Evaluating the Potential of Improving In-Season Potato Nitrogen Status Diagnosis Using Leaf Fluorescence Sensor as Compared with SPAD Meter

Abstract

The petiole nitrate–nitrogen concentration (PNNC) has been an industry standard indicator for in-season potato (Solanum tuberosum L.) nitrogen (N) status diagnosis. Leaf sensors can be used to predict the PNNC and other N status indicators non-destructively. The SPAD meter is a common leaf chlorophyll (Chl) meter, while the Dualex is a newer leaf fluorescence sensor. Limited research has been conducted to compare the two leaf sensors for potato N status assessment. Therefore, the objectives of this study were to (1) compare SPAD and Dualex for predicting potato N status indicators, and (2) evaluate the potential prediction improvement using multi-source data fusion. The plot-scale experiments were conducted in Becker, Minnesota, USA, in 2018, 2019, 2021, and 2023, involving different cultivars, N treatments, and irrigation rates. The results indicated that Dualex’s N balance index (NBI; Chl/Flav) always outperformed Dualex Chl but did not consistently perform better than the SPAD meter. All N status indicators were predicted with significantly higher accuracy with multi-source data fusion using machine learning models. A practical strategy was developed using a linear support vector regression model with SPAD, cultivar information, accumulated growing degree days, accumulated total moisture, and an as-applied N rate to predict the vine or whole-plant N nutrition index (NNI), achieving an R² of 0.80–0.82, accuracy of 0.75–0.77, and Kappa statistic of 0.57–0.58 (near-substantial). Further research is needed to develop an easy-to-use application and corresponding in-season N recommendation strategy to facilitate practical on-farm applications.

1. Introduction

Nitrogen (N) is the most abundantly required plant nutrient, and the application of N fertilizer greatly influences the outcomes of crop production. The proper management of N fertilizer is key to achieving high crop yield and quality, while mismanagement negatively impacts not just crop yield and quality but also the environment [1]. Potato (Solanum tuberosum L.) is a shallow-rooted crop and commonly grown on irrigated coarse-textured soils for efficient tuber development, often resulting in a low N use efficiency [2,3,4]. Agronomic research has focused on improving N use efficiency and developing regional best management practices for potatoes [5]. One of the strategies is an in-season split N application based on potato N status diagnostic results using petiole nitrate analysis [6,7]. The petiole nitrate-N concentration (PNNC) reflects the effect of N fertilizer application and is often correlated with tuber yield, making it a useful tool for in-season potato N status diagnosis [8,9]. However, the analysis involves destructive sampling and on-site (i.e., petiole sap analysis) or laboratory testing (i.e., dry petiole analysis), requires a qualified technician, and takes hours to days until the result is delivered. The result may be highly variable and is not comprehensive because of the dependence on a single plant part. The whole-plant-based approach, according to the concept of a critical N concentration (N_c) and N nutrition index (NNI), provides more stable and comprehensive results [10]. The N_c is the minimum plant N concentration (PNC) required for achieving the maximum dry biomass weight, and the NNI is the ratio of measured PNC to N_c at a certain dry biomass weight [11]. The dynamic allometry of the potato crop has made the determination and interpretation of the NNI an active area of potato N research [12,13,14]. The specific in-season N application rate can be recommended based on the NNI and plant biomass [15]. Nevertheless, this whole-plant-based approach suffers from similar or even more severe problems of destructive sampling.

Recent technological and scientific progress has made proximal and remote sensing technologies a leading candidate for addressing these problems. Soil plant analysis development (SPAD; Konica Minolta, Tokyo, Japan) estimates the leaf chlorophyll (Chl) concentration based on the transmittance of red and near-infrared lights and was used in the earliest effort of proximal sensing-based maize (Zea mays L.) N status diagnosis [16]. The SPAD meter started being investigated for potato N status diagnosis around the same time and has continued to prove its usefulness in predicting the PNNC (R² = 0.85, [17]; r = 0.92, [18]; R² = 0.80, [19]), leaf N concentration (R² = 0.69, [17]; R² = 0.43–0.89, [20]; r = 0.97, [18]), and Chl concentration (r = 0.97, [18]), and by calibrating against total biomass to calculate the critical reading values (R² = 0.83, [21]), tuber yield (R² = 0.56–0.84, [17]; R² = 0.93, [22]), and N sufficiency index [23]. The challenges and limitations of the SPAD meter remain with relativity, sensitivity, and specificity [24].

