sensitive Spectral Extraction

Spectral response under crop N stress varies significantly, and the analysis of original hyperspectral information is an intuitive study in N remote sensing assessment [39,89], but the sensitive bands extracted vary for the same/different crops in different geographical environments [90,91]. As in different studies, the N-sensitive spectral bands of rice include 738 nm, 1362 nm, 1835 nm and 1859 nm [90], and also include NIR (>760 nm), visible (355, 420, 524–534, 583 and 687 nm) and red edge (707 nm) region [39]; the N-sensitive spectral bands of wheat include 440 nm and 610 nm [91], and also include 790.4 nm [92]. Wang et al. [38] explored the best common central bands 822 and 738 nm for LNA estimation in rice and wheat, which can effectively assess the N nutrient status of plants and reasonably reflect the intrinsic N information in different crops. Most studies have extracted several N sensitive bands to estimate crop N by individual spectra or combinations. Although the method is simple, the accuracy is affected by the stability of the spectral information.

Hyperspectral data have hundreds of high-resolution continuous spectral information. When exploring correlations between spectra and N status, using full band data as an input can increase errors and reduce efficiency. Whereas insufficient exploitation and use will lead to data waste, thus losing the significance of high-precision data. Therefore, improving the use efficiency of hyperspectral data still needs to be further explored. Wang et al. [93] divided the spectral data into five groups: blue, green, red, red edge and NIR, and extracted the corresponding N-sensitive bands, in which the red edge (702, 703 and 710 nm) and red edge (706, 733 and 759 nm) correlated with leaf and canopy-scale N status up to 0.92. Yu et al. [94] reduced the spectral data by a discrete wavelet multi-scale decomposition method (DWMD), achieving better results compared to iteratively retaining informative variables, with 16.28–26.23% improvement in coefficient of determination (R²). Hyperspectral data are very similar between adjacent bands, and their dimensionality reduction can help reduce the complexity of feature extraction. Liu et al. [95] demonstrated that using feature bands extracted by improved adaptive ant colony optimization algorithm as input parameters under the same prediction model can reduce the complexity of the model, while improving the prediction capability. In addition, the autocorrelation matrix (R² = 0.86 between N-sensitive bands and N status), the non-negative matrix factorization (R² = 0.83), the successive projections algorithm (R² = 0.66), and the competitive adaptive reweighted sampling (R² = 0.93), etc. are also gradually applied [62,94,96,97,98,99]. Improvements to the N-sensitive band extraction method have resulted in the stability of the retrieval N information, but make the processing process more complicated, with a corresponding increase in model performance and computation time [97]. In addition, the bands extracted by these algorithms integrate more reflectance information, which improves stability and anti-interference. This provides a guideline for constructing a unified and generalizable spectral feature extraction method.

3.2.2. Mathematical Transformations of Spectra

The difference in response to N between spectra can be increased by mathematical transformations of spectra, improving its adaptability to inversion N status. The spectral reflectance curve feature can reflect the N change trend, and the first-order derivative of the spectrum indicates the rate of change in the reflectance, which reduces the effect of background information and is widely used in N estimation [63,92,93]. The morphological differences in crop spectra can be described by characteristic parameters such as slope, angle, and rate of change in the curve. Slope and angle are generally calculated based on areas of reflectance that change abruptly, such as peaks and valleys. The slope indicates the rate of reflectance rise/fall, and the angle formed by the sides of the reflection peak and absorption valley indicates the width between the peak and valley [100]. The rate of change is a generalization of the first-order derivative, which can respond not only to the change in reflectance between successive wavelengths, but also to the rate of change in reflectance between any two wavelengths [101]. Based on two integration metrics, normalized area reflectivity curve and reflectivity integration index, Du et al. [58] combined more wavelengths with improved LNC retrieval performance.

The study showed that the red edge region of 700–780 nm is a sensitive band for responding to the growth status of green crops [102,103], and the characteristic parameters of REP, red-edge slope, red-edge peak, red-edge minimum amplitude and red-edge area obtained by mathematical transformation of the red edge band have also become common parameters for N diagnosis in rice and wheat [36,104,105]. The REP based on linear extrapolation method showed better canopy N concentration correlation at larger canopy cover. Since the first-order derivative has “bimodal” characteristics in the two main spectral regions, the conventional REP is not sensitive enough to canopy LNC only using single-peak maximum, and the spectral reflectance data can be fitted to generate continuous REP values to achieve a continuous relationship between REP and N [105]. Li et al. [106] proposed a continuous wavelet transform-based REP extraction technique, wavelet-based red edge position (WREP), which provides a new idea for understanding the spectral variation in red edge region. Guo et al. [107] constructed an algorithm based on the analysis of red edge features, shifting red edge absorption area (sREA), which enables the construction of N absorption models at the regional scale. Due to the possible discontinuity of changes, insignificant amplitude, and large errors in the RE first-order derivative spectra, the estimation of crop N by red edge parameter features is highly dependent on the feature extraction method. It is necessary to select the appropriate method and parameters according to the research needs. When the canopy cover is too high, the response of the red edge region to N becomes gradually slower and then there is saturation, so only using red edge parameters easily causes misjudgment.

The mathematical transformation of spectra can respond to the trend of N changes, but the spectral information is changed by the influence of environmental stress. Deeply mining the bands with more effective information and using them in combination is the key to overcome the external factors and improve the accuracy of the model.

3.2.3. Spectral Indices

The information presented by specific parameter combinations (mainly difference, ratio and normalized values) is more stable and representative, and it has become the first choice for remote sensing inversion of crop parameters. The development of spectroscopy has improved the accuracy of spectral information, and the methods of first-order derivative, continuum removal, wavelet transform and smoothing algorithm (such as Savitzky-Golay smoothing), etc. are applied to the original spectra, which make the spectral calculation more refined and the results more accurate [95,108,109]. Compared with traditional vegetation indices (VIs), the construction parameters of spectral indices are no longer limited to band reflectance, but can also be red edge parameters, other spectral indices, etc. (Table 1) [26,110,111].

