Retrieval of Vegetation Nitrogen from Hyperspectral Remote Sensing: A Critical Review of Recent Methodological Advances

Highlights

What are the main findings?

• Hyperspectral nitrogen retrieval has evolved from empirical spectral indices and linear regression towards nonlinear machine learning, physically based radiative transfer modeling (RTM), and hybrid RTM–ML retrieval frameworks that provide complementary advances in predictive flexibility, physical consistency, and the potential for robust and transferable retrieval.
• Empirical, machine learning, and physically based approaches are progressively converging into complementary hybrid retrieval frameworks rather than representing competing methodological paradigms.

What are the implications of the main findings?

• Future progress will depend less on incremental algorithmic improvements than on advances in protein-sensitive leaf and canopy RTMs, uncertainty propagation, standardized reference datasets, and hybrid RTM–ML retrieval frameworks.
• The growing availability of operational imaging spectroscopy missions will accelerate robust, transferable, and uncertainty-aware retrieval of nitrogen-related vegetation variables across species, ecosystems, and spatial scales.

Abstract

Hyperspectral retrieval of nitrogen-related vegetation variables has undergone rapid methodological advances driven by the emergence of protein-sensitive radiative transfer models (RTMs), modern machine learning (ML), and operational imaging spectroscopy. This review synthesizes recent developments in hyperspectral retrieval of nitrogen-related vegetation variables across leaf and canopy scales, with particular emphasis on advances reported between 2020 and 2026. We examine the evolution from classical parametric regression and nonlinear ML approaches towards physically based RTM inversion and hybrid RTM–ML frameworks that integrate the complementary strengths of physical modeling and statistical learning. Particular attention is given to protein-sensitive RTMs, advanced ML approaches, and uncertainty-aware retrieval. Recent developments highlight the potential of hybrid RTM–ML frameworks to combine physical consistency with computationally efficient statistical learning, while probabilistic methods such as Gaussian Process Regression provide additional capabilities for uncertainty characterization. The review further discusses the transition from experimental studies to operational applications enabled by airborne and satellite imaging spectroscopy, including PRISMA, EnMAP, and forthcoming missions such as CHIME. Remaining challenges include the inherently ill-posed nature of nitrogen retrieval, limited and insufficiently representative calibration data, uncertainty characterization, and generalization across sensors, species, and ecosystems. Overall, the reviewed evidence points towards increasingly integrated retrieval frameworks, while emphasizing that robust transferability and operational implementation remain dependent on representative data, physical realism, and rigorous uncertainty assessment.

1. Introduction

Nitrogen is one of the most important macronutrients regulating plant metabolism because it constitutes a major component of amino acids, proteins, nucleic acids, pigments, and other nitrogen-containing compounds [1,2,3]. Consequently, vegetation nitrogen strongly influences photosynthetic capacity, carbon assimilation, plant growth, crop productivity, and ecosystem functioning [1,4]. From a remote sensing perspective, vegetation nitrogen can be expressed through several related variables operating at different organizational levels [5]. At the leaf scale, studies commonly retrieve leaf nitrogen concentration (LNC), typically expressed as % dry matter or g N kg⁻¹ dry matter, and leaf nitrogen content (LNA; g N m⁻² leaf). At the canopy scale, retrievals generally target canopy nitrogen content (CNC; g N m⁻² ground or kg N ha⁻¹), which represents the amount of nitrogen contained within the aboveground canopy. Throughout this review, the term nitrogen retrieval collectively refers to these nitrogen-related variables unless stated otherwise.

Unlike pigments or water, nitrogen does not exhibit strong direct absorption features within the optical domain. Instead, a substantial fraction of foliar nitrogen is incorporated into proteins, whose relatively weak absorption features are often masked by stronger water and dry matter absorptions [6,7]. Leaf nitrogen is distributed among multiple photosynthetic and non-photosynthetic compounds and protein classes rather than being dominated by a single protein pool [1,8]. From an optical retrieval perspective, however, protein is particularly relevant because protein-sensitive radiative transfer models (RTMs) such as PROSPECT-PRO explicitly represent proteins as optically active biochemical constituents, thereby providing a physical pathway for estimating nitrogen-related information from leaf spectra [7]. Nitrogen retrieval is further complicated by covariation between leaf nitrogen and chlorophyll and by canopy structural effects, which can confound the attribution of spectral responses specifically to nitrogen [5].

An important distinction is therefore required between the nitrogen variable being retrieved and the spectral information pathway through which the retrieval is achieved. Reference variables such as leaf nitrogen concentration, leaf nitrogen content, and canopy nitrogen content generally represent total nitrogen, whereas hyperspectral observations may contain information related to these targets through several conceptually distinct pathways [5]. Protein absorption provides a relatively direct nitrogen-sensitive pathway because a substantial fraction of foliar nitrogen is incorporated into proteins, with protein-related absorption features providing a physical basis for spectroscopic nitrogen estimation [7,9,10,11]. Chlorophyll can provide a physiological proxy pathway because chlorophyll and nitrogen frequently covary through their association with photosynthetic functioning, although this relationship is not necessarily stable across species, developmental stages, or environmental conditions [1,5]. At canopy scale, leaf area index (LAI), biomass, and canopy architecture can provide additional indirect information related to canopy nitrogen content without being nitrogen-specific [5]. Consequently, the same spectral or structural variable may play different roles depending on the retrieval framework: chlorophyll-related information, for example, may improve the empirical prediction of total nitrogen while acting as a confounding contribution when physically isolating protein-sensitive absorption [7]. These information pathways should therefore be distinguished from the biochemical nitrogen variable ultimately being predicted. Figure 1 summarizes these conceptually distinct pathways linking hyperspectral observations to nitrogen-related retrieval targets.

Consequently, quantitative nitrogen retrieval represents one of the more challenging applications of optical remote sensing. The advent of imaging spectroscopy, providing hundreds of contiguous narrow spectral bands, has nevertheless enabled the development of increasingly sophisticated retrieval methods by exploiting subtle spectral signatures associated with nitrogen-related vegetation variables. Hyperspectral observations acquired from proximal sensors, unmanned aerial vehicles (UAVs), airborne campaigns, and satellite missions have therefore substantially expanded opportunities for quantitative nitrogen estimation across spatial scales (see reviews, refs. [5,12]). Early satellite imaging spectroscopy missions, such as Hyperion [13] and CHRIS-PROBA [14], demonstrated the scientific potential of hyperspectral observations, whereas more recent missions including PRISMA [15], DESIS, EnMAP [16], and forthcoming missions such as CHIME [17] and NASA’s SBG VSWIR component, now developed as EAGLE-VSWIR [18], are ushering imaging spectroscopy into an operational era. These developments provide unprecedented opportunities for quantitative nitrogen monitoring across spatial scales while simultaneously increasing the demand for retrieval methods that are accurate, physically consistent, transferable, computationally efficient, and capable of supporting operational applications.

