Introduction

Potatoes (Solanum tuberosum L.) are important non-grain food crop due to their high nutritional value and energy yield (DeFauw et al., 2012; Devaux et al., 2020; Khalid & Akhtar, 2014; Lutaladio & Castaldi, 2009). However, maximizing potato yield and quality while minimizing nitrogen (N) losses is difficult for growers due to their shallow roots, well-drained soil, and high N fertilizer requirements, which can exacerbate N leaching and decrease N use efficiency (DeFauw et al., 2012; Tanner et al., 1982; Zvomuya et al., 2003). Hence, N management is critical but a difficult aspect of potato production system. Ultimately, growers require effective tools and techniques to regularly monitor crop N status to optimize N management (Goffart et al., 2008; Sharma & Bali, 2017).

The current method for monitoring crop N status involves analyzing nitrate N in petiole samples, which is time-consuming and may not reflect their total N status as petiole analysis does not reflect the cumulative but only the immediate N status of the plants (Bélanger et al., 2003; Porter & Sisson, 1991; Rosen & Eliason, 2005). Although leaves can indicate the cumulative nutritional status of plants (Rosen & Eliason, 2005), the wet lab analysis required for leaf samples is expensive and time-consuming (Kalra, 1998). Hence, a cumulative, cheaper, and rapid assessment of crop N status is needed for accurate and regular monitoring of the N demand of potatoes. Leaf spectral reflectance obtained using narrow band spectral resolution (< 10 nm) correlates better with leaf N contents than with petiole nitrate–N (Cohen et al., 2010; Zhou et al., 2022). Thus, proximal sensor-based leaf analysis is a potential solution to overcome these limitations (Menesatti et al., 2010; Min et al., 2008; Mishra et al., 2017). The use of aerial photographs or satellite imagery is a way to evaluate plant N status and yield (Haboudane et al., 2002; Wooten et al., 1999). However, satellite data offers wide spectral bands (50–100 nm) and low spatial resolution (10–30 m) and cannot replace narrow band (1–2 nm) remote sensing systems. The number of research and commercial hyperspectral satellites are growing, with improved spatial resolution (Gomez et al., 2008; Nigon, 2012; Puschell 2020). While some hyperspectral remote sensing and airborne systems can provide narrow band (1.8–2.2 nm) hyperspectral data and have broader spatial coverage, it is not the focus of our study and hence not widely discussed further.

Vis-NIR spectroscopy is a user-friendly, rapid, and cost-effective analytical technique that provides better reproducibility without emissions or waste (Dedeoglu, 2020; Lopez et al., 2013). Compared to wet lab techniques, vis-NIR technology is faster, cheaper, and more comprehensive since it does not require sample preparation or reagents and can incorporate field-level spatial variations (Capitan-Vallvey & Palma, 2011). Vis-NIR is the region of electromagnetic spectrum that extends between wavelengths of 350 to 780 nm in the visible region and 780 to 2500 nm in the NIR region (McBratney et al., 2003; Viscarra Rossel et al., 2006). Vis–NIR radiations selectively absorb depending on the specific vibration frequencies of the molecules present in the sample, producing a spectrum that can predict the concentration and type of the matter in the object (Murray & Cowe, 2004; Shenk & Westerhaus, 1994; Stenberg et al., 2010). Vis–NIR can provide information on the sample’s biochemical and structural constituents (Richardson et al., 2004), making it useful in agriculture (Lai et al., 1994; Wu et al., 2008). Hence, it enables growers to detect excess or deficient nutrients in the crop and optimize nutrient management in their cropping system by maintaining the synchrony between crop nutrient demand and their supply from soil.

