Nitrogen (N) is a main regulator of several leaf physiological processes, such as photosynthesis, respiration, and transpiration (Reich et al., 2006), and it correlates well with chlorophyll (Chl) content, light use efficiency (LUT), and net primary production (Jiang et al., 2021). N is the principal component which restrains productivity of various terrestrial ecosystems (Richards et al., 2012). This is influenced by the fact that N is a well-known resource that limits plants’ productivity and growth (Ustin, 2013). Spatial patterns of N at leaf level help us to understand the part that terrestrial ecosystems play in the larger Earth system since they are connected to fluctuations of carbon, water, and energy (Ollinger et al., 2008). Canopy N is also associated with the capacity of vegetation to hold atmospheric N deposition (Lindsay, 2017), which has increased together with atmospheric carbon dioxide (CO₂) over the years because of increased use of fossil fuel and production of artificial fertilizer using harmful chemicals (Galloway et al., 2008). The distribution of plant N varies across the landscape due to many variables.

Plant N concentration among plants found in natural ecosystems differs by over 32% (Hobbie et al., 2005) and this varies with topographic gradients and cycling of nutrients, among species (Craine et al., 2009), and in response to manipulations of resource obtainability for experimental purposes (Amundson et al., 2003). Plant N may also be affected by other factors such as land-use history, changes in CO₂ concentration, acid deposition, and limitation of another nutrient element supply (Goodale and Aber, 2001; Hallett and Hornbeck, 1997). Overall, plant N can be used as a measure to generalize patterns of N cycling globally (Lindsay, 2017) as well as a vital indicator to assess changes in N cycling that might accompany anthropogenic influence on ecosystems such as rising atmospheric CO₂ (Huber et al., 2011). As the main nutrient that limits plants’ growth, plant N availability, the way it fluctuates in ecosystems, and how it responds to global change will have long term significant impacts on carbon (C) sequestration in various ecosystems (Luo et al., 2004). Estimations of N at leaf level and canopy N concentration of vegetation allow for a better understanding of ecosystem functioning and biochemical processes (Lee and Nguyen, 2005).

Traditionally, chemical analysis has been used to obtain canopy and foliar N (Serrano et al., 2002). This approach is accurate, but destructive, time-consuming, and costly (Sáez-Plaza et al., 2013). Thus, make it impractical to measure canopy and leaf N for a large area of forests consisting of different vegetation types using the chemical analysis method. To our advantage, remote sensing has turned things around, such assessments can be carried out a low cost, and within a very short space of time across large spatial scales (Walshe et al., 2020). Remote sensing has been used for various applications in vegetation monitoring such as land cover change detection, species classification, land cover estimation, and defoliation (Walshe et al., 2020). Leaf N has been determined in forests (Singh et al., 2015), grassland and crop ecosystems using remote sensing techniques over the years (Yao et al., 2015).

Airborne and spaceborne serve as exceptional tools for temporal and spatial monitoring of leaf N since they provide continuous information in manifold spectral and directional settings (Ustin et al., 2009; Cohen and Goward, 2004). A growing number of hyperspectral imagery reports have been exceptional when it comes to leaf and canopy N prediction for various plants community (Wang and Wei, 2016). However, hyperspectral data are often affected by high dimensionality and are relatively costly to acquire. On the other hand, multispectral data such as Sentinel-2 and field spectroscopy can be used as primary data sources since they are relatively cheaper and easily accessible (Chemura et al., 2018). Launched in 2015, Sentinel-2 is the European Space Agency’s (ESA) state-of-the-art sensor. It has a high spatial resolution (10–60 m range) and 13 bands that span from visible to near-infrared and shortwave infrared (Gholizadeh et al., 2016). Many vegetation characteristics, including the Leaf Area Index (LAI), nitrogen and chlorophyll, biophysical parameters, and Red-Edge Position, can be estimated with the use of Sentinel-2 (Frampton et al., 2013; Clevers and Gitelson, 2013; Verrelst et al., 2013). However, leaf trait contents are distributed differently across vegetation canopies along vertical canopy profile of different species which makes it complex to estimate them (Gara et al., 2018). Other factors such canopy structure, mask of the strong water absorption (Dash and Curran, 2004), viewing geometry and background make it a challenge to retrieve plant N at the canopy level (Grogan and Chapin, 2000).