Retrieving nitrogen from hyperspectral observations requires methods capable of translating complex spectral signatures into meaningful biochemical variables [5]. In essence, nitrogen retrieval requires a retrieval model that establishes a quantitative relationship between spectral observations and nitrogen-related vegetation variables. A wide variety of retrieval approaches have been developed, ranging from empirical statistical relationships to physically based RTM inversion and hybrid retrieval frameworks that integrate RTMs with data-driven learning. Previous reviews have addressed retrieval methods for vegetation traits in general (see reviews, refs. [19,20,21]), whereas nitrogen-specific reviews have focused on the state of hyperspectral nitrogen retrieval up to 2020 [5] or, more recently, on UAV-based crop nitrogen monitoring applications [12]. Here, we specifically review the recent methodological developments underpinning nitrogen retrieval from hyperspectral observations across proximal, UAV, airborne, and satellite platforms, with particular emphasis on advances in machine learning (ML), physically based RTMs, and hybrid RTM–ML retrieval frameworks.

This paper presents a critical narrative review rather than a systematic review or meta-analysis. Accordingly, the literature is used to trace, illustrate, and critically assess methodological developments, with emphasis on representative recent studies and comparative evidence rather than exhaustive enumeration of all published nitrogen retrieval applications. Comparisons among retrieval approaches are therefore primarily qualitative and method-oriented, focusing on characteristics such as predictive flexibility, dimensionality handling, computational efficiency, interpretability, uncertainty characterization, physical consistency, data requirements, and transferability rather than direct comparison of performance metrics across heterogeneous studies. Particular attention is given to developments reported between 2020 and 2026, while earlier seminal studies are retained where needed to establish the methodological context and evolution of the field. Traditionally, retrieval methods have been categorized into either: (1) statistical (or variable-driven) methods or (2) physically based (or radiometrically driven) methods [22]. However, these two fundamental categories have evolved considerably over the last few decades, and elements of both are increasingly combined, giving rise to hybrid approaches. Following the taxonomy outlined in previous reviews [19,20,21], retrieval methods can be organized into the following four categories:

• Parametric regression methodsestablish an explicit mathematical relationship between spectral observations and nitrogen-related vegetation variables. They include (i) user-defined models, such as vegetation indices derived from known spectral relationships, and (ii) statistical regression models whose parameters are estimated from calibration data within a predefined mathematical structure.
• Nonparametric regression methods do not assume a predefined functional form of the relationship between spectra and nitrogen variables. Instead, the relationship is learned directly from the data using nonlinear ML algorithms.
• RTM inversion methods, founded on radiative transfer theory, establish causal relationships between photon interactions and model parameters. These methods are grounded in mathematical representations of light propagation embodied in radiative transfer models (RTMs).
• Hybrid regression methods combine physically based radiative transfer models with statistical or ML approaches, aiming to exploit the physical consistency of RTMs together with the flexibility and computational efficiency of data-driven algorithms.

These categories provide a conceptual framework for organizing the diversity of retrieval methods encountered in contemporary nitrogen estimation studies. Parametric, nonparametric, and hybrid approaches retrieve nitrogen through statistical relationships between spectral observations and target variables, whereas physically based approaches retrieve nitrogen by inverting RTMs. Nevertheless, the boundaries between these categories are not always clear-cut, and methodological overlap frequently occurs. For example, user-defined spectral indices are often used as predictors within ML algorithms (e.g., ref. [23]). Following the taxonomy proposed by Verrelst et al. [21], the retrieval remains nonparametric because the regression model itself learns the input–output relationship directly from the data, regardless of whether the predictors are raw spectra or spectral transformations. It is also important to distinguish between the retrieval methodology and the data used for model development. Parametric and nonparametric approaches are generally calibrated using empirical observations, whereas physically based and hybrid approaches rely on synthetic datasets generated by RTMs, although empirical measurements remain essential for calibration, validation, and performance assessment. Consequently, these four categories provide a practical framework for organizing retrieval methodologies rather than representing mutually exclusive classes.

Because nitrogen-related spectral signatures are weak and frequently confounded by other biochemical and structural properties, hyperspectral nitrogen retrieval constitutes a challenging inverse problem in which multiple combinations of biochemical and structural properties can produce similar spectral responses, making the retrieval inherently ill-posed [5,21]. As a result, achieving accurate, robust, and transferable nitrogen retrieval across different species, sensors, and environmental conditions remains difficult. Nevertheless, recent advances in ML, protein-sensitive RTMs, and hybrid RTM–ML retrieval frameworks have substantially expanded the possibilities for accurate, physically consistent, and transferable nitrogen retrieval. Within this narrative-review framework, this paper provides a critical synthesis of the major methodological developments in hyperspectral nitrogen retrieval, tracing the evolution from classical parametric regression methods to modern ML algorithms, physically based RTM inversion, and hybrid RTM–ML retrieval frameworks. Particular emphasis is placed on advances reported between 2020 and 2026 (post [5]), reflecting the emergence of protein-sensitive leaf and canopy RTMs, uncertainty-aware retrieval methods, and the growing availability of operational imaging spectroscopy from UAV, airborne, and satellite platforms. Finally, the review discusses the remaining methodological challenges and future research directions, including uncertainty quantification, standardized reference datasets, hybrid retrieval frameworks, and the opportunities offered by current and forthcoming imaging spectroscopy missions.

2. Parametric Regression Methods

2.1. Discrete Spectral Band Approaches: Vegetation Indices

Vegetation indices (VIs) represent one of the earliest and most widely adopted approaches for estimating nitrogen-related variables from optical spectroscopy. By combining reflectance measurements from selected wavelengths, they emphasize spectral information related to nitrogen while reducing confounding influences from canopy structure, soil background, and illumination [5]. Owing to their simplicity, computational efficiency, and straightforward interpretation, VIs have remained popular across proximal, UAV, airborne, and satellite hyperspectral observations. Because VIs provide spectral predictors rather than direct nitrogen estimates, they require an empirical calibration step, most commonly based on linear regression (see reviews, refs. [24,25]).