In the past, researchers have used NIR spectroscopy to predict soil properties related to soil fertility and soil nutrient analysis, but the actual plant availability of these soil nutrients is uncertain (Recena et al., 2019; Stenberg et al., 2010). Vis-NIR spectroscopy has been used to determine N contents in plant leaf tissues in agriculture for various crops (Cohen et al., 2010; Lee et al., 1999; Thomas & Oerther, 1972; Ulissi et al., 2011). Most studies have either utilized spectra in the visible or NIR region for predicting leaf N contents for many crops including potatoes (Cohen et al., 2010; Menesatti et al., 2010; Váradi et al., 1987; Young et al., 1997; Zhou et al., 2022). However, there are limited studies determining leaf N content in potatoes utilizing reflectance spectra in the vis-NIR (350–2500 nm) region (Liu et al., 2021; Min et al., 2006). Similarly, most studies on leaf nutrient analysis using NIR spectroscopy have focused on dried, ground leaves and only a few on fresh leaves either intact or removed (Prananto et al., 2020a, 2020b). Furthermore, limited studies have validated the leaf N prediction models for potatoes using vis–NIR spectroscopy, especially when considering factors like maturity classes/varieties, as seen on sweet pepper (Thomas & Oerther, 1972) and corn (Lee et al., 1999). Similarly, for factors like growth stages (Alchanatis et al., 2005), management practices, and environmental conditions (Azadnia et al., 2023; Cohen et al., 2010; Min et al., 2006; Ulissi et al., 2011) there are only a few studies conducted. Hence, more research is needed to understand the reliability of determining N contents of fresh, removed leaves using vis–NIR spectroscopy under above-mentioned conditions.

This study aimed to develop and validate partial least square regression (PLSR) to estimate the total N contents of fresh and removed leaves of Russet potatoes using diffused reflectance spectra in the vis–NIR spectral range generated from a handheld proximal sensor. The first objective was to build a calibration model using data collected from potatoes across different varieties, N rates, growth stages on sandy soils at Hancock Agricultural Research Station (ARS), WI in 2020. The second objective was to validate the PLSR model using samples collected in 2021 for four different conditions, including different locations, potato varieties, N rates, and growth stages.

Materials and methods

Experimental design and measures

The study was conducted over a two-year period from 2020 to 2021 at two locations: Hancock Agricultural Research Station (HARS; 44°8′23" N, 89°31′23" W; elevation: 328 m) (both years 2020–21) and a farmer's field (Coloma; 44°1′22" N, -89°35′50" W) (in year 2021) (Table 1). Both locations had Plainfield loamy sand soils (Mixed, mesic Typic Udipsamments) and were situated in the central sand region of WI, USA. In 2020, each treatment plot measured 30 m (m) in length, while in 2021, they were 35 m long. Each plot consisted of four rows, approximately 4 m in width, with 1 m between rows and 30 cm between plants within rows. For both years and locations, leaf samples were randomly collected from the middle two rows, covering the entire length of each plot. More information regarding sampling dates and the number of leaves sampled has been reported in Table 1. At Hancock, four popular Russet potato varieties (Russet Burbank, Goldrush, Russet Norkotah, and Silverton Russet) were sampled under both N-fertilized (308 kg N ha−1) and unfertilized conditions in a randomized complete block design, with three replicates per treatment and a total of 24 experimental plots. At the farmer's field, the Lakeview Russet potato variety was grown under three different N-fertilization rates (67, 135, and 202 kg N ha−1), with four replicates per treatment and a total of 12 experimental plots in a randomized complete block design. The N-unfertilized plots did not receive any additional N apart from the starter N fertilizer (37 kg N ha−1). Based on previous trials, it was expected that 308 kg N ha−1 was a non-limiting application of N when split applied three times during the growing season. All the fertilized plots received three applications of N fertilizers at hilling (27% of the total N input), tuber initiation (36% of the total N input), and two weeks after tuber initiation (remaining N input) with ammonium sulfate and ammonium nitrate-based N fertilizers. The first N application was broadcast applied then incorporated via hilling. The final two applications were hand applied on top of the hills. Additionally, all plots were irrigated and fertilized with other essential nutrients such as P, K, S, and Ca to avoid any nutrient limitations to the plants. For our experimental plots, we have gathered data on various soil characteristics, including soil organic matter (was mostly below 1%), soil pH (ranged from acidic to neutral), and soil total levels of P and K. However, we did not collect data on pre-planting status of nitrate or ammonium levels in these irrigated sandy soils.