Fluorescence-based sensing technologies were proposed to address these challenges and limitations in plant N status diagnosis using the SPAD meter, as well as other proximal or remote-sensing technologies [25,26]. Dualex Scientific+ (Dualex; METOS^® by Pessl Instruments, Weiz, Austria) and Multiplex (Force-A, Orsay, France) measure flavonols (Flav) and anthocyanins (Anth), N- and phosphorus-induced phenolic secondary metabolites, using one of the Chl fluorescence-sensing mechanisms involving the screening effect of these phenolic compounds. This mechanism avoids relying on a more fundamental and time-consuming Chl fluorescence mechanism called variable Chl fluorescence and enables more efficient data collection [26,27,28]. The Dualex leaf sensor and Multiplex canopy sensor demonstrated comparable performance in plant N status assessment or prediction, despite some differences in their mechanisms [29,30]. Dong et al. [31,32] found that the Dualex readings modified by days after sowing could predict leaf N concentration, PNC, and above-ground biomass for maize with R² values of 0.61–0.79, 0.62–0.83, and 0.58–0.63, respectively. Padilla et al. [33] used Multiplex to predict the NNI for cucumber (Cucumis sativus) with R² values of 0.65–0.99. Multiplex was also used to predict leaf N concentration, PNC, and NNI for rice (Oryza sativa L.) with R² values of 0.40–0.78 [34]. However, the evaluation of Dualex or Multiplex for potato N status assessment or prediction has been limited. SPAD and Dualex are both leaf-clip sensors and use a similar transmittance-based mechanism for the Chl reading. Thus, comparing SPAD and Dualex will clarify the potential benefits of fluorescence-based sensing features for potato N status prediction.

Recent research has illustrated the effectiveness of the multi-source data fusion approach through machine learning (ML) models in predicting the N status or yield of maize, rice, potato, and wheat (Triticum aestivum L.) using different proximal and remote-sensing technologies, including Dualex and Multiplex [19,30,35,36,37]. The benefits and incentives of upgrading sensors should also be evaluated, considering the improvement magnitude of adding easily available ancillary information to potato N status prediction models. Wang et al. [36] found that the differences between two-band and three-band active canopy sensors, GreenSeeker and Crop Circle ACS-430, could be significantly reduced when multi-source data were used in ML models, reducing the need to upgrade to more expensive sensors. However, similar analysis has not been reported for potato N status diagnosis using SPAD and Dualex sensors.

Therefore, the objectives of this study were to (1) determine if the Dualex sensor can perform better than the SPAD meter for predicting potato N status indicators when only sensor data are used, and (2) evaluate the potential of improving potato N status prediction using multi-source data fusion compared with only using leaf sensor data.

2. Materials and Methods

2.1. Experiment Sites

The plot-scale experiments were conducted at the Sand Plain Research Farm, Becker, Minnesota, USA, in 2018, 2019, 2021, and 2023. The farm was located at 45°23′N, 93°53′W and characterized as a Hubbard loamy sand (sandy, mixed, frigid Entic Hapludolls) until 2018 and was relocated to 45°20′N, 93°49′W in 2019 and characterized as a Hubbard (Sandy, mixed, frigid Entic Hapludolls)–Mosford (Sady, mixed, frigid Typic Hapludolls) complex sand soil. The average air temperature and total precipitation during the potato-growing season at the farm (i.e., mid-April to early-October) from 2013 to 2023 were 18.7 °C and 466.2 mm, respectively, according to the on-site weather station (Figure 1). Soil samples were collected at 0–60 cm for soil N (NO₃⁻ + NH₄⁺) and 0–15 cm for pH and other standard macro- and micro-nutrients before planting each year and analyzed at the Research Analytical Laboratory at the University of Minnesota. Due to the coarse soil texture, the organic matter content and soil N concentration were relatively low: 14.7 g kg⁻¹ and 5.86 mg kg⁻¹ on average, respectively (

Table 1. Summary of replant soil tests.

	Max	Min	Mean	Median
OM	22.0	10.0	14.7	14.0
pH	7.4	6.0	6.7	6.8
N	11.7	1.7	5.9	5.9
P	69.0	18.0	46.2	55.0
K	157.0	74.0	100.5	94.0
S	12.2	4.4	8.0	7.0
Ca	958.8	620.2	781.2	731.7
Mg	185.1	115.2	150.6	154.6
B	0.3	0.1	0.2	0.2
Fe	33.4	10.4	20.5	17.5
Mn	25.7	3.9	11.1	7.9
Zn	11.9	1.1	5.6	3.4
Cu	1.2	0.5	0.8	0.8

A total of 0–60 cm for N and 0–15 cm for other elements. OM in g kg⁻¹, pH in unitless, the rest in mg kg⁻¹.