The transition from broadband to hyperspectral observations substantially expanded VI development by enabling the systematic optimization of narrowband combinations, typically based on simple ratios or normalized differences. Exhaustive evaluation of two- and three-band combinations has been widely used to identify informative spectral regions, and optimized narrowband indices often outperform conventional broadband formulations for nitrogen-related variables (see reviews, refs. [5,26]). For example, Guo et al. [27] demonstrated improved leaf nitrogen estimation using optimized narrowband combinations. Recent studies have further optimized crop-specific indices, central wavelengths, bandwidths, and automated band selection for canopy nitrogen retrieval (e.g., refs. [23,28,29,30]). Beyond agricultural applications, Farella et al. [31] evaluated vegetation indices together with full-spectrum imaging spectroscopy for foliar nitrogen across multiple vegetation communities and biomes, demonstrating that these empirical retrieval concepts also extend to more diverse vegetation types. Nevertheless, optimized VIs are increasingly complemented by ML and hybrid RTM–ML approaches when stronger transferability across crops, sensors, and environmental conditions is required (see review, refs. [12]).

VI-based approaches nevertheless remain limited by the reduction of hyperspectral observations to only a few selected bands and by the application-specific nature of optimized formulations. Their performance depends jointly on band selection and bandwidth, index formulation, and regression function [19,20,21], while optimal configurations often vary among species, growth stages, sensors, and environmental conditions (e.g., refs. [28,32,33]). At the canopy scale, structural attributes such as leaf area index (LAI), biomass, and canopy architecture further confound the subtle nitrogen signal [34]. Consequently, although VIs remain attractive for their simplicity and interpretability, their limited use of the available spectral information has motivated the development of full-spectrum regression approaches.

2.2. Spectral Transformation Techniques

Because VIs use only selected wavelength combinations, spectral transformation techniques have been developed to exploit contiguous hyperspectral observations more fully. These methods generate alternative spectral representations that enhance subtle nitrogen-related information while reducing noise, redundancy, and background effects, and are now widely used as preprocessing techniques for nitrogen retrieval [5]. Figure 2 summarizes the principal approaches considered here: optimized band combinations, derivative spectroscopy, and wavelet decomposition.

Derivative and wavelet-based representations have been progressively refined to improve sensitivity to weak nitrogen-related spectral information (see reviews, refs. [5,26]). Recent developments include crop-specific nitrogen-sensitive VIs, automated band selection, optimized wavelength combinations, and derivative-based transformations embedded within retrieval workflows [28,29,30,32,33,35,36]. Derivative spectroscopy enhances weak absorption features while reducing baseline variability and background effects. Soltanikazemi et al. [35], for example, reported improved wheat nitrogen estimation using derivative preprocessing, while Xu et al. [36] combined derivative spectra with optimized feature selection across multiple growth stages. Wavelet decomposition provides a complementary multiscale representation that suppresses noise while preserving localized spectral information; Li et al. [37] demonstrated its value for wheat nitrogen estimation.

Overall, spectral transformations are best regarded as preprocessing steps rather than retrieval methods in their own right. Contemporary studies increasingly combine them with full-spectrum ML and hybrid RTM–ML frameworks (e.g., refs. [36,38,39]), reflecting a broader transition from manually engineered spectral descriptors towards data-driven exploitation of hyperspectral observations.

2.3. Parametric Regression Methods with Implicit Parameterization

Whereas the preceding approaches rely on explicitly defined spectral transformations, the methods discussed here are parameterized through model optimization. Following Verrelst et al. [21], they are classified as parametric regression methods because they assume a predefined linear relationship between predictors and response variables described by a finite set of model parameters. Their predictive capability is achieved through latent-variable decomposition while retaining linear regression as the underlying framework.

Hyperspectral datasets typically comprise hundreds of highly correlated spectral bands, whereas calibration datasets often contain only a limited number of field observations. Under these conditions, conventional multiple linear regression becomes unstable because of severe multicollinearity and overfitting [40,41]. Consequently, latent-variable regression methods have become well-established parametric approaches and continue to serve as important benchmarks for hyperspectral nitrogen retrieval [5,12].

Principal Component Regression (PCR) first projects the original spectral variables onto a reduced set of orthogonal principal components derived from principal component analysis, after which linear regression is performed on the leading components [42]. Because these components are selected solely to explain spectral variance, PCR does not explicitly account for the response variable during dimensionality reduction. Partial Least Squares Regression (PLSR), by contrast, simultaneously projects predictor and response variables into a common latent space, extracting latent variables that maximize their covariance [41]. Consequently, the resulting components are optimized for prediction rather than merely representing the dominant spectral variance. Figure 3 illustrates the conceptual differences between both methods.

Although PCR has been applied to hyperspectral nitrogen retrieval, PLSR has achieved substantially wider adoption. Comparative studies generally report better predictive performance for PLSR because its latent variables are constructed to explain variation in the target variable while mitigating spectral redundancy. For example, Yin et al. [43] found that PLSR consistently outperformed PCR for cotton leaf nitrogen estimation, while Wang et al. [39] identified PLSR as one of the strongest linear baselines across multiple crops and growth stages.

PLSR has therefore become the dominant parametric regression method and a principal benchmark for hyperspectral nitrogen retrieval (see also, refs. [12,26]). Its adoption spans proximal, UAV, airborne, and satellite imaging spectroscopy across agricultural and natural vegetation applications (e.g., refs. [31,32,38,39,44]). Its popularity derives from simultaneously reducing spectral dimensionality and extracting latent variables that maximize covariance with the target, thereby addressing hyperspectral multicollinearity while retaining full-spectrum information.

Recent studies continue to use PLSR as a reference model when evaluating more advanced ML approaches (e.g., refs. [38,39,44,45,46,47]). Although nonlinear algorithms sometimes achieve higher predictive accuracy, PLSR often remains highly competitive while offering computational efficiency, numerical stability, and interpretability. It therefore continues to provide a transparent and reliable baseline against which more complex retrieval approaches can be evaluated.

3. Nonparametric Regression Methods

Unlike parametric approaches, nonparametric regression methods learn the relationship between spectral observations and target variables directly from the data without assuming a predefined functional form. Their flexibility enables complex nonlinear relationships to be represented, but also requires appropriate regularization to avoid overfitting and maintain generalization (see also, ref. [48]). A broad range of these algorithms, commonly referred to as machine learning (ML) methods, have been adopted for hyperspectral remote sensing. They are well suited to exploiting subtle information distributed across multiple wavelength regions, although feature selection or dimensionality reduction can improve computational efficiency and generalization when spectral dimensionality is high (see reviews, refs. [5,20]).

Nitrogen retrieval provides an excellent example where nonlinear learning becomes advantageous. Nitrogen-related information can be expressed in hyperspectral reflectance through multiple pathways, including relatively direct protein-sensitive absorption and indirect relationships with chlorophyll, leaf structure, biomass, and canopy architecture [5,7]. Nonlinear ML algorithms can exploit interactions among these different information sources distributed across multiple spectral regions, although the resulting statistical relationships are not necessarily nitrogen-specific. As a result, ML methods have become the dominant empirical framework for hyperspectral nitrogen retrieval, with numerous comparative studies reporting improved predictive performance over traditional parametric approaches (see review, ref. [12]). Recent comparative studies using UAV hyperspectral observations of maize [32], wheat [36,46], rice [49], perennial ryegrass [50], cotton [51], and multiple crop species [39] generally report improved performance of nonlinear ML over classical linear regression, although the magnitude of improvement depends on crop type, measurement scale, preprocessing, and training sample size.