Table 1 Leaf sampling information (no. of leaves, sampling dates, potato varieties sampled, typical growth stages, nitrogen fertilization rates) collected from Hancock and Coloma. The information on typical growth stages of potato was referenced from Rosen et al., 2020
Full size table

Throughout the growing season, leaf samples were randomly collected for reflectance measurement from the plants covering the entire sampling row of each treatment plot representing four growth stages of the plants (vegetative, tuber initiation, tuber bulking, and tuber maturation) (Table 1). Afterward, the leaves were dried in a plant dryer at 60 °C until a constant weight was achieved. The dried samples were then ground to pass a 0.63-mm screen and stored prior to laboratory analyses. Total N (%) contents were analyzed using the dry combustion-based Dumas method at the Kansas State Soil Test Lab. These varieties were chosen not only based on their popularity but also to represent a range of characteristics such as maturity classes and tuber density commonly found in commercial potato production. Similarly, leaf samples were collected at different sites under different N fertilization inputs may reflect differences in environmental or management practices.

Spectra acquisition

Once, the leaves were collected from the field, they were transported to the UW-Madison soil laboratory for spectral data collection. A portable handheld proximal sensor, the PSR + 3500 vis–NIR spectroradiometer (Spectral Revolution Inc., Lawrence, MA, USA) was used to measure leaf diffuse reflectance spectra in the range of 350–2500 nm with varying spectral resolutions: 3 nm at 350–1000 nm, 8 nm at 1000–1900 nm, and 6 nm at 1900–2500 nm. The experiment was conducted at room temperature (~ 25 °C). To control the effects of external environment such as temperature and weather on the scans, conducting scans in removed leaves under controlled environments is a better method than conducting scans in intact leaves in the field (Prananto et al., 2020a, 2020b; Read et al., 2002). Fresh leaf samples were scanned using a leaf clip probe was attached to the vis–NIR spectrometer via a bifurcated fiber optic cable. The leaf clip included a 5 W halogen illuminator, and a high-diffuse reflectance polytetrafluoroethylene (PTFE) white plate was used to calibrate the spectrometer. Calibration was conducted once every ten sample scans to ensure baseline correction before scanning the samples from the next treatment plot. A portable PC (Getac) was connected to the vis–NIR spectrometer to manage the measurement and store data. The adaxial epidermis of the leaf was secured in the leaf clip and scanned by the vis–NIR spectrometer. To minimize noise, 30 readings were taken for each measurement and averaged to produce one spectrum. All the leaves were scanned twice, one scan perpendicular to the other, and the two spectra were averaged to obtain one spectrum per sample for further analysis. After scanning, the leaves were dried at 60 °C and analyzed in the lab for total N measurements using the dry combustion-based Dumas method.