).

2.2. Experiment Designs

The experiment in 2018 (Experiment 1) involved three N rates (i.e., 134.5, 269.0, and 403.5 kg N ha⁻¹) as the main plot treatment and six cultivars (i.e., Clearwater Russet, Ivory Russet, Lamoka, MN13142, Russet Burbank, and Umatilla Russet) as the subplot treatment in a split-plot design with three replications. The experiment in 2019 (Experiment 2) used the same design but involved five cultivars except for Ivory Russet. Russet Burbank was included twice more than other cultivars in Experiment 2.

The experiment in 2021 (Experiment 3) involved two cultivars (i.e., Hamlin Russet and Russet Burbank) as the main plot treatment and five N rates (i.e., 44.8, 89.7, 179.3, 269.0, and 358.7 kg N ha⁻¹) as the subplot treatment in a split-plot design with three replications. The experiment in 2023 (Experiment 4) involved three irrigation blocks (i.e., 60%, 80%, and 100% irrigation based on water balance). Each irrigation block included the same two cultivars as Experiment 3 (i.e., Hamlin Russet and Russet Burbank) as the main plot treatment in a split-plot design with three replications. The 60% and 80% irrigation blocks used four N treatments (i.e., 89.7, 179.3, 269.0 kg ha⁻¹ and a sensor-based precision N management treatment) as the subplot treatment, while the 100% irrigation block included nine N treatments (i.e., 44.8, 89.7, 179.3, 269.0, 358.7 kg ha⁻¹, fixed-split, and three sensor-based precision N management treatments). Leaf sensors (i.e., SPAD or Dualex) were used to make decisions on the in-season N applications in the precision N management treatments based on potato N status diagnosed through PNNC or NNI prediction or N sufficiency index calculation.

Experiments 1 and 2 and Experiments 3 and 4 were conducted for different objectives. Furthermore, Experiment 4 evolved from Experiment 3 and investigated objectives in a more integrated system. This study took advantage of the rich data from these experiments to evaluate the Dualex and SPAD sensors.

Table 2. Summary of experiment designs.

ID	Year	Plant Date	Harvest Date	Cultivars	Irrigation	N Rates (kg N/ha)
						Plant (DAP)	Emerge (ENS)	Post-Emerge (UAN)	Total
1	2018	5/14	9/25	Clearwater Russet Ivory Russet Lamoka MN13142 Russet Burbank Umatilla Russet	100%	44.8	89.7 179.3 269.0	0 11.2 * 4 22.4 * 4	134.5 269.0 403.5
2	2019	5/6	9/27	Clearwater Russet Lamoka MN13142 Russet Burbank Umatilla Russet	100%	44.8	89.7 179.3 269.0	0 11.2 * 4 22.4 * 4	134.5 269.0 403.5
3	2021	4/16	9/23	Hamlin Russet Russet Burbank	100%	44.8	0 44.8 134.5 224.2 313.8	0	44.8 89.7 179.3 269.0 358.7
4	2023	4/26	10/5	Hamlin Russet Russet Burbank	60% 80%	44.8	44.8 134.5 224.2 44.8/134.5	0 0 0 16.8 * ~4	89.7 179.3 269.0 ~156.9/246.6
					100%	44.8	0 44.8 134.5 224.2 313.8 44.8/134.5 44.8/134.5 44.8/134.5 44.8/134.5	0 0 0 0 0 16.8 * 4 16.8 * ~4 16.8 * ~4 16.8 * ~4	44.8 89.7 179.3 269.0 358.7 156.9/246.6 ~156.9/246.6 ~156.9/246.6 ~156.9/246.6

Note: all between-row spacing was 0.9 m, while the within-row spacing for Ivory Russet in 2018, Hamlin Russet in 2023, and the rest of the cultivars were 0.23, 0.25, and 0.3 m, respectively. Diammonium phosphate (DAP; 18-46-0), Environmentally Smart N (ESN; Nutrien, Canada; 44-0-0), and urea-ammonium nitrate (UAN; 28-0-0). UAN was applied immediately before irrigation as simulated fertigation every 7–14 days 4 or up to (~) 4 times (*). The two different N rates in a cell in the Emerge and Total columns are for Hamlin Russet/Russet Burbank.

summarizes the details of the experiment designs. All of the other cultural practices were implemented according to the regional recommendations [38].