This section focuses on ML algorithms trained exclusively on empirical observations; hybrid RTM–ML approaches are discussed separately in Section 5. We distinguish three principal families: (1) decision-tree and ensemble-learning methods, (2) kernel-based regression methods, and (3) neural networks (NNs). Figure 4 illustrates representative methods from each family.

3.1. Decision-Tree and Ensemble-Learning Methods

Decision trees represent an established nonlinear ML paradigm and have evolved into a broad family of ensemble-learning algorithms (

Table 1. Principal decision-tree and ensemble-learning methods used in hyperspectral nitrogen retrieval and their relevance to high-dimensional spectral data.

Method	Description	Ref.
Decision Trees (DTs)	DTs recursively partition the predictor space into decision rules and can represent nonlinear relationships and spectral interactions without requiring linearity assumptions. They are readily interpretable but can exhibit high variance when applied to high-dimensional hyperspectral predictors.	[52]
Random Forests (RFs)	RF combines trees trained on bootstrap samples with random predictor selection at each split, improving robustness and reducing overfitting. This makes RF well suited to high-dimensional and correlated hyperspectral predictors and enables assessment of spectral-variable importance.	[53]
Boosted Trees	Boosted-tree methods sequentially correct residual errors and can capture complex nonlinear relationships in hyperspectral data. GBDT, XGBoost, and LightGBM provide regularized and computationally efficient implementations, although performance remains dependent on representative training data and appropriate model tuning.	[54,55,56]

). Because individual trees can exhibit high variance and limited predictive performance on complex hyperspectral datasets, contemporary nitrogen retrieval predominantly relies on ensemble methods, particularly Random Forest (RF).

RF is widely used for hyperspectral nitrogen retrieval because it combines nonlinear modeling capability with comparatively limited tuning requirements. Studies using UAV hyperspectral imagery have reported strong performance for maize [32], wheat [36], and multiple crop species [39]. Boosted-tree methods have also gained attention. Yang et al. [57] reported that GBDT outperformed PLSR and SVR for wheat leaf nitrogen content, while Tian et al. [58] found GBDT among the best-performing methods for rice leaf nitrogen content from UAV hyperspectral imagery.

Recent work increasingly combines ensemble learning with spectral preprocessing and model interpretation. RF and boosted-tree methods have been coupled with derivative spectra, VIs, wavelet-derived features, and optimized spectral inputs [32,36,38,39], while more integrated pipelines combine preprocessing, feature selection, and stacking ensembles [59]. Tree-based methods also provide predictor-importance measures, and SHapley Additive exPlanations (SHAPs) [60] provide additional attribution of individual predictors to model outputs. Recent applications demonstrate their value for identifying influential wavelength regions and improving the interpretation of hyperspectral nitrogen retrieval [51,61,62].

3.2. Kernel-Based Regression Methods

Kernel-based regression methods represent complex nonlinear relationships through kernel functions that implicitly map observations into a high-dimensional feature space (

Table 2. Principal kernel-based regression methods used in hyperspectral retrieval and their relevance to high-dimensional spectral data.

Method	Description	Ref.
Support Vector Regression (SVR)	SVR uses kernel functions to model nonlinear relationships in high-dimensional hyperspectral data while controlling model complexity through regularization and an ϵ-insensitive loss. It is particularly suitable when calibration datasets are limited relative to spectral dimensionality.	[64]
Kernel Ridge Regression (KRR)	KRR combines kernel-based nonlinear regression with regularized least squares, enabling efficient modeling of complex relationships among highly correlated hyperspectral predictors while controlling overfitting through an L2 penalty.	[65]
Gaussian Process Regression (GPR)	GPR provides nonlinear probabilistic regression for high-dimensional hyperspectral data, yielding predictive estimates together with associated uncertainty. Kernel formulation and ARD can additionally support spectral-variable relevance analysis, making GPR attractive for uncertainty-aware retrieval and active learning.	[66]

) [63]. They are particularly attractive for hyperspectral trait retrieval under limited calibration data and high spectral dimensionality (see reviews, refs. [19,20,21]). Three principal kernel-based algorithms considered here are SVR, Kernel Ridge Regression (KRR), and Gaussian Process Regression (GPR), which differ in optimization strategy, computational complexity, and uncertainty characterization.

SVR has long served as a benchmark for hyperspectral nitrogen retrieval because of its strong generalization under limited calibration data [67]. Optimized implementations have achieved strong performance for winter wheat [68] and wheat leaf nitrogen concentration [69], highlighting the importance of preprocessing, feature selection, and hyperparameter optimization. KRR has received comparatively little attention in empirical nitrogen retrieval despite its theoretical suitability for high-dimensional hyperspectral data.

GPR occupies a distinctive position because it combines nonlinear regression with Bayesian inference, providing predictive means and associated uncertainty estimates [66]. This is particularly attractive for nitrogen retrieval, where weak spectral signatures, measurement uncertainty, and limited field observations often lead to ambiguous solutions. Although empirical nitrogen-specific evaluations of GPR remain comparatively limited, available evidence is encouraging. Wang et al. [39] found GPR among the best-performing methods across multiple crops and growth stages. Beyond predictive accuracy, GPR provides uncertainty estimates through posterior predictive variances [66,70], while Automatic Relevance Determination (ARD) can identify influential spectral variables through predictor-specific kernel length scales [71,72].

Overall, SVR remains an established benchmark, GPR has received increasing attention owing to its probabilistic formulation, and KRR has been explored more extensively within hybrid RTM–ML frameworks (Section 5). The uncertainty and relevance-analysis capabilities of GPR also provide a natural link to the hybrid approaches discussed later.

3.3. Neural Networks

Neural networks (NNs) have long been applied to hyperspectral remote sensing because of their ability to approximate highly nonlinear relationships. Architectures have evolved from conventional feed-forward networks towards deep-learning models capable of learning hierarchical spectral representations directly from the data.

Table 3. Representative neural network architectures used for hyperspectral regression and their relevance to spectral and spectral–spatial data.