Spectral preprocessing

We conducted spectral preprocessing to prepare the vis–NIR reflectance spectra for modeling. Initially, the spectra were averaged for each leaf sample. Any poor or distorted spectra were removed during spectra processing. The reflectance (R) spectrum was then resampled by the spectrometer to 1-nm resolution for export, resulting in each spectrum comprising 2151 spectral points. Due to the collinearity of the spectra, we retained only every 10th wavelength from 360 to 2490 nm resulting in leaves with 214 spectral points were used for analysis. To enhance predictive accuracy, we converted the reflectance (R) of all the samples into absorbance (A = Log (1/R)) for further analysis (Fig. 1), aligning with findings from previous research (Kokaly & Clark, 1999; Min et al., 2006). We also observed, in accordance with the work of Viscarra Rossel et al. (2009), the absorbance spectra tend to exhibit higher prediction accuracy than reflectance spectra. Similarly, we followed the recommendation of Ge et al. (2019) and Zhai et al. (2013) and observed that spectral data from 350 to 450 nm and 2450 to 2500 nm exhibited higher noise-to-signal ratio and were removed and excluded from downstream spectral analysis (Rodrigues et al., 2020) (Fig. 2). The spectra pre-processing such as smoothening, resampling, and normalizing before developing predictive models was conducted using prospectr-package in R. The spectral data was smoothened using a Savitzky–Golay smooth filter (window size = 5 and polynomial order = 2) (Savitzky & Golay, 1964) and was resampled using a spline interpolation method to every five nm to reduce the dimensionality of the predictors and enhance computational efficiency to predict total N without over fitting (Ge et al., 2019). The spectrometer has three sensors with different spectral resolutions (3 nm to 8 nm). The spectrometer automatically resamples spectra to 1-nm resolution to export the spectrum. Since there are over 2000 spectral bands, which are highly collinear, we resampled them to a coarse resolution to reduce the numbers from thousands to hundreds. Thus, eliminating the issue of multi-collinearity due to highly correlated spectral bands. The spectral data was normalized by using a standard normal variate function which subtracts each treatment row by its mean spectral absorbance and then divides by its standard deviation (Barnes et al., 1989).

Fig. 1
figure 1

All the absorbance spectra calculated as the log of the inverse of the reflectance (R) spectra obtained from scanning the leaves using vis-NIR spectroradiometer in the range of 350–2500 nm before spectral processing

Full size image
Fig. 2
figure 2

All the absorbance spectra calculated as the log of the inverse of the reflectance (R) spectra obtained from scanning the leaves using vis–NIR spectroradiometer in the range of 500–2450 nm after spectral processing

Full size image

Model development and evaluation

In our study, we implemented a train-test split approach based on the year of data collection. The data collected in 2020 were utilized for training the model, whereas the data collected in 2021 were utilized for evaluating the model's performance. It should be noted that the data from 2021 were further subdivided into subsets based on various factors, including varieties, sites, fertilizer rates, and growth stages. This subdivision allowed us to explore potential variations within the testing dataset. Using data from a separate year for validation ensures that the model is evaluated on an independent dataset, reducing the risk of overfitting, and providing a more realistic assessment of the model's performance in predicting leaf N content.

We used Partial Least Squares Regression (PLSR), a multivariate modelling approach, to build and evaluate the model (Menesatti et al., 2010; Rodrigues et al., 2020) using ‘pls’ package in R 3.6.3 (R Core Team, 2020). Our preliminary data analysis showed that PLSR provided better prediction of leaf N contents in our samples compared to other machine learning models, such as cubist and random forest. Similarly, PLSR have been successfully and widely used to develop prediction models for the leaf N contents using vis–NIR hyperspectral data as demonstrated by Rodriques et al. (2020) and Tallada et al. (2018). The independent variables used to build the PLSR model were the spectral signatures obtained from the vis–NIR spectroscopy. The model used these independent variables to predict the dependent variable, i.e., total N. Once the model was built, it was used to predict the total N of new samples based on their spectral signatures, which were not used to build the model.

Repeated k-(10) fold cross-validation was used to tune the model, determine the tuning parameter, and reduce overfitting issues (Helland, 2006). The best tuned model was selected to predict leaf total N to the test (validation) dataset. To perform cross-validation, the trainControl function with repeated k-fold cross-validation resampling method from caret package was used in R (3.6.3). The parameters were set as tenfold cross-validation and repeated ten times. In the PLSR, the maximum 30 PLS components were considered, while the best pls model was selected by the smallest RMSE. Similarly, the coefficients of the spectral variables used in the prediction model was extracted from the best pls model using the “coefficients” function in R (3.6.3). The coefficients represented the importance of each spectral variable in the model for predicting the response variable. These combined cross-validation techniques ensured a robust assessment of the model's accuracy and generalization, enhancing the reliability of our findings.