Nonetheless, retrieval of many vegetation parameters such as plant N, leaf Chl, and others are dominated by empirical techniques such as spectral indices (Gitelson et al., 2003), traditional regression techniques such as stepwise multiple linear regression and partial least square regression (PLSR), to several machine learning approaches such as support vector regression (SVR), neural network and random forest (RF) (Zhao et al., 2013). Spectral indices are one of the simplest and frequently utilized methods to estimate leaf biochemical contents such as N, Chl and vegetation leaf water content (Gitelson A. A. et al., 1996; Wang B. et al., 2018). Spectral indices calculated from the Sentinel-2 product were investigated in this study. The Normalized Difference Vegetation Index, Modified Simple Ratio, and Atmospherically Resistant Vegetation Index (ARVI) are among the broadband indices that are obtained from Sentinel-2 imagery. Normally, the Normalized Difference Vegetation Index (NDVI) is the most widely used vegetation index, and it is typically used to analyse vegetation and its changes over time at both local and regional levels (Ali et al., 2022). The Ratio Vegetation indicator (RVI), the most fundamental indicator used in vegetation study and monitoring, has been refined into the Modified Simple Ratio (MSR) (Calvão and Palmeirim, 2011). N in leaf cells mostly occurs in proteins and chlorophylls (Kokaly et al., 2009). Plant N and Chl correlate well across various plant species (Homolova et al., 2013), therefore, spectral indices formulated for Chl have been employed as a means of N estimation (Le Maire et al., 2008). This study aimed to map plant N concentration using different categories of spectral indices and stepwise multiple linear regression in the Luvuvhu River Catchment.

The aim of this study was to estimate and map plant nitrogen concentration of indigenous vegetation in the Luvuvhu River Catchment in Limpopo province using Sentinel-2 data. Remote sensing makes it feasible to monitor plant N concentration at a large scale effectively and at a low cost. However, remote sensing mostly obtains spectral reflectance of the canopy of vegetation which requires the need to understand spectral traits of vegetation to be mapped. This allows for the establishment of the relationship between physiological, ecological variables, and spectral reflectance which will allow rapid N estimation and vegetation health monitoring (Vigneau et al., 2011; Osbome et al., 2011). Spectral reflectance of vegetation is influenced by many factors which result in changes in spectral reflectance across the study area (Liu et al., 2019). Therefore, this presents a challenge to come up with spectral indices which are simple and applicable for plant N estimation.

However, Sentinel-2 allows for the construction of chlorophyll-related spectral indices using the red-edge part of the spectrum, and since chlorophyll serves as a proxy to nitrogen, such spectral indices can be explored for indirect nitrogen estimation in the forest, which has yielded good results in many ecosystems such as grassland and for agricultural crops (Clevers and Gitelson, 2013). However, spectral indices also have many factors that mostly affect them (Zheng et al., 2018). For vegetation parameters estimations, some of spectral indices have excessive capability for soil noise reduction (Rondeaux et al., 1996), while others are very useful by avoiding at saturation for vegetation spectra (Jordan, 1969). In cases where the ecosystem is nitrogen-rich, the correlation between chlorophyll and nitrogen becomes weak (Hallik et al., 2009). Hence, it is very crucial to be considerate when choosing spectral indices that are related to chlorophyll for nitrogen estimation, and to consider the features that are sensitive to nitrogen when selecting spectral indices to estimate plant N (Osco et al., 2020; Zheng et al., 2018). NIR serves as a principal constituent for various spectral indices (Rondeaux et al., 1996). However, in practice, the red edge is one of the most utilized spectral features for vegetation health evaluation (Vigneau et al., 2011).

Therefore, in this study, three groups of spectral indices linked to the biochemical and physical properties of vegetation such as nitrogen, structure, and chlorophyll were used with Sentinel-2 imagery to estimate plant N in the study area. Three models were built for plant N concentration in this study, namely, red edge, nitrogen based and combined spectral indices to explore the relationship between predicted and observed N values for plant N concentration estimation. Spectral indices constructed from the combination of bands (red edge and nitrogen based) in Sentinel-2 data have the potential to extract vegetation properties and be utilised for a wide range of application such as plant N, chlorophyll, and water content estimation (Li et al., 2014a; Hunt Jr et al., 2013). A combination of spectral regression models which are suitable for plant N estimation are very useful for vegetation health monitoring. The SWIR and the red edge region had the greatest impact on plant N estimation among the evaluated spectral indices. The red edge is a useful band for tracking vegetative metrics, as is the case with most of the research (Jiang et al., 2021). A sensitive area for tracking the amount of chlorophyll in plants is the red edge, which exhibits abrupt variations in reflectance. The degree of chlorophyll absorption determines the red edge zone (Li et al., 2014b). Nevertheless, the study’s findings also show that SWIR can successfully raise plant N estimation accuracy. Similarly, the SWIR was proposed by Perich et al. (2021) as the optimal zone for plant N status. This could be because plants have nitrogen in their protein and amino acid molecules. Generally, short-wave infrared light is absorbed by proteins. Proteins in the plant absorb light in this band, which causes absorption peaks to appear in the spectrum when using the SWIR to track plant growth.