For multispectral/hyperspectral remote sensing data, spectral indices are usually constructed by selecting appropriate bands in the visible red region and NIR region. For different varieties of crops, there are significant differences in the slopes of normalized difference vegetation index (NDVI)-based LNC models, and it is difficult to achieve uniform regression analysis across different crops varieties with conventional indices. It has been shown that NDVI_{(1220, 610)} is a good index to estimate LNC in both rice and wheat, with RMSE all less than 13.04% [34], and NDVI_{(1220, 710)} achieves high precision estimation of N status for different varieties of rice [112]. However, the model has not been extended to other regions for validation, and the accuracy and stability of the spectral index remains to be explored. In addition, the use of a multi-band vegetation index for LNC monitoring in rice and wheat is more effective [92,113,114]. Wang et al. [113] used a three-band vegetation index combining NIR, red edge and blue bands to estimate LNC for rice and wheat with R² of 0.866 and 0.883, respectively, which were 17.66% (rice) and 7.68% (wheat) more accurate than NDVI, and 40.13% (rice) and 16.18% (wheat) more accurate than RVI. Tan et al. [92] explored the relationship between LNA and parameters such as first-order derivative sum (SD), first-order derivative maximum (D), etc. in VIS and red edge regions, and the new normalized index (SD_r − SD_b)/(SD_r + SD_b), which was constructed by integrating information from multiple bands, was a good fit for wheat LNA (R² = 0.935) and was applicable to wheat N inversion in different varieties and regions.

The N estimation model used fixed VIs performed consistently over the same growth stage, but when pooling data from multiple growth stages, accuracy decreased significantly. Canopy structure and background conditions changed as the growth stage progressed, but in all cases, whole leaf pixels showed more stable performance than light and shade leaf pixels [87,115,116]. Traditional spectral indices cannot fully capture the intrinsic relationship between canopy spectra and N in growth stage, and how the predicted intensity of N models varies across growth stages has not been fully explored [117]. The red edge chlorophyll index (CI_{red edge}) is sensitive to canopy N content and can effectively mitigate the effect of canopy structure on canopy N estimation. Li et al. [118] combined NDVI and CI_{red edge} to construct a nitrogen planar domain index (NPDI) with good predictive ability for canopy N uptake in wheat, corn and both combinations. Palka et al. [119] constructed a regression model by combining the canopy chlorophyll content index (CCCI) and the canopy nitrogen index (CNI), and modified the CCCI-CNI to extend N estimation to the end of booting stage for wheat. Quantifying the spectral contribution in the mixed image elements using spectral mixture analysis (SMA) can improve the spectral accuracy, and the spectral indices constructed in this way show superior capability in all growth stages and even in the early evaluation of LNC [116,120]. The selection of sensitive multi-band and multi-index information fused to form a spectral index improves the sensitivity of N inversion, and to some extent overcomes the inconsistency between indices and N for different varieties of crops and different fertility stages.

Based on RGB data, the response of different image indices to N varies widely and has limited response [121]. The RGB three-color channels of UAV images each contain luminance information and are susceptible to interference from lighting conditions. In contrast, in the HSV and Lab color spaces, V and L denote value and luminosity, respectively, and the transformation of RGB to them weakens the influence of luminance information through nonlinear changes, resulting in a more sensitive response to LNC [122]. Nevertheless, RGB is limited by spectral accuracy and still has difficulties in index improvement. In addition to the color information extracted from spectral data, the texture information can be obtained from the local variance function, which reflects the variation relationship between several pixel points and characterizes the canopy structure. Fusing two or more datasets with different feature information can provide a more comprehensive interpretation of the relationship between remote sensing information and N status [123]. Therefore, for RGB data, the “image-spectrum” fusion indices formed by fusing image indices and texture features improves the sensitivity of image data to N features, and the investigation of its ability to diagnose N status is a major development direction at present [60,82]. When extracting texture features through Gray-level Co-occurrence Matrix (GLCM), crop cultivation patterns may cause differences in texture information metrics for N content in different directions, and using texture information calculated along the perpendicular to the row direction to monitor row-grown crops has the best results [85]. In addition, adding depth information to RGB images can break through the limitations of extracting canopy structural features from 2D images [86,124]. Xu et al. [124] fused texture features with 3D structural information and RVI to invert N status with better accuracy and stability, the LNA prediction accuracy of 0.74 during whole growth stages. By adding 2D and 3D structural information, the background and saturation effects can be better reduced, and the N inversion information of the crop canopy can be enhanced.

Table 1. Spectral indices for assessing crop N status.

Index	Formula	Reference
Nitrogen Reflectance Index	$\frac{{\left(R_{800}/R_{550}\right)}_{\mathit{target}}}{{\left(R_{800}/R_{500}\right)}_{\mathit{reference}}}$	[125]
Red Edge Position: Linear Extrapolation Method	linear extrapolation of two straight lines on the derivative spectral curve (lines formed by 680 nm and 694 nm, and formed by 732 nm and 760 nm)	[126]
Normalized Difference Red Edge	$\frac{R_{790}\hspace{0.17em}-\hspace{0.17em}{R}_{720}}{R_{790}\hspace{0.17em}+\hspace{0.17em}{R}_{720}}$	[33]
Double-peak Canopy Nitrogen Index	$\frac{\left(R_{720}\hspace{0.17em}-\hspace{0.17em}{R}_{700}\right)/\left(R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\right)}{R_{720}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\hspace{0.17em}+\hspace{0.17em}0.03}$	[127]
Nitrogen Planar Domain Index	$\frac{\mathit{Measured}C{I}_{\mathit{red}\mathit{edge}}\hspace{0.17em}-\hspace{0.17em}C{I}_{\mathit{red}\mathit{edge}}\_MIN}{C{I}_{\mathit{red}\mathit{edge}}\_MAX\hspace{0.17em}-\hspace{0.17em}C{I}_{\mathit{red}\mathit{egde}}\_MIN}$	[118]
Water Resistance Nitrogen Index	$\frac{\left(R_{735}\hspace{0.17em}-\hspace{0.17em}{R}_{720}\right)R_{900}}{R\left(R_{930}\hspace{0.17em}-\hspace{0.17em}{R}_{980}\right){\left(R_{735}\hspace{0.17em}+\hspace{0.17em}{R}_{720}\right)}_{min}}$	[128]
Canopy Chlorophyll Content Index	$\frac{\mathit{NDRE}\hspace{0.17em}-\hspace{0.17em}{\mathit{NDRE}}_{\mathit{min}}}{{\mathit{NDRE}}_{\mathit{max}}\hspace{0.17em}-\hspace{0.17em}{\mathit{NDRE}}_{\mathit{min}}}$	[129]
Modified Chlorophyll Absorption Ratio Index	$\frac{R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\hspace{0.17em}-\hspace{0.17em}0.2\left(R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)}{R_{700}/R_{670}}$	[130]
MCARI/MTVI2	$\mathrm{MCARI}=\frac{R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\hspace{0.17em}-\hspace{0.17em}0.2\left(R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)}{R_{700}/R_{670}}$  $\mathrm{MTVI}2=\frac{1.5\left(1.2\left(R_{800}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)\hspace{0.17em}-\hspace{0.17em}2.5\left(R_{670}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)\right)}{\mathit{sqrt}\left({\left(2R_{800}\hspace{0.17em}+\hspace{0.17em}1\right)}^2\hspace{0.17em}-\hspace{0.17em}\left(6R_{800}\hspace{0.17em}-\hspace{0.17em}5\mathit{sqrt}\left(R_{670}\right)\right)\hspace{0.17em}-\hspace{0.17em}0.5\right)}$	[131]