Model performance evaluation

We evaluated the quality and accuracy of both the calibration and validation models using various statistical parameters such as coefficient of determination (R2), root mean squared error (RMSE), ratio of performance to deviation (RPD), ratio of performance to interquartile range (RPIQ), mean absolute error (MAE), and bias. A model with high R2, low RMSE, high RPD and RPIQ, and low bias were considered a good model. RPD and RPIQ provide a measure of the model's predictive ability relative to the natural variation in the dataset. RPD is calculated as the ratio of the standard deviation of the measured values and the root mean squared error of the model predictions whereas RPIQ is the ratio of the interquartile range of the measured data and the root mean squared error of the model predictions. Hence, both RPD and RPIQ help assess the model’s performance in relation to the variability of the measured data.

As per the classification rubric by Chang et al. (2001), if RPD > 2 and R2 between 0.8 and 1, RPD between 1.4 and 2 and R2 between 0.5 and 0.8, and RPD < 1.4 and R2 < 0.5, then the prediction models were categorized into good (category A), moderate (category B), and poor (Category C) respectively. If the model fell in category B, then the model prediction may be improved by using different calibration models. Whereas, if the model prediction fell in category C, then total leaf N might not be reliably predicted using vis–NIR spectroscopy under respective conditions. We also included the Lin's Concordance Correlation.

Coefficient (LCCC) values to assess the 1:1 fit of predictions providing a comprehensive evaluation of accuracy and was calculated as \(LCCC= \frac{2\times \mathrm{cov }\left(\mathrm{actual},\mathrm{ predicted}\right)\times \mathrm{mean }\left(\mathrm{actual}\right)\times mean (predicted)}{var \left(actual\right)+var \left(predicted\right)+\left(mean \left(actual\right)\right)2+\left(mean\left(predicted\right)\right)2}\), where, cov = covariance between actual and predicted and var = variance of actual or predicted values. LCCC values greater than 0 but less than 1 indicates a positive linear relationship, with values closer to 1 indicates a stronger relationship between observed and predicted values whereas, values less than 0 indicate a negative linear relationship, with values closer to 0 indicate a poor relationship between observed and predicted values.

Results

The summary statistics of the training and testing datasets on leaf N contents are shown in Table 2. The model evaluation statistics showed that the PLSR model had a high predictive performance for estimating leaf total N contents with R2 of 0.95, MAE of 0.2%, RMSE of 0.3%, bias of 0%, RPD of 4.3, and RPIQ of 7.8 (Table 3). The validation of the model had R2 of 0.89, MAE of 0.4%, RMSE of 0.5%, bias of -0.3%, RPD of 2.5, and RPIQ of 4.1 all the samples from the year 201 (Table 3). Hence, as per the classification rubric by Chang et al. (2001), the calibration and validation models for both sites were categorized into Category A as they had R2 ≥ 0.8 and RPD ≥ 2. The validation of the PLSR model for different Russet potato varieties grown at Hancock in 2021 showed a high predictive performance, with R2 values ranging from 0.82 to 0.91 and RPD values ranging from 1.9 to 2.4 (Table 3). The model was also validated for two extreme N rates (0 and 308 kg N ha-1), and the results showed that the predictive.