2.3. Collection of Plant Samples and Sensor Data

Between late-June and early-August, corresponding to the growth stages of tuber initiation and tuber bulking, vines and tubers of three plants in each plot were sampled two to four times each year (i.e., 26 June, 10 July, 18 July, 1 August in 2018; 26 June, 11 July, 24 July, 7 August in 2019; 30 June, 28 July in 2021; 20 June, 18 July, 26 July in 2023). The vines and tubers were separated, and the whole and sub-sampled fresh weights were obtained. The sub-samples were dried in the oven at 60 °C to a constant weight and weighed again to determine percent dry matter (%DM). The dried sub-samples were ground to pass through a 2 mm sieve using a Wiley mill and analyzed for vine and tuber N concentration using an Elemental CNS analyzer (Elementar Vario EL III; Elementar Americas, NY, USA). The plant N concentration (PNC) was determined as follows:

PNC = (VNC ∗ W_v + TNC ∗ W_t)/(W_v + W_t),

(1)

where PNC is in g 100 g⁻¹, VNC is the vine N concentration in g 100 g⁻¹, TNC is the tuber N concentration in g 100 g⁻¹, W_v is the dry vine biomass (Mg DM ha⁻¹), and W_t is the dry tuber biomass (Mg DM ha⁻¹).

Before or after the whole-plant sampling campaigns, twenty petioles from the fourth leaf from the shoot apex were sampled in each plot. The petioles were dried, ground, and analyzed for nitrate-N concentration using water extraction and conductimetric procedures [39]. The SPAD and Dualex data were also collected on the same day as petiole sampling or as close as possible. Twenty or thirty SPAD readings were taken on the fourth leaf from the shoot apex and manually averaged for each plot. Fifteen Dualex readings were taken on the top fully expanded leaves, and Dualex provided the average Chl, Flav, Anth, and N balance index (NBI) values, where NBI is the Chl/Flav ratio [27].

2.4. Data Wrangling

Plant nitrogen uptake (PNU) and the NNI were used as N status indicators along with PNNC, VNC, and PNC. PNU was calculated as follows:

PNU = 10 ∗ PNC ∗ (W_v + W_t),

(2)

where PNU is in kg ha⁻¹. The critical N dilution curves define the relationship between N_c and plant dry biomass (W) using an allometric negative power function as follows:

N_c = aW^−b,

(3)

where N_c is in g 100 g⁻¹ and W is in Mg DM ha⁻¹, and a and b are the empirical parameters. Parameter a is numerically equivalent to the N_c concentration at W = 1 Mg DM ha⁻¹, and Parameter b is the dimensionless dilution parameter defining the rate of N_c decline with an increase in W. When W was less than 1 Mg DM ha⁻¹, the value of Parameter a was always used as N_c, assuming a constant total N concentration [11]. The vine and whole-plant critical N dilution curve coefficients were derived from Giletto et al. [14]. The parameters for Russet Burbank and Umatilla Russet were directly available. The parameters for Russet Burbank, Shepody, and Umatilla Russet were used for Hamlin Russet; Ivory Russet, MN13142; and Clearwater Russet, Lamoka, according to the maturity class. The parameters are summarized in

Table 3. Critical N dilution curve parameters for vine and whole-plant.

Cultivar	Vine a	Vine b	WP a	WP b
Russet Burbank Hamlin Russet	5.08	0.28	4.57	0.42
Umatilla Russet Clearwater Russet Lamoka	5.44	0.27	5.04	0.42
Ivory Russet MN13142	5.17	0.18	5.19	0.25

a and b are the empirical parameters in the N_c definition.

.

The NNI was calculated as follows:

NNI = PNC/N_c,

(4)

where PNC and N_c are in g 100 g⁻¹, PNC is measured on the plant of interest (i.e., vines for Vine NNI, and vines + tubers for whole-plant NNI), and N_c is derived from the critical N dilution curve based on the dry biomass weight of the plant.

The cultivar information was organized categorically using the cultivar names, which were coded using dummy variables as needed. The beginning-of-the-season (initial) soil samples were collected on a replication basis. When the soil test results were collected only from a subset of replications, the rest of the replications were imputed with the average soil test result values within each site-year. Daily weather information was recorded by the on-site weather station. Air temperature data were used to calculate accumulated growing degree days (GDDs) as follows:

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{G}\mathrm{D}\mathrm{D}\mathrm{s}=\sum {\displaystyle \frac{\left({\mathrm{T}}_{\mathrm{m}\mathrm{a}\mathrm{x}}+{\mathrm{T}}_{\mathrm{m}\mathrm{i}\mathrm{n}}\right)}{2}}-7,$$

(5)

where T_max and T_min are daily maximum and minimum air temperatures in °C, and 7 °C is the base air temperature for potatoes [40]. The GDDs were summed up from the planting date to each sampling/sensing date.