$R_i$ stands for reflectance at wavelength i nm.

Table 1. Spectral indices for assessing crop N status.

Index	Formula	Reference
Nitrogen Reflectance Index	$\frac{{\left(R_{800}/R_{550}\right)}_{\mathit{target}}}{{\left(R_{800}/R_{500}\right)}_{\mathit{reference}}}$	[125]
Red Edge Position: Linear Extrapolation Method	linear extrapolation of two straight lines on the derivative spectral curve (lines formed by 680 nm and 694 nm, and formed by 732 nm and 760 nm)	[126]
Normalized Difference Red Edge	$\frac{R_{790}\hspace{0.17em}-\hspace{0.17em}{R}_{720}}{R_{790}\hspace{0.17em}+\hspace{0.17em}{R}_{720}}$	[33]
Double-peak Canopy Nitrogen Index	$\frac{\left(R_{720}\hspace{0.17em}-\hspace{0.17em}{R}_{700}\right)/\left(R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\right)}{R_{720}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\hspace{0.17em}+\hspace{0.17em}0.03}$	[127]
Nitrogen Planar Domain Index	$\frac{\mathit{Measured}C{I}_{\mathit{red}\mathit{edge}}\hspace{0.17em}-\hspace{0.17em}C{I}_{\mathit{red}\mathit{edge}}\_MIN}{C{I}_{\mathit{red}\mathit{edge}}\_MAX\hspace{0.17em}-\hspace{0.17em}C{I}_{\mathit{red}\mathit{egde}}\_MIN}$	[118]
Water Resistance Nitrogen Index	$\frac{\left(R_{735}\hspace{0.17em}-\hspace{0.17em}{R}_{720}\right)R_{900}}{R\left(R_{930}\hspace{0.17em}-\hspace{0.17em}{R}_{980}\right){\left(R_{735}\hspace{0.17em}+\hspace{0.17em}{R}_{720}\right)}_{min}}$	[128]
Canopy Chlorophyll Content Index	$\frac{\mathit{NDRE}\hspace{0.17em}-\hspace{0.17em}{\mathit{NDRE}}_{\mathit{min}}}{{\mathit{NDRE}}_{\mathit{max}}\hspace{0.17em}-\hspace{0.17em}{\mathit{NDRE}}_{\mathit{min}}}$	[129]
Modified Chlorophyll Absorption Ratio Index	$\frac{R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\hspace{0.17em}-\hspace{0.17em}0.2\left(R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)}{R_{700}/R_{670}}$	[130]
MCARI/MTVI2	$\mathrm{MCARI}=\frac{R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{670}\hspace{0.17em}-\hspace{0.17em}0.2\left(R_{700}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)}{R_{700}/R_{670}}$  $\mathrm{MTVI}2=\frac{1.5\left(1.2\left(R_{800}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)\hspace{0.17em}-\hspace{0.17em}2.5\left(R_{670}\hspace{0.17em}-\hspace{0.17em}{R}_{550}\right)\right)}{\mathit{sqrt}\left({\left(2R_{800}\hspace{0.17em}+\hspace{0.17em}1\right)}^2\hspace{0.17em}-\hspace{0.17em}\left(6R_{800}\hspace{0.17em}-\hspace{0.17em}5\mathit{sqrt}\left(R_{670}\right)\right)\hspace{0.17em}-\hspace{0.17em}0.5\right)}$	[131]

$R_i$ stands for reflectance at wavelength i nm.

3.3. Retrieval Methods of Canopy N Status

For crop N status assessment and monitoring, the modeling methods can be divided into statistical analysis, physical analysis, and hybrid methods. Statistical analysis is to obtain high-precision N diagnosis by establishing mathematical relationships between ground truth data and remote sensing spectral information, which can be further classified into two categories: traditional statistical methods and machine learning methods. Physically based methods take the structural information and physicochemical properties of plants and leaves as input parameters, simulate the process of radiation absorption and scattering inside the crop, and form the reflection spectrum and output, so as to establish the correlation between crop parameters and ground reflection spectrum. The models are divided into radiative transfer models (RTM) and geometric optical models, and RTM are often used in research because of the continuous and uniform distribution of rice and wheat in cultivation. Hybrid models have advantages by using multiple categories of models in combination but are currently used rarely in N retrieval (discussed in Section 5).

3.3.1. Traditional Statistical Methods

Traditional statistical methods have simple mechanisms, and their inverse accuracy depends more on the rationality of modeling parameters, thus they cannot overcome the influence of environmental and other factors, and there are difficulties in transferring prediction models to other datasets. However, due to the simplicity and convenience of the model and its usability, it is still widely used in the field of crop N monitoring at present. Univariate linear regression and its nonlinear transformation, as the basis of statistics, have achieved good results in exploring whether there is a significant correlation between crop N indicators and individual VIs, and occupy an important position in N status inversion [104,132,133]. Hansen et al. [89] found that applying partial least squares regression (PLSR) to fit N status gave equal and better results than exponential regression, with the R² maximum increase of 23%. PLSR was considered a good alternative to univariate statistical models. When multiple growth stages are involved or when there is a lack of phenological information, individual index that do not fully utilize spectral data cannot describe the relationship between N and spectral information, so using multiple linear regression (MLR) is a better choice [134]. MLR can reduce model chance. Whether the parameters involved in modeling exist in response to N and whether there is overfitting between parameters is the key to influencing the multiple regression model. Pearson correlation coefficient is generally used as a measure in the study, and the combination of parameters with larger correlation coefficients is selected to construct MLR models with good interannual scalability [64,69]. PLSR, principal component analysis (PCA), stepwise multiple linear regression (SMLR), and Ridge Regression, etc., play an advantage in dealing with the collinearity problem between parameters. Many co-linear spectral variables were reduced to a few uncorrelated latent variables to avoid overfitting problems [91,123,135]. The selection of a suitable variable dimensionality reduction method based on the quantitative relationship between the number of samples and the dimensionality of variables, combined with multiple regression, significantly improves the inversion accuracy and applicability.