Table 2 Summary statistics of the training (calibration and testing (validation) datasets on leaf N contents where n = number of samples, min = minimum total N contents, max = maximum total N contents, SD = standard deviation of the data
Full size table
Table 3 Model evaluation criteria for leaf total N contents using vis–NIR spectroscopy includes coefficient of determination (R2), concordance, mean absolute error (MAE), root mean square error (RMSE), residual prediction deviation (RPD), ratio of performance to interquartile (RPIQ) range, and bias
Full size table

performance was good, with R2 of 0.83 and 0.84 and RPD values of 2.3 and 1.4 for N unfertilized and fertilized conditions respectively (Table 3). Finally, the model was also validated for the four growth stages of Russet potato varieties, including vegetative, tuber initiation, tuber bulking, and tuber maturation. The results showed that the model performance was high (category A) for tuber bulking and tuber initiation, with R2 of 0.95 and 0.89 and RMSE of 0.7 and 0.3 respectively. The performance of the model for vegetative (category C) and tuber maturation (category B) stages was relatively low, with R2 values of 0.41 and 0.58, respectively. The predicted and actual total N values in the leaves were visually represented through various graphs on Figs. 3, 4, 5, 6, 7, and 8. We obtained the LCCC value of 0.97 indicating a strong positive linear relationship between the predicted and observed values in the calibration model. We obtained LCCC value of 0.90 between the predicted and observed values in the validation of the model using entire dataset from 2021. Similarly, using the validation dataset for locations, varieties, N rates, and growth stages, we obtained LCCC values of 0.92 (Coloma), 0.87 (Hancock); 0.82 (Goldrush), 0.86 (Burbank), 0.82 (Norkotah), 0.83 (Silverton); 0.88 (unfertilized), 0.73 (fertilized); and 0.29 (vegetative), 0.88 (tuber initiation), 0.77 (tuber bulking), 0.69 (tuber maturation).

Fig. 3
figure 3

A linear regression between predicted and measured leaf N contents for a calibration model built using data collected from Hancock in 2020 (n = 167). Solid line represents the 1:1 regression line between measured and predicted total N contents (intercept = 0, slope = 1)

Full size image
Fig. 4
figure 4

A linear regression between predicted and measured leaf N contents for the entire data collected in 2021 (n = 144) for validation of the model built using samples from Hancock in 2020. Solid line represents the 1:1 regression line between measured and predicted total N contents (intercept = 0, slope = 1)

Full size image
Fig. 5
figure 5

Linear regression between predicted and measured leaf N contents for different locations (Coloma (n = 48) and Hancock (n = 96)) for validation of the model built using samples from Hancock in 2020. Solid line represents the 1:1 regression line between measured and predicted total N contents (intercept = 0, slope = 1)

Full size image
Fig. 6
figure 6

Linear regression between predicted and measured leaf N contents for Goldrush (n = 24), Russet Burbank (n = 24), Russet Norkotah (n = 24), and Silverton Russet (n = 24), potato varieties for validation of the model built using samples from Hancock in 2020. Solid line represents the 1:1 regression line between measured and predicted total N contents (intercept = 0, slope = 1)

Full size image
Fig. 7
figure 7

Linear regression between predicted and measured leaf N contents for fertilized (308 kg N ha.−1) (n = 48) and unfertilized (n = 48) conditions for validation of the model built using samples from Hancock in 2020. Solid line represents the 1:1 regression line between measured and predicted total N contents (intercept = 0, slope = 1)

Full size image
Fig. 8
figure 8

Linear regression between predicted and measured leaf N contents for vegetative (41 DAP, n = 24), tuber initiation (62 DAP, n = 24), tuber bulking (77 DAP, n = 24), and tuber maturation (97 DAP, n = 24) growth stages of Russet potatoes for validation of the model built using samples from Hancock in 2020. Solid line represents the 1:1 regression line between measured and predicted total N contents (intercept = 0, slope = 1)

Full size image

The line plot showing the variation of the spectral coefficients from 13 pls components across different wavelengths, overlaid with a linear regression line was graphed in Fig. 9. The wavelengths: 500, 650, 660, 680, and 700 nm (in the visible spectral range) and 1940, 1980, 2110, 2120, 2130, 2170, 2180, 2190, 2200, 2330, and 2390 nm (in the Near-Infrared spectral range) significantly contributed to the model development. All the wavelengths with values (> 1) were considered significant as per Zhou et al. (2022). However due to the larger number of significant wavelengths, only the wavelengths with model coefficients ≥  ± 2.2 were identified in the study and presented in Fig. 9.

