**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 sensingbased 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. 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 R<sup>2</sup> 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.

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



#### **Table 2.** *Cont.*

Table 2 covers case studies from different regions.

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, R<sup>2</sup> 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.) [185–187], 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.