Traditional statistical methods can describe different rates of change between N status and spectral information, providing fitting models for a variety of change conditions. For region-specific datasets, models with good inversion effects can be obtained by comparing different regression methods; for datasets with different growing environments, the best regression models derived from the study will be different [64,89,104,134]. Although the statistical models are not stable enough to overcome environmental problems, the different regression methods in the statistical models are more consistent in principle, have specific mathematical relationships, are easy to understand and apply, and are extremely convenient to use in fixed research areas. Therefore, the traditional statistical methods still dominate the existing N remote sensing monitoring studies.

3.3.2. Machine Learning Methods

Machine learning models are gaining widespread attention for their ability to handle large amounts of input data from multiple platforms and to solve nonlinear tasks. Artificial Neural Network (ANN) and Back Propagation Neural Network (BPNN) are commonly used models for remote sensing estimation of crop N status, which can automatically extract relevant features from data. However, in practical applications, a large training dataset is required, and the number and size of the implied layers, training efficiency, and overfitting are considered. Yang et al. [96] used Gaussian radial basis function as the implied layer of the neural network to avoid the tedious calculation and overfitting phenomenon of BPNN, with structural adaptive features, good generalization ability and fast learning convergence speed, and more stable and reliable application. When constructing neural network models directly, the differences in results for different types of parameters are not obvious, but the accuracy is significantly improved after using PCA for model input parameters [136,137,138,139]. The combined use of PCA and machine learning methods shows unique advantages and promising applications.

Support Vector Machine (SVM) is extremely effective for analytically solving high-dimensional data problems. Yao et al. [140] applied traditional regression analysis, ANN and SVM, to compare the prediction accuracy, computational efficiency and complexity level of different methods for inversion of wheat LNC. The results showed that the machine learning models were more accurate, with the SVM method being more stable in dealing with potential confounding factors for most varieties, ecological niches, and growth stages. The kernel function in SVM is the focus of attention, and the multiple-kernel support vector regression (MK-SVR) plays an advantage in estimating N status at different growth stages because it combines the advantages of local kernel function and global kernel function [141]. However, complex optimization algorithms can reduce the computational efficiency of SVM and using a combination of least squares and SVM methods, LS-SVM can solve linear or nonlinear multivariate estimation capability in a relatively fast way, significantly improving the computational efficiency of SVM [17,141,142].

In the presence of weak a priori knowledge, Gaussian Processes Regression (GPR) can perform adaptive nonlinear fitting of complex datasets with flexible probabilistic Bayesian models and simpler parameter optimization applied to crop N status inversion [11,93,143]. Random Forest (RF) integrated with decision trees as the basic unit can rank the importance of variables, reduce redundancy in high-dimensional datasets, and have high stability, with vast application prospects [37,144]. When the entire spectral range of a single band is used as an input variable, the accuracy of regression by RF inversion (R² = 0.89) is higher than that of univariate regression with existing VIs; when VIs is used as input features, model accuracy is improved with R² of 0.95 [145]. Determining the appropriate input dataset is a key element to exploit the predictive power of the model.

Machine learning methods techniques can be used to reveal the physiological and structural characteristics of plants, and can respond to dynamic differences in physiology due to environmental influences [144]. The study is no longer limited to conventional machine learning models, but improves the models starting from input datasets [138,145], model parameters [96,144], functions and structures [17,96,141], which not only improves the efficiency of data analysis, but also enables higher accuracy N status analysis, making it more efficient and flexible in N monitoring. The input variables have diversified from single spectral information to mathematically transformed spectral data, spectral indices, and texture information, etc. while the machine learning methods are expanding toward efficiency, accuracy, and speed.

3.3.3. Physically Based Methods

RTM use optimization algorithms for inversion to infer the N content of crops from observed spectral data. Depending on the scale of the object of study, it is mainly divided into leaf radiative transfer models and canopy models. At the leaf scale, the PROSPECT model is the most widely used and has been continuously optimized and improved [146,147,148]; at the canopy scale, the SAIL model is one of the first models applied, mainly for uniformly distributed continuous vegetation surfaces [149]. Combining PROSPECT and SAIL models to invert vegetation physiological and biochemical parameters is a common approach nowadays, mostly around canopy and LCC, water content and leaf area [29,75,150,151]. Most studies derive crop N status through empirical relationships based on significant associations between leaf area or Chl and N. Despite the long-standing stability and reliability of RTM in the inversion of physicochemical parameters, different configurations of the model by users may lead to equally plausible results [152,153], so it is necessary to constrain the model using a priori information. Combining the DSSAT cropping system model CSM and PROSAIL model, complementing the interaction between crop growth stages and the environment for the constraints of the input parameters of the PROSAIL model, plays a unique advantage in the inversion of crop physicochemical parameters, not only with high accuracy, but also with the statistics of physicochemical parameters among different varieties of crops [154].

Yang et al. [155] used N uptake coefficients to equivalently replace the Chl uptake coefficients in the original PROSPECT model, and established the N-PROSPECT model based on the PROSPECT model to directly invert leaf N content. The N-PROSAIL model, established by combining the N-PROSPECT model and the SAIL model, achieves the diagnosis of N status at the leaf and canopy scales, and reduces the model error by setting a priori parameters at different growth stages [156]. The RTM expresses the crop growth process from a physical point of view, which is more stable in the inversion, but has the problem of being time-consuming. Combined with the Lookup Table (LUT) it can reduce the computational demand. Li et al. [157] constructed a multi-LUT for wheat LAI, LND and two spectral indices (MSR and MCARI/MTVI2), which not only reduced the LUT size and improved the computation time, but also had better accuracy of N estimation. On the other hand, since protein is also a major N-containing component in crops, coupling protein specific absorption coefficients into the PROSPECT model to form PROSPECT-PRO, which is combined with the 4SAIL model to form PROSAIL-PRO, can also be used for crop N status diagnosis [31,158]. RTM with strong explanations is better expressed in the inversion, but because the model expression depends on the input of more parameters and complex computational process they are less used in current research. Reducing the complexity of models and complementing the advantages of statistical models, hybrid RTM and machine learning models have become a future research need.