Fig. 9
figure 9

Important spectral ranges for predicting leaf total N contents: A graph of coefficients from the PLSR model, highlighting wavelengths (nm) in the visible and near-infrared region with model coefficients greater than ± 2.2

Full size image

Discussion

The PLSR models developed and validated using vis–NIR spectroscopy (500–2450 nm) showed overall good predictive performance for estimating total N contents of fresh and removed leaves of Russet potatoes as suggested by Card et al. (1988), Cohen et al., (2010), Liu et al. (2021), Min et al. (2006), Váradi et al. (1987), Young et al. (1997), and Zhou et al. (2022). Our study was established upon Cohen et al., (2010) claims which was also supported by Zhou et al. (2022) where they found that leaf spectral reflectance obtained using spectrometer with narrow band spectral resolution (< 10 nm) correlated better with leaf N contents than with petiole nitrate–N. Hence, in this study, leaf N was predicted rather than petiole nitrate–N using vis–NIR spectroscopy.

While most of the spectral studies either utilized or suggested the best prediction using reflectance spectra in the visible (Cohen et al., 2010; Zhou et al., 2022) or NIR (Váradi et al., 1987; Young et al., 1997) region alone, however, Min et al. (2006) measured reflectance in the larger range of 400–2500 nm and Liu et al. (2021) found the best PLSR model prediction in the similar range (400–2350 nm) as compared to the spectra in the range of 400–1300 nm or 1400–2350 nm alone for predicting potato N status. Therefore, our study aligned with Min and Liu et al.’s results, suggesting that using spectra in a wider range of wavelengths instead of using spectra only in either visible or NIR region could improve the accuracy of PLSR-based model predictions.

The study also found that leaf samples collected from potatoes growing under two extreme N fertilization conditions and two different sites may reflect differences in environmental conditions and management practices, but this did not impact the model’s predictive ability. Most leaves collected from fertilized plants had total N contents greater than 6%, while leaves collected from plants under unfertilized conditions had N contents less than 6%. The models, however, underpredicted the measured total N contents greater than 6%. Consequently, the models may have performed slightly better for plants under unfertilized (RPD > 2, LCCC = 0.88) than fertilized (RPD < 2, LCCC = 0.73) conditions. Despite some variations, the model R2 was > 0.8 under both N fertilization conditions, indicating a good model prediction as suggested by Thomas and Oerther (1972) (Sweet pepper). Additionally, the model prediction was similar and good for both sites, as represented by R2 > 0.8, RPD > 2, and LCCC closer to 1.0. While the selected varieties differed in their characteristics such as tuber specific gravity and maturity, which could have an impact on the spectral characteristics of leaf samples. However, the results demonstrated that the model had a good prediction ability for all the Russet varieties. Finally, the model showed a good to moderate prediction of leaf N for all the growth stages except for vegetative stage where the model had relatively low prediction performance exhibited through lower R2, RPD, RPIQ, and LCCC values as compared to later growth stages when belowground tuber biomass started to grow and bulk. Potato crop typically accumulate more N in their aboveground (leaf and stem) biomass earlier in the growing season (Gayler et al., 2002), potentially leading to underprediction by our models during vegetative stage as leaf N contents were greater than 6%. Hence, the discrepancy could be because total leaf N measurements were highest (mean TN = 6.8%) during the vegetative stage as compared to the later stages of the plant growth (Fig. 10). To improve model’s performance for fertilized plants and for plants during their vegetative growth stages, we propose calibrating our model using a training dataset that includes spectral data and corresponding leaf N measurements greater than 6%, enabling the model to better capture the patterns associated with high N and ultimately improving its performance.