Method	Description	Ref.
Neural Networks (NNs)	NNs use interconnected layers to learn nonlinear relationships between high-dimensional hyperspectral predictors and target variables. Their flexibility allows complex spectral responses to be represented, although performance depends strongly on training-data quantity and representativeness.	[73]
Convolutional Neural Networks (CNNs)	CNNs use convolutional filters and parameter sharing to learn local spectral patterns and, for image data, spectral–spatial features directly from hyperspectral observations, reducing reliance on manually engineered features.	[74]
Transformer Networks	Transformers use self-attention to capture long-range dependencies among spectral bands and, in spectral–spatial implementations, across image features, making them promising for modeling complex hyperspectral relationships.	[75]

summarizes the principal architectures considered here.

NNs have long been explored for hyperspectral nitrogen retrieval because of their ability to model nonlinear interactions distributed across multiple wavelength regions, and often outperform classical parametric approaches such as PLSR (see review [12]; see also, e.g., refs. [39,44,57]). When calibration datasets are limited, NNs are often trained using engineered spectral features such as VIs, derivatives, wavelets, continuum removal, or principal components (e.g., refs. [38,39,57]). With larger datasets, deep-learning architectures can increasingly learn useful representations directly from raw hyperspectral observations [38,76]. Conventional feed-forward NNs therefore remain useful baselines but are less central than modern deep architectures.

Among deep-learning approaches, convolutional neural networks (CNNs) are widely used because they automatically learn hierarchical spectral representations. Applications include canopy and leaf nitrogen retrieval in field crops [57,77], greenhouse tomato [47], and, more recently, joint structural and biochemical retrieval from UAV hyperspectral imagery over individual trees [78]. Recent developments extend CNNs beyond fully supervised learning. Gallo et al. [76] proposed a self-supervised framework for estimating chlorophyll and nitrogen content in maize, while Yin et al. [79] showed that active transfer learning can improve potato nitrogen estimation under limited labeled data. These approaches aim to reduce dependence on extensive field observations and improve transferability across sensors, environments, and vegetation types. Despite these advances, there is no consistent evidence that increasingly complex NN architectures universally outperform established ML algorithms. Comparative studies indicate that performance differences are often modest and strongly dataset dependent [39,44,46].

Overall, the main advantage of deep learning lies in its ability to learn informative spectral representations directly from hyperspectral observations. Current developments therefore increasingly emphasize self-supervised and transfer-learning strategies, while predictive performance remains strongly dependent on the quality, quantity, and representativeness of the training data.

4. Radiative Transfer Model Inversion

Unlike empirical regression methods, physically based inversion estimates vegetation properties by explicitly modeling the interaction between radiation and vegetation canopies. Retrieval is achieved by matching observed spectra with RTM simulations rather than learning purely statistical relationships from calibration data, thereby providing greater physical consistency and potential transferability across sensors, acquisition conditions, and ecosystems [80]. RTMs describe the propagation, absorption, transmission, and scattering of radiation within vegetation canopies and range from computationally efficient one-dimensional models to detailed three-dimensional representations of canopy architecture and multiple scattering [80,81,82,83]. Increasing model complexity generally improves physical realism but also increases computational demands and complicates inversion.

For hyperspectral nitrogen retrieval, the most relevant RTMs are leaf and canopy models capable of representing protein-sensitive optical properties. Classical versions of PROSPECT [84,85] successfully simulate leaf reflectance and transmittance using biochemical constituents such as chlorophyll, carotenoids, water, dry matter, and brown pigments, but they do not explicitly represent proteins [7]. This limitation considerably restricted their applicability for physically based nitrogen retrieval because protein-associated absorption provides one of the more direct optical pathways through which nitrogen-related information can be represented in leaf reflectance [7,71].

The introduction of PROSPECT-PRO [7], which explicitly represents protein absorption, substantially strengthened this physical pathway. Coupling PROSPECT-PRO with canopy models such as SAIL yields PROSAIL-PRO, enabling physically based simulation of leaf and canopy reflectance as a function of protein content [7,71]. Importantly, PROSPECT-PRO parameterizes protein rather than total nitrogen; conversion therefore requires an additional protein-to-nitrogen relationship whose stability can vary among species, developmental stages, nutritional conditions, and environmental stresses [1,8,86]. Protein-sensitive RTMs should therefore be understood as strengthening the physical basis of nitrogen retrieval rather than providing a direct observation of total nitrogen. Their introduction has nevertheless renewed interest in RTM-based inversion, with applications across crops and sensing platforms (e.g., refs. [71,86,87,88,89]).

From a computational perspective, the PROSAIL family remains the principal operational framework because it provides a favorable compromise between physical realism and inversion efficiency [80,90]. More sophisticated three-dimensional RTMs, including Raytran [91], FLIGHT [92], librat [93], DART [94], and LESS [82], represent canopy architecture and radiative interactions more explicitly. Coupled with PROSPECT-PRO, they can generate physically consistent protein-sensitive canopy simulations (e.g., ref. [83]), but their computational demands currently limit routine inversion.

Physically based nitrogen retrieval remains more challenging than retrieval of structurally dominant traits such as LAI because protein-sensitive absorption features are weak and partly masked by water, dry matter, chlorophyll, and canopy structure [7,71]. Nitrogen-related information can therefore arise through distinct pathways: relatively direct protein-sensitive absorption, physiological proxy relationships with chlorophyll, and indirect structural or biomass-related information at canopy scale. These pathways should not be interpreted as equivalent evidence of nitrogen absorption. Reflectance-based nitrogen inversion is also intrinsically ill posed because multiple biochemical and structural parameter combinations can produce similar spectra. Accordingly, the principal challenge in RTM inversion is not predictor multicollinearity but parameter equifinality and the identification of sufficiently constrained solutions [80,95]. The protein-to-total-nitrogen conversion introduces an additional uncertainty beyond radiative transfer itself, while interactions among biochemical and structural variables further contribute to retrieval ambiguity [7,8,71,86,87]. The principal inversion strategies considered here are numerical optimization and lookup-table (LUT) inversion, while hybrid RTM–ML approaches are discussed separately in Section 5. Their underlying concepts are illustrated in Figure 5.

RTM Inversion Strategies

Numerical optimization iteratively adjusts RTM parameters until simulated spectra match the observations according to a predefined objective function, whereas LUT inversion identifies suitable solutions within a precomputed simulation database. LUT approaches substantially reduce computational costs and have therefore become the principal physically based retrieval strategy in optical remote sensing [19,80].

For hyperspectral nitrogen retrieval, recent progress has been driven primarily by advances in the underlying RTMs rather than by inversion algorithms themselves. PROSPECT-PRO and PROSAIL-PRO provide a more direct protein-sensitive pathway between spectral reflectance and a major nitrogen-containing biochemical pool, although conversion from protein to total nitrogen remains subject to the allocation uncertainty described above [7,8,86]. Experimental studies have demonstrated the potential of these frameworks across a range of crops and sensing platforms [71,87,88].