4. Influential Factors on Accuracy of Remote Sensing Monitoring of Canopy N

Spectral reflectance information has been shown to be sensitive to the N content of canopy leaves, but differences in data acquisition, vertical distribution of leaf N, dynamic changes in N during the growth stages, and physiological differences between different plants can all have an impact on the correlation between crop spectral information and canopy N indicators.

4.1. Differences in Data Acquisition Angles

Both portable spectrometers and UAV sensors usually acquire spectral data for crop N monitoring within a range of VZAs, which can vary by 30° and more. The spectral indices in crop N status studies are usually developed from vertical angle data. However, due to the variation in the angle of view of data acquisition caused by different experimental conditions, and the anisotropy of vegetation reflectance, the accuracy and robustness of using these indices directly to estimate the N content are not sufficient. The reflectance in the VIS, red edge and NIR bands decreases gradually from VZAs from −60° to 0°, with relative changes in reflectance ranging from 34.7% (+60°) to 265.5% (−60°), and 81.7% (+60°) to 89.3% (−60°) in the VIS and NIR bands, respectively [53]. Therefore, developing an index that is sensitive to N content and insensitive to VZAs is of great practical importance to adapt to different experimental conditions, improve prediction accuracy and enhance model stability.

Higher viewing angles allow better extraction of crop biochemical information compared to the nadir orientation [159]. The change in view angle significantly affects canopy reflectance, especially in the red and NIR bands, which in turn makes VIs based on these spectral bands sensitive to angle [160]. The introduction of angle-insensitive bands to construct indices, such as the normalized difference red edge (NDRE), the green and blue bands, and the green band Chlorophyll Index (CIgreen), can improve the accuracy of canopy N inversion of different VZAs remote sensing images, significantly expanding the range of suitable viewing angles for determining crop N status by remote sensing, and thus adapting to the differences between different experimental conditions [53,161,162]. For the angular insensitivity index cannot be simply attributed to the effect of a single band; green, blue and NIR bands may have played a joint role in improving the index adaptation. It is difficult to obtain accurate spectral collection perspectives in the applications. A unified N monitoring model under a range of perspectives can help with the flexible application of crop N diagnosis. Like other VIs, Li et al. [53] developed angular insensitivity vegetation index (AIVI) to have the best LNC estimation accuracy at −20° view angle, but at the same time the correlation between AIVI and LNC has high stability at −10° to −40° with R² of 0.83. Similarly, floating-position water band index (FWBI) has the highest correlation with LNC at −10° view angle (R² = 0.852), also has superior N content estimation accuracy at 0° to 30° (R² = 0.835) [161]. The statistics on angular differences show that back-scatter direction has better LNC prediction accuracy than the forward-scatter view angle.

However, the spectral information obtained by whatever VZAs inevitably has information such as soil background, light shading, etc. Different growth stages and different light conditions will change the crop spectral reflectance, which is a common noise in inversion. The study reduces their effects by spectral preprocessing such as first-order differentiation and wavelet transform, suitable vegetation index, and threshold segmentation [27,63,82,93]. The water background of rice is a unique feature that differs from other crops; water has an absorption effect on the NIR band, and when the canopy cover is small, the water depth and turbidity have an isotropic effect on the spectral reflectance of the red-edge region [163]. Therefore, when converting reflectance to vegetation index, this effect can be eliminated or attenuated by calculating between multiple bands. The individual N content was significantly improved in accuracy before and after removal, and the group indicators indicated the total amount per unit area, which was less influenced by background noise and had a smaller enhancement effect [82].

4.2. Vertical Distribution of Leaf N

When the canopy leaves are taken as a whole object, most of the studies are carried out for the upper leaves, ignoring the vertical heterogeneity of the canopy. The LNC within the canopy is not constant from the top to the bottom of the leaf layer, and it varies with growth stages. At jointing stage, flowering stage and filling stage, LNC decreases from top to bottom; and at booting stage, it tends to increase and then decrease [43]. Moreover, it is difficult to effectively identify the information at the bottom of the canopy by acquiring the crop reflectance spectra vertically due to the influence of canopy leaf cover at different growth stages. At present, we have achieved better results in exploring the canopy N content by acquiring reflectance spectrum vertically or from multiple angles to overcome the stratification differences.

The PLS algorithm has better estimation capability for different levels of leaf N status, then becomes an effective tool for early N monitoring [40,164,165,166]. Huang et al. [164] combine NRI and NPCI to construct a PLSR model which could better retrieve foliage N density in different leaf layers (R² > 0.67). He et al. [166] also demonstrated that PLSR estimates LNC accuracy better than BPNN and eXtreme gradient boost (XGBoost). However, for the studied spectral information, which must contain information from different leaf layers, it is especially important to determine the contribution of different leaf layers to the spectral reflectance of the canopy. Studies have begun to explore the characteristics of the vertical distribution of N in the canopy of crops at different growth stages, and to develop an effective method for estimating N in each leaf layer or total N in the canopy by determining the correlation between different leaf layers and N status [40,43,44]. Duan et al. [43] used a calibration coefficient to adjust the relationship between the effective layer of remote sensing detection and the whole canopy, and then developed a method for estimating the overall canopy LNC based on GI, mND705 and NDVI. He et al. [44] estimated the canopy top LNC by NDRE, then inputting the results into the LNC vertical distribution model to get the model coefficients; thus the model based on the relative canopy height could obtain LNC in different leaf layers (LNC_Li), which was superior because of fewer parameters and higher accuracy. The short plant size of rice and wheat crops and the small vertical distance between different leaf layers can easily mask differences in the spectral response of canopy N status changes [54]. Compared with vertical remote sensing observation, multi-angle observation can reduce the information bias of fixed viewpoints. Using different combinations of VZAs, Wu et al. [54] were able to retrieve LCC in the upper-layer (VZA 10°), middle-layer (VZA 10° and 30°) and bottom-layer (VZA 10°, 30° and 50°) of the plant, respectively. Based on the response of spectral indices to each leaf layer of N status at different VZAs, selecting the best VZAs or combination of VZAs can realize the complementation of canopy spectral information so that the accuracy of crop N monitoring can be more robust and accurate [40,54].

