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. [68–71]. 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 [72–74]. 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. [77–81]. 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 [78–81], 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.

**Figure 3.** Remote sensing platforms, including ground-based platform, UAV-based platform, and satellite-based platform.

**Figure 4.** Spectral reflectance properties of crop canopy and spectral range of different platform sensors.

### *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" (Narea, per unit area); another is mass-based measure "nitrogen concentration" (N%, per unit dry matter) [28]. N% can be converted to Narea 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 (R<sup>2</sup> = 0.86 between N-sensitive bands and N status), the non-negative matrix factorization (R<sup>2</sup> = 0.83), the successive projections algorithm (R<sup>2</sup> = 0.66), and the competitive adaptive reweighted sampling (R2 = 0.93), etc. are also gradually applied [62,94,96–99]. 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 singlepeak 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.