Numerical optimization remains valuable for methodological studies because it supports sensitivity analysis, Bayesian inversion, and uncertainty characterization. Bayesian approaches estimate posterior parameter distributions rather than single optimal solutions, but repeated RTM evaluations can make them computationally demanding for operational nitrogen retrieval.

LUT inversion reduces this computational burden through spectral matching against precomputed simulations. However, weak protein-sensitive signals and interactions with chlorophyll, dry matter, and canopy structure make LUT retrieval susceptible to equifinality. Recent developments therefore emphasize protein-sensitive RTMs, parameter constraints, optimized cost functions, and averaging multiple acceptable solutions rather than relying on a single best match [49,96,97,98]. Incorporating physiological nitrogen allocation within PROSAIL-PRO provides an additional route for improving canopy nitrogen realism [86].

Overall, RTM inversion provides a physically consistent framework for nitrogen retrieval and a controlled environment for sensitivity analysis and generation of synthetic training data. Its practical performance nevertheless remains constrained by computational cost, parameter equifinality, model realism, and the representativeness of simulated conditions. These limitations motivate the hybrid RTM–ML frameworks discussed in the following section.

5. Hybrid RTM–ML Retrieval Frameworks

Hybrid retrieval has become an increasingly important framework for hyperspectral nitrogen estimation by combining the physical consistency of RTMs with the predictive capability of ML algorithms. Following early implementations based on feed-forward NNs, hybrid approaches have expanded to ensemble methods, kernel algorithms, and deep-learning architectures [19,20,86].

5.1. NN-Based Hybrid Retrieval

Hybrid retrieval was initially developed using feed-forward NNs trained on RTM simulations to avoid the computational expense of direct RTM inversion [19]. The introduction of protein-sensitive models such as PROSPECT-PRO and PROSAIL-PRO subsequently enabled physically consistent retrieval of nitrogen-related variables through a more direct protein-sensitive pathway (e.g., refs. [7,71,86]). Recent hybrid nitrogen studies increasingly employ deep-learning architectures trained on synthetic RTM simulations. CNNs are particularly attractive because they learn hierarchical spectral representations while reducing the need for manual preprocessing. Gallo et al. [76], for example, proposed a self-supervised convolutional framework for maize chlorophyll and nitrogen retrieval, while Bhadra et al. [99] evaluated CNN, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) architectures using PROSAIL simulations and multi-angular UAV hyperspectral observations. More recent frameworks integrate physical knowledge, biological constraints, and transfer learning more explicitly. Dehghan-Shoar et al. [100] proposed a physically informed multi-scale deep network for foliar nitrogen estimation, while Yang et al. [101] combined PROSPECT-PRO simulations, a nitrogen allocation model, wavelet decomposition, spectral similarity selection, and Conv-Transformer learning. Yang et al. [102] further integrated RTM simulations with deep transfer learning for winter wheat nitrogen density estimation, improving generalization while reducing calibration requirements. Overall, progress in hybrid NN retrieval increasingly depends less on network complexity itself than on the realism of RTM simulations, integration of physical and biological knowledge, and transferability of learned representations.

5.2. Kernel-Based Hybrid Retrieval

Kernel-based algorithms have become widely used for hybrid RTM inversion, with GPR receiving particular attention because it combines nonlinear prediction, predictive uncertainty, and feature-relevance analysis through ARD [71]. Early studies demonstrated the feasibility of coupling PROSAIL simulations with GPR for vegetation trait retrieval [103,104], while protein-sensitive RTMs subsequently strengthened the physical basis for nitrogen-specific applications. Using EnMAP simulations, Berger et al. [71] demonstrated hybrid GPR retrieval of aboveground crop nitrogen content while identifying informative spectral regions through ARD. Ranghetti et al. [87] subsequently reported robust GPR performance for canopy chlorophyll and nitrogen retrieval from PRISMA imagery, while Peanusaha et al. [88] demonstrated hybrid GPR retrieval of leaf nitrogen using PROSPECT-PRO simulations.

An additional advantage of kernel-based hybrid retrieval is the possibility of optimizing the synthetic training dataset itself. Active Learning (AL) iteratively selects informative RTM simulations, reducing training-set size while maintaining predictive performance [105,106]. Combined with PCA, AL enabled regional canopy nitrogen mapping from PRISMA [107], and related strategies have been applied to maize nitrogen estimation [108], operational canopy nitrogen monitoring [109], and multi-trait retrieval from EnMAP under limited field sampling [110].

Overall, hybrid GPR combines physically based simulation with probabilistic regression and intrinsic uncertainty estimation [70]. Together with protein-sensitive RTMs, AL, and improved RTM parameterization, this makes GPR attractive for uncertainty-aware hybrid retrieval, although nitrogen-specific comparative evidence remains comparatively limited.

5.3. Emerging Hybrid RTM–ML Retrieval Frameworks

Hybrid retrieval is increasingly extending beyond the conventional paradigm of training ML exclusively on RTM-generated simulations. Emerging frameworks combine physically based models, empirical observations, complementary environmental information, and advanced learning strategies to improve robustness and transferability. One form of hybridization integrates physically derived features with empirical learning. In heterogeneous grassland, Dehghan-Shoar et al. [111] combined PROSAIL-derived features with empirical spectral information and GPR for nitrogen prediction, illustrating an alternative route for integrating radiative-transfer information with probabilistic regression.

Another direction is the integration of complementary observations such as thermal infrared data, meteorological variables, multi-angular measurements, and ancillary environmental information. These can help constrain confounding effects associated with water stress, canopy structure, and environmental variability. For example, Chen et al. [46] showed improved winter wheat nitrogen estimation using meteorological information, while Fazli et al. [112] demonstrated the value of combining hyperspectral and thermal observations for separating nitrogen and water stress.

Physics-assisted empirical learning provides a further pathway by combining field observations with physically based simulations during model development. Li et al. [89] combined empirical measurements with PROSAIL-PRO simulations for maize nitrogen estimation, while Li et al. [86] integrated PROSAIL-PRO with a physiological nitrogen allocation model to improve the representation of canopy nitrogen dynamics. Overall, current developments indicate a gradual convergence of physical modeling, empirical observations, complementary data, and modern ML. Although many of these approaches remain at an early stage, they represent promising directions for more robust and transferable operational nitrogen retrieval.

6. Discussion

The rapid development of imaging spectroscopy across proximal, UAV, airborne, and satellite platforms has substantially expanded opportunities for quantitative nitrogen retrieval. Building upon earlier reviews [5,19,20,21], this review has focused on recent methodological developments in hyperspectral nitrogen retrieval. Rather than identifying a universally superior algorithm, the reviewed literature indicates that performance remains conditional on the target variable, calibration data, preprocessing, sensing scale, vegetation type, and validation strategy [5,12,39]. Reported accuracies across individual studies should therefore not be interpreted as directly comparable benchmarks.