Multi-angle stereoscopic observation can obtain more vertical information about the plant, but when the bottom leaves are too low, the influence of soil background and crop residues, etc., will increase. Therefore, it is necessary to determine the effective depth of crop canopy spectroscopy observation and realize the inversion model of vertical distribution of N content in canopy from “surface” to “three-dimensional”. At the same time, the multi-angle measurement will generate a huge amount of data, and how to quickly extract the effective information from it has become an urgent problem to be solved.

4.3. Dynamic Changes in N during the Growth Stages

N content in crops is a long-term accumulation process that changes as growth stages. The correlation between N and spectral information varies at different stages, and an ideal N inversion model needs to overcome the effects of phenological variability and accurately estimate the N content of the crop at different growth stages. Throughout the crop’s reproductive stages, temperature levels affect photosynthesis and metabolic processes that are closely related to N assimilation and utilization. Therefore, it is necessary to introduce meteorological information to construct a dynamically changing model.

Crop models such as CERES and APSIM, which are widely used around the world, simulate crop growth processes by inputting meteorological data and field management data, etc., and are important guides for real-time diagnosis of crop N nutrition status under different cropping conditions [167,168,169]. The construction of growth models relies on numerous experimental parameters, which are data-intensive and cumbersome to process. Cao et al. [170] dynamic obtained relative growing degree days (GDD) based on the physiological development time of crops to participate in modeling, and quantified the model parameters that reduced the effects brought about by different indicators. In addition, the field spectral information obtained based on remote sensing data is accurate, and quantitative analysis of temporal variation between VIs is more conducive to the dynamic monitoring of N status. Double Logistic functions and Gaussian curves fitted to time-series data can effectively describe crop growth and senescence processes [171,172,173]. By combining effective accumulated temperature and crop growth parameters, such as the spectral index NDRE, a model for monitoring the entire growing period of the winter wheat canopy, constructed with the growing degree-days as a moderating factor, offers the possibility for N estimation throughout the growth stages [174]. Dynamic curves of indices such as NDVI, constructed using accumulated growing degree days (AGDD) as a time driver, provide a reference for N nutrition diagnosis at different periods [172]. Combining multi-temporal VIs with key phenological indicators, the constructed dynamic model has clear biological significance, which not only facilitates crop N monitoring but also significantly enhances the ability of N status early prediction.

4.4. Physiological Differences in Plants

N synthesis is influenced by canopy structure, photosynthesis-related pigments, and water content, etc., and confounding effects are common in canopy reflectance under physiological stress, making N remote sensing monitoring challenging.

Water is the carrier of N transport in the crop and is closely related to the N status of the canopy. Exploring the spectral response to N and water shows that the spectral reflectance of wheat treated under low water conditions increases in both regions of SWIR, with less difference in the NIR region under water differences [27,59]. Thus, indices constructed based on SWIR and NIR reflectance are better able to show differences between N levels in different water treatment environments. The SWIR region contains more water-related information, but sensors that include SWIR are costly so it is relevant to explore water-sensitive bands in the VIS–NIR range. Under the condition of constant N content, the red edge reflectance of crops with different water treatments tends to be the same [128]. Based on the red-edge correlation indices such as NDRE and normalized pigment chlorophyll index (NPCI), the introduction of water-related indices, such as floating-position water band index (FWBI), crop water stress index (CWSI), etc., can significantly improve the interaction between water deficit and N nutrition [128,175,176,177]. Whether the multi-analysis based on specific VIs or the whole spectrum, studies have been conducted to separate N and water information in the spectrum by reducing spectral mixing effects, thus improving the estimation accuracy of crop N under the influence of water.

Crop growth parameters are not independent of each other and may correlate under different circumstances, such as correlation between N, Chl and LAI when the canopy cover is small. However, when Chl and LAI change driven by other external conditions, there will be errors in estimating N status based on their correlation with N. Research needs indices that are both sensitive to N and resistant to interference from other factors. The combination of two VIs with different sensitivities to Chl and LAI, whose ratios can minimize the effect of LAI and have a better correlation with Chl, such as the joint indices modified chlorophyll absorption ratio index and second modified triangular vegetation index in ratio (MCARI/MTVI2) [131], the red-edge-chlorophyll absorption index and the triangular vegetation index in ratio (RECAI/TVI) [111], the transformed chlorophyll absorption reflectance index and optimized soil-adjusted vegetation index in ratio (TCARI/OSAVI) [177]. The red edge region is influenced by LAI which cannot estimate N well in complex situations. Chen et al. [127] found that there is a double-peaked phenomenon in the first-order derivative spectrum of the red edge region, and that changes in N concentration can be amplified as changes in the relative height of the double-peaked peaks, with which the proposed double-peaked canopy nitrogen index (DCNI) can overcome the influence of LAI. There is a certain similarity between different spectral indices and different physiological parameters, which show hierarchy and aggregation in statistical analyses. Exploring the different relationships that may exist between parameters is therefore important for exploring N status in crops grown under different growing conditions.

5. Challenges and Perspectives

After decades of development, the techniques for remote sensing monitoring of canopy N have made rapid progress and achieved good results, but there are inevitably many difficulties and challenges that need to be addressed.

(1) The development of multi-source data integration from “satellite–airborne–ground” to meet the needs of high-precision monitoring at all scales. In recent years, remote sensing-based crop N monitoring and assessment research has been conducted mainly at the laboratory and field scale, applying small sensor platforms based on ground-based spectral instruments and UAV to acquire data. Research on large farms, counties, cities, or larger regional scales rely on satellite-based multispectral data. Multispectral data are affected by radiation and atmosphere during acquisition, making processing more difficult. In the face of complex geography, the low-resolution images are prone to mixed pixels, making it difficult to achieve accurate estimates of N status. Hyperspectral data are limited at large scales due to their access and data volume, so how to achieve high accuracy monitoring and assessment of crop N over large scale areas is a major challenge currently faced.