Precipitation data were used with the irrigation log to calculate accumulated total moisture as follows:

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{m}\mathrm{o}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{e}=\sum (\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{p}\mathrm{i}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{I}\mathrm{r}\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}),$$

(6)

where precipitation and irrigation are in mm. Irrigation was scheduled using the checkbook method and applied a few times a week [41]. In Experiment 4, irrigation was reduced in the 60% and 80% treatments proportionately, except before 29 May and on 2 August, because of too much drought pressure or mechanical issues with the irrigator. The summation was applied in the same way as accumulated GDDs.

The as-applied N rate was calculated by summing up the amount of N that had been applied until each sampling/sensing date. When slow-release Environmentally Smart N (ESN) fertilizer was applied at emergence, the N release rate of the fertilizer was considered based on the work by Wilson et al. [42] as follows:

Percent N release = −0.008 DAS² + 2.0 DAS − 37.8,

(7)

where DAS is days after sowing. When other N fertilizer types were used, all of the N credit was added to as-applied N rate at once on the following sampling/sensing dates. All of the N status indicators and genetic, environmental, and management variables were combined with leaf sensor data. After handling missing values, the dataset amounted to 656 observations.

2.5. Statistical Analysis

Regression models with varying complexities were used to compare the two leaf sensors, including simple regression (SR) models, multiple linear regression (MLR) or the least absolute shrinkage and selection operator (LASSO) regression model, the random forest regression (RFR) model, extreme gradient boosting (XGBoost), and support vector regression (SVR) models. Two different scenarios were considered: (1) only the leaf sensor data were available, and (2) the leaf sensor data and the available genetic, environmental, and management data. The SR and MLR models were originally fitted using the whole dataset to select the best-performing models based on the coefficient of determination (R²) values. The best-performing models were trained and tested using a 4-fold cross-validation, where each fold holds data from a different site-year. The other models with hyperparameters were trained and tested using nested cross-validation with 4 outer folds and 10 inner folds. The 4 outer folds were the same as those in a 4-fold cross-validation, whereas the 10 inner folds were created randomly. This data partitioning design aimed to realize more robust model evaluation. The variabilities of the N status indicators were considered when determining which metric to use for the hyperparameter tuning of the ML models in Bayesian optimization. Meanwhile, using dummy variables for nominal values (i.e., cultivar information) in the geometric models necessitated all levels of nominal values to be present in the training dataset. Ivory Russet, Lamoka, and MN13142 were, therefore, removed from the dataset, resulting in 568 observations. Important features were selected using the coefficients of the LASSO regression models and the permutation-based importance analysis using the random forest called Boruta 8.0.0 for Scenario 2 [43].

Model development was conducted using an R framework, Tidymodels 1.2.0 [44]. Bayesian optimization was initialized with 10 sets of random hyperparameters and iterated up to 50 times for hyperparameter tuning using the expected improvement with a trade-off value of 0.1 as an acquisition function. The “glmnet 4.1.8”, “ranger 0.16.0”, “xgboost 1.7.8.1”, and “kernlab 0.9.32” packages were used as engines for LASSO regression, RFR, and SVR in Tidymodels [45,46,47,48]. The following hyperparameters were tuned: L-1 regularization term for LASSO regression; the number of predictor variables randomly selected at each node (mtry), the number of trees (trees), and the minimum node size (min_n) for RFR; mtry, trees, min_n, learning rate, the proportion of observations sampled for growing each tree, and L-2 regularization term; cost, margin, degree, scale, offset, and sigma for SVR.

R², mean absolute error (MAE), and root mean square error (RMSE) were used for model evaluation:

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}{{\sum _{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}^2}},$$

(8)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right|,$$

(9)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2},$$

(10)

where n is the number of observations, y_i is the actual value of the ith observation, $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ is the predicted value of the ith observation, and $\overline{\mathrm{y}}$ is the mean of all the observations. The PNNC values were tentatively assigned deficient, sufficient, or excessive categories using the sufficiency thresholds established for Russet Burbank by Rosen and Bierman [5]: 17,000–22,000 mg kg⁻¹ for 15–30 June, 11,000–15,000 mg kg⁻¹ for 1–15 July, and 6000–9000 mg kg⁻¹ for 15 July–15 August. The NNI values were also categorized using the sufficiency threshold of 0.95–1.05. Accuracy and Kappa statistics were used to evaluate diagnostic capability as follows:

Accuracy = (TP + TN)/(TP + TN + FP + FN),

(11)

Kappa statistic = (P_o − P_e)/(1 − P_e),

(12)

where TP, TN, FP, and FN are true positive, true negative, false positive, and false negative, and P_o and P_e are the probability of agreement observed and probability of agreement by chance, respectively. Kappa statistics measure the level of agreement between observed and predicted categories while considering the chance agreement. The Kappa values of < 0, 0–0.2, 0.21–0.40, 0.41–0.60, 0.61–0.80, and 0.81–1.00 corresponds with poor, slight, fair, moderate, substantial, and almost perfect levels of agreement [49].