Fig. 10
figure 10

Scatter plot of measured total nitrogen (N) contents across four different growth stages (vegetative, tuber initiation, tuber bulking, and tuber maturation) in potato plants collected at Hancock in 2021, along with the mean total N contents. Growth stage information was referenced from Rosen et al. (2020)

Full size image

Herrmann et al (2010) found that potato N content was correlated with the combined reflectance from 1510 and 660 nm which supported our result where 660 nm was one of the significant wavelengths predicting potato leaf N contents (Fig. 9). Our results had corroborated findings by Zhou et al. (2022) who had found the wavelengths between 400 and 700 nm and 650 and 830 nm significantly predicted potato leaf N contents using PLSR models as our study had found that wavelengths between 500–780 nm mostly contributed to the N prediction in the visible range. Whereas wavelengths between 800 and 1000 nm had not contributed to the model prediction in our study. Min and Lee (2005) supported our results as they had identified 2231 nm wavelength using PLSR technique as a significant wavelength for N detection on citrus trees. Since only every 10th wavelength from 360–2490 nm was retained in our study during analysis. Hence, the study found that a wavelength of 2240 nm (which had been closer to 2231 nm) was significant in the prediction of leaf total N content on potatoes.

An R.

2

Of 0.95 (p \ 0.01) and overall accuracy of 80.5% (Kappa = 74%) were.

determined for both vegetative and tuber-bulking periods.

An R.

2

Of 0.95 (p \ 0.01) and overall accuracy of 80.5% (Kappa = 74%) were.

determined for both vegetative and tuber-bulking periods.

N is a major component of photosynthetic pigments, internal structures, and various organic compounds in plant leaves. Jensen (2000) had reported that chlorophyll a absorbed light at wavelengths of 430 and 660 nm, and chlorophyll b absorbed light at 450 and 650 nm. Hence, the peaks were associated with absorption of light by organic compounds that contain N, such as chlorophyll and nucleic acids. Similarly, wavelengths in the region of 1400–2500 nm are related to leaf biochemistry such as sugars and starches, proteins, and ligno-cellulose (Curran, 1989; Féret et al., 2021). Liu et al. (2021) had shown that N was correlated with the intensity of signature peaks in this spectrum range. Therefore, by measuring the intensity of these peaks in the vis–NIR range of the electromagnetic spectrum of a leaf sample, it had been possible to estimate the N content in the sample. While our study did not directly measure leaf biochemistry or chlorophyll content, the literature had suggested that using wavelengths in the range of 500–2450 nm and identifying significant wavelengths in that range could accurately predict leaf N contents. Thus, our study’s use of these wavelengths aligned with the existing research, indicating the potential for accurately predicting N contents in potato leaves using spectral data.

Conclusions

Vis–NIR spectroscopy generated from a handheld proximal sensor was found to be a reliable technique to determine N contents in fresh and removed leaves of potatoes at different N fertilization, maturity classes, growth stages, and locations. Partial least square regression effectively predicted N contents using diffused vis–NIR spectra in the range of 500–2450 nm. This proximal sensor based precision agriculture technique could help optimize N application rates and timing in potatoes through faster, accurate, and user-friendly detection of in-season N status in the crop. When used beyond the initial 40 days after planting, it can also enhance the prediction of utilization of N from aboveground to belowground biomass and ultimately yield. While the study primarily focused on specific region and potato varieties, the model's fundamental principles make it adaptable to areas with similar potato types and conditions. Given the prevalence of Russet varieties in key potato-producing states in the USA, like Idaho and Washington, this technique holds promise for broader use. However, regional variations and potato types (such as red, white, blue/purple, and chipping) may influence model performance, necessitating further validation for new regions. Our study's use of wider spectral ranges enhances the model's adaptability, and future research can explore its fine-tuning for specific geographic needs. Overall, these results could have implications for the development of sustainable N management practices through precision in the detection of N status in the crop.