An important limitation of the current evidence base is its strong emphasis on agricultural vegetation. Although recent studies demonstrate applications across grasslands, forests, and broader vegetation types and biomes [31,111,113], much post-2020 methodological development has focused on crops, particularly for UAV-based ML and hybrid RTM–ML retrieval. Transferability to structurally complex and taxonomically diverse natural and semi-natural vegetation therefore remains less extensively evaluated and represents an important research priority.

A major source of performance variation is the quality and representativeness of calibration and validation data. Many studies rely on relatively limited datasets collected within individual experiments, seasons, or sites, while genuinely independent validation across years, sites, sensors, or vegetation types remains less frequent (see review, ref. [12]). Increasingly complex models do not consistently outperform simpler approaches under different data conditions [39,44]. PLSR therefore remains a useful transparent baseline, whereas nonlinear ML approaches become particularly attractive when sufficiently representative observations are available. Method selection should consequently reflect the retrieval problem rather than algorithmic complexity alone. VIs and PLSR remain useful when datasets are limited or transparency and computational efficiency are priorities [5,12]. Ensemble-learning and kernel-based methods offer greater nonlinear flexibility with moderate sample sizes, while deep-learning architectures generally require larger and more diverse datasets to realize their representation-learning advantage [47,76]. Current comparative evidence does not show that deep learning universally outperforms established ML approaches [39,44].

Transferability represents a second major limitation. Empirical relationships between spectra and nitrogen may exploit both nitrogen-sensitive information and correlated proxy variables, and these relationships can change across species, developmental stages, management conditions, sensors, and environments [5]. Nitrogen-related information may arise through relatively direct protein-sensitive absorption [7], physiological relationships with chlorophyll, and indirect structural or biomass-related information at canopy scale. Chlorophyll should therefore not be regarded simply as an interference term: it can provide an effective physiological proxy for total nitrogen in empirical retrieval, while acting as a confounding contribution when physically isolating protein-sensitive absorption [1,5,7]. These pathways differ in their stability across domains, which helps explain why empirical models may lose skill when transferred beyond their calibration conditions.

A related conceptual limitation concerns an overly Rubisco-centered interpretation of vegetation nitrogen. Although Rubisco represents an important photosynthetic nitrogen pool, foliar nitrogen is distributed among multiple photosynthetic and non-photosynthetic proteins and other nitrogen-containing compounds [1,8]. Neither Rubisco nor protein alone should therefore be interpreted as equivalent to total leaf nitrogen. The relevance of protein for hyperspectral retrieval instead arises because protein-sensitive absorption provides a relatively direct optical pathway that can be represented by current RTMs.

Physically based and hybrid retrievals introduce their own limitations. Protein-sensitive RTMs such as PROSPECT-PRO and PROSAIL-PRO strengthen the physical representation of nitrogen-related optical properties [7,71,86], but inversion remains affected by weak protein absorption, parameter equifinality, and canopy parameterization uncertainty. Because PROSPECT-PRO represents protein rather than total nitrogen, conversion additionally requires a protein-to-nitrogen relationship whose stability may vary across species, developmental stages, nutritional conditions, and environmental stresses [7,8,86]. Hybrid RTM–ML approaches further depend on how well the simulated training domain represents real observations [87,107]. Incomplete RTM parameterization, instrumental effects, canopy heterogeneity, unrealistic parameter distributions, and imperfect nitrogen allocation can all contribute to a simulation–observation domain gap. Hybrid approaches should therefore be regarded as promising combinations of physical constraints and statistical learning rather than inherently superior retrieval solutions.

Uncertainty characterization remains another major weakness. Most nitrogen retrieval studies report point estimates and conventional accuracy metrics, whereas comparatively few provide predictive uncertainty [70,107,108]. Conceptually, aleatoric uncertainty arises from inherent variability and observational noise, whereas epistemic uncertainty reflects limited knowledge of the model or training domain and can, in principle, be reduced through improved models or more representative data. The latter is especially relevant when retrievals are applied beyond their calibration or simulation domain. GPR provides predictive uncertainty intrinsically, while Bayesian inversion, ensembles, and related approaches extend uncertainty estimation to other retrieval families. Nevertheless, predictive uncertainty should not automatically be interpreted as complete retrieval uncertainty when domain shift or model inadequacy remains unrepresented [114].

Complementary observations provide another route for reducing retrieval ambiguity. Solar-induced fluorescence (SIF) has been explored as a physiological constraint on nitrogen-related functioning [115,116,117,118], while Li et al. [113] demonstrated in coniferous nutrient trials that combining hyperspectral information with far-red SIF and RTM-derived vegetation traits can provide complementary information for needle nitrogen assessment. Meteorological, thermal, and ancillary environmental information have also been incorporated [12,46,61]. Their added value, however, should be demonstrated through controlled comparisons rather than assumed a priori. Across all methodological families, the quality and representativeness of reference nitrogen measurements may ultimately be as important as further algorithmic development [5,12]. Harmonized reference protocols, broad sampling of vegetation and environmental conditions, and genuinely independent validation datasets are needed to evaluate transferability, uncertainty calibration, and methodological improvements under common conditions.

Taken together, methodological choice should reflect available calibration data, required physical consistency, computational constraints, interpretability or uncertainty needs, and the expected degree of extrapolation beyond the training domain. In practical terms, PLSR remains a suitable baseline for limited datasets, ensemble and kernel methods are attractive when nonlinear relationships are expected, GPR is particularly useful when predictive uncertainty and feature relevance are required, deep learning benefits from large and diverse datasets, RTM inversion is preferable when physical consistency is central, and hybrid RTM–ML is especially promising when realistic simulations can provide representative training data. In all cases, method selection should be validated against independent data whenever possible.

Synthesis and Outlook

The preceding sections reveal a methodological landscape spanning parametric regression, nonlinear ML, physically based RTM inversion, and hybrid RTM–ML retrieval. Figure 6 and

Table 4. Qualitative comparison of the principal methodological families for hyperspectral nitrogen retrieval across selected methodological dimensions. The descriptors summarize relative methodological characteristics based on the literature discussed throughout this review and are not intended as quantitative performance rankings.