Shapley additive explanation (SHAP) was used to interpret the contribution of each feature to an output with respect to the expected output in the developed models [50]. One hundred sub-samples were randomly selected to determine SHAP values for each observation using the iml 0.11.3 package [51]. SHAP values were visualized using shapviz 0.9.6 [52]. All of the other statistical analyses and visualizations were also conducted in R 4.4.1. [53].

3. Results

The summary statistics of the N status indicators indicated that the PNNC was more variable than the other N status indicators (

Table 4. Summary statistics of N status indicators.

	PNNC	VNC	PNC	PNU	Vine NNI	NNI
	(mg kg−1)	(g 100 g−1)	(g 100 g−1)	(kg ha−1)
Min	5	1.02	0.87	41.05	0.29	0.25
Mean	10,896	3.77	3.16	155.02	0.94	0.96
Median	9984	3.65	2.73	143.84	0.96	0.93
Max	31,410	7.22	7.12	405.37	1.65	2.11
SD	7898	1.28	1.4	59.37	0.28	0.37
CV	1	0.34	0.44	0.38	0.3	0.39

). The NNI had median and mean values close to 1 with 0.3–0.4 standard deviations, allowing for the evaluation of the diagnostic accuracy across all three N status categories. As for the metrics for hyperparameter tuning, the MAE was selected for the PNNC prediction models to minimize the potential outlier effect, while the RMSE was selected for the rest of the prediction models.

3.1. Scenario 1: Leaf Sensor Data Only

The SR models were applied with or without axis log-transformation (i.e., linear, logarithmic, power, exponential, and quadratic). Across the N status indicators, the SPAD and Dualex NBI always had the highest R² values using the power or quadratic forms (

Table A1. The relationship between SPAD or Dualex and different N status indicators.

Sensor	Type	Equation	R2
PNNC
SPAD	power	y = 4.32 × 10−10 x8.11	0.55
DuxChl	power	y = 8.52 × 10−5 x5.46	0.38
DuxFlav	quadratic	y = −10,342.21 x2 − 129,710.52 x + 10,896.26	0.41
DuxAnth	quadratic	y = −29,994.25 x2 − 73,319.01 x + 10,896.26	0.15
DuxNBI	quadratic	y = −6568.29 x2 − 149,803.83 x + 10,896.26	0.55
VNC
SPAD	power	y = 4.15 × 10−3 x1.81	0.48
DuxChl	power	y = 7.65 × 10−2 x1.16	0.31
DuxFlav	quadratic	y = −0.47 x2 − 23.40 x + 3.77	0.51
DuxAnth	quadratic	y = −3.29 x2 − 13.68 x + 3.77	0.19
DuxNBI	quadratic	y = −0.08 x2 + 25.33 x + 3.77	0.60
WPNC
SPAD	power	y = 5.84 × 10−4 x2.28	0.52
DuxChl	power	y = 1.98 × 10−2 x1.50	0.35
DuxFlav	exponential	y = 21.26 e−1.34 x	0.45
DuxAnth	quadratic	y = −5.52 x2 − 13.24 x + 3.16	0.16
DuxNBI	quadratic	y = 3.37 x2 + 27.71 x + 3.16	0.61
PNU
SPAD	quadratic	y = −488.13 x2 − 247.01 x + 155.02	0.11
DuxChl	quadratic	y = −127.65 x2 − 71.00 x + 155.02	0.01
DuxFlav	quadratic	y = −213.96 x2 + 80.41 x + 155.02	0.02
DuxAnth	quadratic	y = 306.50 x2 + 76.48 x + 155.02	0.04
DuxNBI	quadratic	y = −411.87 x2 − 154.86 x + 155.02	0.08
Vine NNI
SPAD	power	y = 1.89 × 10−3 x1.65	0.49
DuxChl	quadratic	y = −2.00 x2 + 3.39 x + 0.941	0.31
DuxFlav	quadratic	y = −0.791 x2 − 5.20 x + 0.941	0.55
DuxAnth	quadratic	y = −1.06 x2 − 1.83 x + 0.941	0.09
DuxNBI	power	y = 5.24 × 10−2 x0.967	0.60
NNI
SPAD	power	y = 3.21 × 10−4 x2.12	0.52
DuxChl	power	y = 8.95 × 10−3 x1.39	0.34
DuxFlav	exp	y = 5.42 e−1.21 x	0.42
DuxAnth	quadratic	y = −2.43 x2 − 1.18 x + 0.96	0.08
DuxNBI	power	y = 2.99 × 10−2 x1.16	0.54