Currently, the research on crop N estimation from UAV data is beginning to bear fruit, with good validation in small farm applications. A summary of the spectral data acquisition platforms and their inversion N in existing studies is shown in

Table 2. Summary of the data platforms, retrieval methods, and research results of the studies cited in the body.

Crop	N Status	Data Platforms	Indices	Retrieval Method	Results	References
Wheat	LNA	ASD FieldSpec Pro	Spectral bands VIs	PLSR SVM RF	R2 = 0.895 RMSE = 0.903 g/m2	[109]
Rice	LNC CNC	ASD FieldSpec Pro2500 ASD FieldSpec3	Spectral bands VIs	GPR RF, GPR-RF SVR, GPR=SVR	R2 > 0.94 NRMSE < 6%	[52]
Wheat	CND	ASD FieldSpec Handheld UAV (UHD 185 Firefly)	VIs	N-PROSAIL	Field: R2 = 0.83 RMSE = 0.23 UAV: R2 = 0.74 RMSE = 0.26	[157]
Wheat	N concentration N content	ASD FieldSpec HandHeld	Spectral bands VIs	Statistical analysis	N concentration: R2 = 0.81 RMSE = 0.72% N content: R2 = 0.96 RMSE = 0.83 g/m2	[119]
Wheat	LNA	ASD FieldSpec HandHeld 2 RealSense depth camera D435i	VIs Texture	MLR BP	R2 = 0.74 RRMSEs = 40.13%	[124]
Rice	LNC	UAV (AIRPHEN multispectral camera)	VIs	Linear spectral mixture analysis Statistical analysis	R2 = 0.78 RMSE = 0.26% RMSE = 10.4%	[116]
Wheat	LNC	UAV (hyperspectral camera)	VIs Texture	RR PLSR SVR RF	R2 = 0.84 RMSE = 0.25	[123]
Wheat	LNC LNA	ASD FieldSpec Handheld 2 RealSense depth camera D435i	VIs Canopy structural	PLS RF	LNC: R2 = 0.78 RMSE = 0.35% LNA: R2 = 0.79 RMSE = 1.54 g/m2	[86]
Rice	NNI	UAV (Parrot Sequoia camera)	Spectral bands VIs	LR, SMLR RF SVM ANN	RF accuracy is the highest: R2 = 0.61 (stem elongation stage) R2 = 0.79 (heading stage) RMSEs = 0.09	[37]
Rice	LNC PNC LNA PNA	UAV (Cubert S185 hyperspectral camera)	Spectral bands VIs	LR, MLR PLSR ANN RF SVM	At single growth stage, LR estimation N status based on VIs has the highest accuracy; at multiple growth stages, PLSR and ML are better.	[64]
Wheat	N content	HyMap sensor	Spectral bands	PROSAIL-PRO GP Heteroscedastic GP	RMSE = 2.1 g/m2 The optimal N retrieval spectral bands are in the SWIR.	[31]
Wheat	LCC	Landsat8 ASD FieldSpec Pro	VIs	LR PROSPECT SAIL	Use hyperspectral leaf reflectance data to simulate Landsat-8 bands LR: R2 = 0.59 PROSPECT: R2 = 0.64	[75]
Wheat	LCC	Sentinel-2 RapidEye EnMAP	Spectral bands	PLSR	Sentinel-2: R2 = 0.755 RapidEye: R2 = 0.689 EnMAP: R2 = 0.735	[74]
Wheat	LNC	ASD FieldSpec HandHeld	VIs	Statistical analysis	AIVI could overcomes the impact of VZAs: R2 = 0.84 at −20° R2 = 0.83 at −10° to −40°	[53]
Wheat	LNC	ASD FieldSpec	VIs Spectral bands	VIs BPNN XGBoost PLSR	R2 ≥ 0.83 at 0° to −30° VZAs range The accuracy of PLSR is better than VIs (16–17%), BPNN (15–16%) and XGBoost (29–58%) at VZAs ±60°	[166]
Rice	LNCLi	ASD FieldSpec4	VIs	Vertical distribution model	LNCL1: R2 = 0.768 LNCL2: R2 = 0.700 LNCL3: R2 = 0.623 LNCL4: R2 = 0.549	[44]
Wheat	NNI	ASD FieldSpec Micro-Hyperspec and NIR-100 imager SC655 thermal camera	VIs Thermal indices	Statistical analysis	The combination of CCCI and DWI can overcome the influence of water to retrieve NNI, and the RMSE is reduced to 0.109.	[59]

Table 2 covers case studies from different regions.

. UAV-based research can not only analyze spectral information (sensitive spectra, spectral mathematical transformations, spectral combination calculations, etc.), but also extract image information (texture information, color information, etc.), which shows advantages of high precision due to its high spatial and temporal resolution and the amount of representation information. Satellite-based imaging data are limited in depth, mostly only extract VIs result in low inversion accuracy, yet satellite data is still the most important source of data for large-scale studies. In the existing research situation, the rapid development of sensor technology and remote sensing platforms has extended the scale of research to medium and large farm areas. The results of existing UAV-scale research results are translated to municipal, provincial, national, and even larger scales through algorithms such as multi-scale analysis and reconstruction, and spatiotemporal data fusion, thus enabling N monitoring over large areas. Therefore, the fusion of multiple sources of data from “satellite–airborne–ground” is the basis for large scale applications with inversion accuracy [178]. At present, the spectral resolution of the red edge and NIR bands (N-sensitive bands) in satellite-based hyperspectral sensors is insufficient, and the spatial resolution and revisit period are not advantageous. In this context, the transformation of ground-based research results and the development of high-precision satellite-based hyperspectral sensors deserve even more attention. From the perspective of data acquisition, the complementary advantages of the multiple types of data from “satellite–airborne–ground” could break the limits of geographical scope, and then enable high-precision monitoring and assessment of crop N status at all scales.