Method Family	Physical Consistency	Learning Capacity	Operational Efficiency	Transfer- Ability	Interpret- Ability	Uncertainty Quantification
Vegetation indices	Limited	Limited	High	Moderate	High	Limited
Spectral transformations	Limited	Low	High	Moderate	High	Limited
PLSR/PCR	Limited	Moderate	High	Moderate	High	Limited–moderate
Ensemble learning	Limited	High	High	Moderate	Moderate	Moderate
Kernel-based regression	Limited	High	High	Moderate–high	High	High
Deep learning	Limited	Very high	High	Moderate	Limited–moderate	Moderate
RTM inversion	High	Limited	Low	Moderate–high	High	Moderate–high
Hybrid RTM–ML retrieval	High	High	High	Moderate–high	High	Moderate–high

Descriptors provide a qualitative synthesis of methodological characteristics discussed throughout the review and should not be interpreted as standardized performance scores or cross-study accuracy rankings.

and

Table 5. Synthesis of the principal methodological families used for hyperspectral nitrogen retrieval, including current status, main advantages and limitations, and anticipated future role.

Method Family	Current Status	Main Advantages	Main Limitations	Likely Future Role
Vegetation indices	Established empirical approaches	Simple; computationally efficient; interpretable; transferable between sensors with compatible bands.	Limited spectral information; often require crop- or application-specific optimization.	Operational indicators; benchmark methods; complementary predictors in advanced retrieval frameworks.
Spectral transformations	Established preprocessing techniques	Enhance weak nitrogen-related signals; reduce noise; improve hyperspectral feature extraction.	Do not directly retrieve nitrogen variables; require subsequent regression or inversion.	Preprocessing for ML, RTM inversion, and hybrid retrieval frameworks.
PLSR/PCR	Established statistical regression approaches	Computationally efficient; robust to multicollinearity; transparent; suitable as benchmark methods.	Limited representation of complex nonlinear relationships; reduced transferability across sensors and vegetation conditions.	Interpretable baseline methods; benchmarking of more advanced retrieval algorithms.
Ensemble learning (RF, GBDT, XGBoost, LightGBM)	Mature nonlinear ML approaches	Strong predictive performance; robustness; limited preprocessing; compatible with feature selection and model interpretation.	Dependence on representative calibration data; limited extrapolation beyond the training domain; no intrinsic uncertainty estimates without additional strategies.	Operational empirical retrieval; feature selection; model interpretation.
Kernel-based regression (SVR, KRR, GPR)	Mature nonlinear regression approaches	Strong generalization with limited samples; effective for high-dimensional spectra; applicable to empirical and hybrid retrieval. GPR additionally provides predictive uncertainty and ARD-based relevance analysis.	Computational scaling for large datasets; sensitivity to kernel selection and calibration-data representativeness.	Uncertainty-aware retrieval; hybrid RTM–ML inversion; active learning; scalable operational implementation.
Deep learning (CNNs, transformers)	Rapidly developing approaches	Direct learning of spectral and spectral–spatial representations; reduced reliance on manually engineered features.	Large representative training datasets required; challenges in transferability, interpretability, and uncertainty characterization.	Growing role with expanding hyperspectral archives, transfer learning, and self-supervised learning.
Physically based RTM inversion	Established physically based framework	Physically consistent retrieval; improved transferability across sensors, acquisition geometries, and vegetation conditions.	Ill-posed nitrogen retrieval due to weak protein-sensitive signals, parameter equifinality, and sensitivity to RTM assumptions.	Protein-sensitive RTMs; improved parameter constraints; Bayesian inversion; more realistic simulation strategies.
Hybrid RTM–ML retrieval	Rapidly maturing retrieval framework	Combines physically consistent RTM simulations with ML flexibility and computational efficiency; potential for improved transferability and uncertainty-aware retrieval.	Dependence on RTM realism, parameterization, simulation representativeness, and reduction of the synthetic–observed spectral gap.	Promising framework for transferable, scalable, uncertainty-aware, and operational hyperspectral nitrogen retrieval.

summarize their complementary characteristics, relative methodological capabilities, current limitations, and anticipated future roles.

Three broad methodological transitions are evident. First, nitrogen retrieval has expanded from discrete spectral features towards fuller exploitation of contiguous hyperspectral information. Second, linear statistical approaches have increasingly been complemented by nonlinear ML capable of representing complex spectral–nitrogen relationships. Third, physically based RTMs and statistical learning have converged within hybrid RTM–ML frameworks. Together, these developments have expanded predictive flexibility, physical consistency, interpretability, transferability, and uncertainty characterization. These methodological families should not be regarded as mutually exclusive or sequential replacements. Parametric methods remain valuable benchmarks, nonlinear ML provides flexible predictive capability, RTMs contribute physical constraints, and hybrid frameworks increasingly combine these complementary strengths with biological process knowledge.

Future progress is therefore likely to depend less on incremental algorithmic improvements than on advances in protein-sensitive leaf and canopy RTMs, uncertainty propagation, standardized reference datasets, and robust hybrid RTM–ML frameworks. Together with expanding imaging spectroscopy observations from PRISMA, EnMAP, ESA CHIME, and NASA SBG/EAGLE-VSWIR, these developments are expected to support more transferable and uncertainty-aware nitrogen retrieval across sensors, ecosystems, and spatial scales (see also, refs. [5,119,120]).

7. Conclusions

Hyperspectral nitrogen retrieval has progressed from empirical spectral approaches towards nonlinear ML, protein-sensitive RTMs, and increasingly integrated RTM–ML frameworks. Established methods such as PLSR remain valuable benchmarks, while nonlinear and probabilistic approaches provide additional flexibility and uncertainty-aware capabilities. Protein-sensitive RTMs have strengthened the physical basis of nitrogen retrieval, although their interpretation remains constrained by weak spectral signals and uncertainty in the relationship between protein and total nitrogen.

The reviewed evidence does not support a universally superior retrieval method. Method suitability depends on the target variable, sensing scale, vegetation type, calibration-data representativeness, validation strategy, and requirements for interpretability, uncertainty characterization, and extrapolation beyond the training domain. Likewise, the benefits of hybrid RTM–ML approaches depend strongly on RTM realism, the representativeness of simulated training data, and the correspondence between simulated and observed spectral domains.

Three research priorities emerge particularly clearly: (1) independent and standardized validation across sites, years, sensors, and vegetation types; (2) uncertainty frameworks that distinguish and propagate measurement, model, calibration, and domain-shift uncertainties; and (3) improved realism of protein-sensitive leaf and canopy RTMs, including better representation of the protein-to-total-nitrogen relationship and realistic structural and environmental variability.

As a critical narrative review, this study provides a qualitative methodological synthesis rather than a standardized benchmarking of retrieval accuracies. Future progress will therefore depend not only on methodological innovation, but also on stronger validation, improved physical realism, and more representative reference and training data. Together with expanding imaging spectroscopy observations, these advances should support more robust, transferable, and uncertainty-aware nitrogen retrieval across vegetation types and spatial scales.