). The R² values were up to 0.12 higher for the Dualex NBI than the SPAD in the PNNC, VNC, PNC, Vine NNI, and NNI prediction models. The R² values for the PNU prediction models were much lower for both SPAD and Dualex NBI (i.e., 0.11 and 0.08). Figure 2 shows the results of 4-fold cross-validation, and the scatter plots visualize the prediction results in the testing datasets. The performance difference between the SPAD and Dualex NBI in every N status indicator prediction was negligible. Concisely, the PNNC, VNC, PNC, Vine NNI, and NNI prediction models all achieved an R² value of approximately 0.6, while the PNU models remained to have a low R² value of approximately 0.15. The systematic under- and over-estimation at low and high values with the power regression models suggest insufficient explanatory variables or a lack of model flexibility.

The multi-parametric functionality of Dualex enables the development of MLR models. The derivation of the NBI (i.e., Chl/Flav) led to extremely high variance inflation factor values, justifying the omission of the NBI in the MLR model development. The MLR models did not demonstrate any improvements over the SR models. Excluding Chl and Flav instead of the NBI produced similar results (i.e., ±0.002–0.026 in adjusted R²). The advanced ML models (i.e., RFR, XGBoost, and SVR) were also fitted using all of the Dualex parameters.

Table 5. The summary of advanced ML model performance metrics using all Dualex parameters.

N Indicator	Dataset	Model	R2	MAE	RMSE	Acc	Kappa
PNNC	Training	RFR	0.94	1600.58	2052.82	0.77	0.65
	Testing	RFR	0.66	3898.12	4864.5	0.56	0.32
VNC	Training	SVR L	0.65	0.6	0.75	-	-
	Testing	SVR L	0.57	0.69	0.88	-	-
PNC	Training	RFR	0.95	0.24	0.32	-	-
	Testing	RFR	0.62	0.72	1	-	-
PNU	Training	SVR L	0.09	44.11	56.31	-	-
	Testing	SVR L	0.11	54.14	69.72	-	-
Vine NNI	Training	SVR L	0.62	0.14	0.17	0.71	0.51
	Testing	SVR L	0.55	0.16	0.2	0.69	0.45
NNI	Training	SVR P	0.53	0.2	0.26	0.7	0.5
	Testing	SVR P	0.54	0.26	0.32	0.64	0.4

summarizes the performance metric values of the best ML model for each N status indicator prediction in the training and testing datasets. Despite the ability to characterize complex non-linear relationships, little improvement was found. The PNNC and NNI predictions were marginally improved using the RFR and polynomial SVR models (i.e., testing R² = 0.66 and 0.60, respectively).

The within-year performance of the two leaf sensors was additionally evaluated, as the between-year variability of environmental factors might have obscured the advantages of the fluorescence sensor.

Table 6. The R² values of the SR models fitted using SPAD, Dualex Chl, or Dualex NBI in each year.

Year	N indicator	SPAD	DuxChl	DuxNBI
2018	PNNC	0.66	0.60	0.69
	VNC	0.57	0.58	0.74
	PNC	0.53	0.55	0.72
	PNU	0.06	0.06	0.15
	Vine NNI	0.48	0.51	0.69
	NNI	0.40	0.43	0.61
2019	PNNC	0.43	0.47	0.53
	VNC	0.38	0.71	0.76
	PNC	0.38	0.71	0.76
	PNU	0.04	0.29	0.22
	Vine NNI	0.39	0.37	0.53
	NNI	0.43	0.48	0.66
2021	PNNC	0.70	0.79	0.84
	VNC	0.87	0.86	0.67
	PNC	0.90	0.87	0.62
	PNU	0.14	0.12	0.23
	Vine NNI	0.83	0.86	0.73
	NNI	0.83	0.84	0.69
2023	PNNC	0.71	0.34	0.50
	VNC	0.69	0.33	0.47
	PNC	0.74	0.30	0.46
	PNU	0.34	0.03	0.07
	Vine NNI	0.58	0.33	0.45
	NNI	0.65	0.34	0.45

shows the highest R² values of the SR models fitted using SPAD, Dualex Chl, or Dualex NBI in each year. Dualex Chl had higher R² values than SPAD in 2019, whereas SPAD had higher R² values than Dualex Chl in 2023. Dualex NBI had higher R² values than Dualex Chl in most cases. However, the improvement of the Dualex NBI over Dualex Chl in N status indicator prediction within each growing season was not greater than across all the growing seasons (Table A1 and Table 6).