Furthermore, stepwise regression was used to determine the category of spectral indices with better performance when it comes to predicting plant nitrogen. Regression model of combined spectral indices from the red edge and the broadband was used in this study since it showed higher correlation between the observed and predicted values of plant N which presents reference for the plant N monitoring using spectral indices in the subtropical environment. The combination of broadband and narrowband spectral indices performed best for this study with R² = 0.59. According to the Darvishzadeh et al. (2008) study, foliar N and indices have been found to correlate, and both narrowband and broadband regions appear to be significant predictors of foliar N. The method of using the whole spectrum has yielded the best results since it was able to leverage important regions across the spectrum and discern differences in foliar N both between and within varying biomes (Darvishzadeh et al., 2008). The broadband spectral indices came second with R² = 0.44. Foody et al. (2001) study supported our results according to which there is poor correlation between the observed and predicted plant N using the broadband spectral indices. Main problem is saturation hence the accuracy is less; saturation can be due to complex structure of forest in result of which broadband VIs did not observe the plant N increase. Finally, the narrowband spectral indices with R² = 0.35. Wang et al. (2016) studied that reason of saturation is the computation of indices through NIR and red spectral band and the red band absorbs electromagnetic radiation and remains constant while the canopy cover increases hence in result at 100% canopy cover red energy absorption reaches to peak. Also, the soil reflectance and topography influence plant N estimation using spectral indices therefore assessment of different spectral indices is very useful (Heiskanen, 2006). However, the study by Imran et at., (2020) has shown that a relationship of plant N with these vegetation indices is best due to the reason that red-edge band is located in high chlorophyll reflectance and absorption area which is between red and NIR region due to which the change in chlorophyll as well as leaf properties have greater effect on the red-edge spectral band. It was found that red edge indices have more accuracy and correlation as compared to broadband indices (Xie et al., 2018). The NIR region can contain spectral variation due to structural variables that covary with foliar N concentrations and can influence overall reflectance and absorption from this region (Knyazikhin et al., 2013; Lepine et al., 2016; Ollinger et al., 2008). For this investigation, summed reflectance values over the whole NIR area did not account for more than 35% of the variance in foliar N, even though narrowband within this region consistently emerged as significant predictors of foliar N. Since only specific portions of the NIR were important, this might be due to atmospheric effects rather than broad structural changes with changing N concentrations. Variation of plant nitrogen concentration is well explained by functional and species composition in the study area. This is dependable looking into studies from various ecosystems, such as temperate, tropical, boreal, and Mediterranean ecosystems (Asner and Martin, 2009).

The resultant map shown in Figure 6 Indicates density of forest is higher at northern part of study area since have higher biomass as well as towards southwestern and northeastern side vegetation is less hence the plant N is relatively higher compared to other parts. The southwestern aspect is cool hence have more vegetation cover and plant N as compared to northern aspect which is warmer.

This study compared the performance of three groups of spectral indices which are mostly related to nitrogen, chlorophyll, and structural properties, derived from Sentinel-2 imagery to estimate the concentration of plant N in a subtropical environment of South Africa. Further, it was found that there are still very few studies in South Africa dedicated to utilizing Sentinel-2 data for plant N estimation. Plant N varies greatly across the study area depending on species composition. The resultant map shown in Figure 6. Indicates density of forest is higher at northern part of study area since have higher biomass as well as towards southwestern and northeastern side vegetation is less hence the plant N is relatively higher compared to other parts. The southwestern aspect is cool hence have more vegetation cover and plant N as compared to northern aspect which is warmer.

Results from stepwise regression showed that showed that plant N can be estimated using different groups of indices with acceptable accuracy in the subtropical environment. Results have shown that combined spectral indices performed better with R² = 0.59, RMSE = 0.47% and MAE = 0.38%, followed by N related spectral indices with R² = 0.44, RMSE = 0.65% and MAE = 0.48%, and the last category is red edge based spectral indices with R² = 0.35, RMSE = 0.81% and MAE = 0.65%. Most of the structural related indices could estimate nitrogen at a moderate to good accuracy, due to the functional link between canopy structure and nitrogen resulting from functional type and species differences. Improvements could be accomplished by using multiple VIs with differing feature relevance according to different methods and by using hyperspectral data instead of Sentinel-2 imagery. Although the VIs from different ways could have different feature relevance, stacking them while considering different methodologies could prevent overfitting and should be used to monitor vegetation biophysical parameters and improve classification accuracy. We expect that these predictions will also be very helpful in limiting the uncertainty of biogeochemical models for a variety of terrestrial ecosystems around the world.