(2) Research still has bottlenecks in monitoring crop N in the presence of confounding factors. Under N stress, the spectral properties of vegetation leaves change, and N monitoring is achieved through crop spectral information obtained from ground-based observations by remote sensing technology. However, in practical applications, the inconsistency of crop growth conditions can lead to irregular overall crop deficiencies in water, fertilizer, and cause pests and diseases, all of which could generate yellowing and wilting of crop leaves. The changes in the external structure and intrinsic characteristics of the canopy, result in corresponding changes in the spectral reflectance characteristics [178]. To simply attribute spectral changes to canopy N content would be a misjudgment. Studies are generally set up with variable conditions for different growth stages, locations, field management, species, or plant types to test the stability of the model. However, there is insufficient evidence that the method is effective in overcoming spectral variation due to physiological differences. The primary way to achieve interpretability and practical application of the model is to start with the principles and isolate the influencing factors [179,180]. When considering only N and another stress factor, overcoming the effects of water can increase R² to 0.843 [128]. However, few studies have quantified and differentiated the contribution of leaf biochemical content (including water, diseases, other pigments, etc.) to the spectral band from spectral perspective.

In addition to physiological characteristics of vegetation, differences in soil background and canopy structure can cause difficulties in extracting whole leaf pixels, and noise from atmospheric transport processes also affects spectral accuracy. Existing studies mainly consider the effects of soil background and fractional vegetation cover (FVC). Before and after removing the background pixels, the accuracy of remote sensing inversion of crop biochemical parameters was significantly improved; for example, R² in LAI inversion could improve 0.27 [181], in N status inversion could improve 0.11 [82], and in Chl status inversion could improve 0.10 [182]. However, the applicable method of background elimination is also extremely important, as it requires high performance to adapt to the complex and changing field environment [182]. FVC correlates with background information and combining this information with spectra can also improve N estimates in different environmental contexts [183,184]. In addition, most imaging systems use top or side views to collect data, and the anisotropy of spectral information leads to different responses to crop N, and a suitable angle of spectral acquisition is important for accurate N monitoring. The blue band has atmospheric function, when the N content estimation model combined with blue band and other N-sensitive bands can improve the adaptability of the angle of data acquisition [53]. Under multiple conditions such as data acquisition, environmental stress and crop physiological stress, the spectral information is mixed with numerous non-target factors, which need to be decomposed to determine the precise response of crop N to the spectrum. Therefore, various influencing factors should be considered when estimating crop N by remote sensing to achieve high precision diagnosis.

(3) Improving the generalizability of models is key to crop N monitoring and assessment. To summarize the currently used models for inversion of N status and their effectiveness, a statistical model is currently the most commonly applied method in research experiments. When using statistical models, it is first necessary to determine the crop N indicator, the determination of which may result in experimental bias due to equipment or operational practices; subsequently, in the phase of selecting characteristic bands with high correlation to the crop N indicator, there is the possibility of wrong band selection. In essence, the reliability of statistical models to assess crop N status depends on the dataset used to train the algorithm and the model. When applied to separate datasets under different conditions, the models are less generalizable. To achieve regional scalability and explore the influence of environmental differences on modeling, datasets can be constructed by combining spectral information and ecological factors so that they contain a large amount of variability data to improve model generalizability. However, improving the predictive accuracy of the model by adjusting the input parameters still has limits. Considering the principles, there is a trade-off between model interpretation and model performance. The predictive principles of traditional statistical models are intuitive and easy to understand, but at the expense of model performance; some machine learning models produce better predictive accuracy, while they are considered black box models because explaining how these models make decisions is a very difficult task (partial model prediction accuracy show in Table 2).

Physical models are highly advantageous in achieving model generalization, as they simulate the interaction between physical–chemical parameters and light from the physiological mechanisms of crops, thus providing explanations for the complex relationships between spectra and physical–chemical parameters at different fertility periods and under different growing conditions. However, the tedious and time-consuming inversion process limits its application. Existing studies usually analyze crop status at a small number of growth stages, so the statistical model has limited transferability across crops with different phenological status. The physical model overcomes this limitation and allows a wider range of crop canopy properties to be simulated. A hybrid model combining the mechanisms of statistical and physical models is not only efficient and flexible, but also explanatory for parameter inversion. In a hybrid model, the physical model is used to generate simulated spectra, which describes intra-canopy radiative transfer and interactions according to physics laws, thus providing information on spectral reflectance in relation to crop physicochemical variables [28]. Using simulated spectral data as input to train statistical models can provide physical constraints and explanations, and give a wider range of suitability [185]. Hybrid models have been applied successfully to estimate crop physical–chemical parameters (LAI, LCC, FVC, etc.) [185,186,187], but have rarely been used to invert N status. In recent years, physical models have moved from the previous indirect inversion of N through the physiological relationship between Chl and N to explore direct modeling of crop N status from spectral information, with more stable models and high response efficiency. Berger et al. [31] and Verrelst et al. [158] combined GP and PROSAIL-PRO models for inversion of crop N content, and confirmed the efficiency of hybrid models for direct N estimation. Therefore, how to achieve the complementary advantages between statistical and physical models, then construct a crop N estimation model with both mechanics and accuracy, will be the focus of future research.

6. Conclusions

Remote sensing technology is developing at a rapid pace and non-destructive monitoring and assessment of crop N status is gaining importance. This paper analyses the physiological mechanisms and spectral response characteristics of remote sensing monitoring for canopy N. Taking the remote sensing monitoring platform, the correlation between remotely sensed data and N status, and the remote sensing retrieval methods as the entry point, this paper provides an in-depth summary of the research techniques in the field of remote sensing monitoring for canopy N. The factors affecting the accuracy of remote sensing monitoring are also discussed. To date, the research at field scale has been well validated. The development of sensors and spectral carrying platforms facilitates high-precision remote sensing monitoring of crop N at farm scale. Due to the amount of information that can be extracted from remote sensing data, the efficiency of model use has become a key research concern. The efficiency and flexibility of machine learning models and the explanatory nature of physical models have their own advantages. The hybrid of the two models is beginning to show results in improving model stability. In addition, the effective use of multi-source data, and the removal of confounding factors in crop N monitoring need to be further explored. In-depth understanding of the limitations of current technology will be necessary to enhance the understanding of the link between canopy optical properties and crop N status, and to identify more appropriate N retrieval methods. In the context of the current rapid development of smart agriculture, the combination of sensors, remote sensing platforms and the Internet of Things results in the initial formation of a crop growth monitoring IoT platform. It provides the development direction for real-time monitoring and early forecasting of crop N, making it more widely application in the fields of growth monitoring, yield prediction and precision fertilization.