3.2. Scenario 2: Multi-Source Data Fusion

The SPAD or Dualex data were used with the genetic, environmental, and management data to improve the N status indicator prediction. LASSO regression was used due to the extremely high variance inflation factor values of some variables (e.g., initial soil test results). The NBI was removed because the regularization parameter tuning was negatively affected. The model performance greatly improved the testing R² values for all of the N status indicator predictions (i.e., testing R² = 0.69 − 0.85 for PNNC, VNC, PNC, Vine NNI, and NNI; testing R² = 0.49 − 0.5 for PNU; Figure 3). However, some of the LASSO regression models (e.g., PNC, PNU, and Vine NNI prediction) had greater degrees of deviation from the 1:1 relationship, likely due to overfitting.

Figure 4 illustrates the LASSO regression coefficients for all of the eight regression models in Figure 3. The absolute values of the coefficients were used, separating the positive and negative coefficients with colors. Figure 4a,b were on a logarithmic scale, and the coefficient values less than 1 were replaced with 0 for practical and visualization purposes. The intercept values were also excluded. The information provided by SPAD and Dualex Chl was neither outstanding nor consistent. The most important features for predicting different N status indicators were accumulated GDDs and the as-applied N rate, supporting the effectiveness of the data fusion approach. Dualex Flav and Anth also provided useful information, although to a lesser extent. Soil properties and nutrients that influence plant PNU or protein synthesis were other informative features (e.g., organic matter, S, Mg, Zn). Evaluating the importance of cultivar information in the same fashion is not appropriate because cultivar information was incorporated using the dummy variable technique. The positive and negative correlation between the N status indicators and predictor variables was reasonable.

The Boruta results were pooled across all of the eight RFR models and visualized in Figure 5. Accumulated GDDs and the as-applied N rate remained the most important features. In contrast with the feature importance analysis results based on the LASSO regression coefficients, accumulated total moisture demonstrated similar importance to accumulated GDDs. All of the leaf sensor data, particularly SPAD and the Dualex NBI, were the other most informative features. The Z-score of cultivar information was generally high but variable, reflecting the inconsistency of its contribution depending on the N status indicators. In light of the feature importance analyses and general data accessibility, the following sophisticated ML models were developed using leaf sensor data (i.e., SPAD or Dualex Chl, Flav, Anth, and NBI) and the four types of auxiliary information (i.e., cultivar information, accumulated GDDs, accumulated total moisture, and as-applied N rate).

Table 7. The summary of sophisticated machine learning model performance metrics using leaf sensor data and auxiliary information.

(a) SPAD Meter
N Indicator	Dataset	Model	R2	MAE	RMSE	Acc	Kappa
PNNC	Training	SVR L	0.81	2625.87	3513.96	0.71	0.56
	Testing	SVR L	0.79	4189.66	5285.45	0.64	0.42
VNC	Training	SVR L	0.84	0.39	0.50	-	-
	Testing	SVR L	0.85	0.56	0.68	-	-
PNC	Training	SVR R	0.94	0.25	0.34	-	-
	Testing	SVR R	0.90	0.40	0.50	-	-
PNU	Training	SVR L	0.62	26.43	36.54	-	-
	Testing	SVR L	0.55	34.57	45.39	-	-
Vine NNI	Training	SVR L	0.80	0.09	0.12	0.79	0.65
	Testing	SVR L	0.80	0.11	0.14	0.75	0.57
NNI	Training	SVR L	0.81	0.12	0.16	0.82	0.68
	Testing	SVR L	0.82	0.16	0.20	0.77	0.58
(b) Dualex Sensor
PNNC	Training	RFR	0.99	653.64	891.39	0.91	0.86
	Testing	RFR	0.75	3399.62	4266.46	0.63	0.43
VNC	Training	SVR L	0.87	0.36	0.46	-	-
	Testing	SVR L	0.85	0.51	0.63	-	-
PNC	Training	SVR L	0.90	0.35	0.45	-	-
	Testing	SVR L	0.87	0.47	0.58	-	-
PNU	Training	SVR L	0.64	25.70	35.32	-	-
	Testing	SVR L	0.57	32.74	43.21	-	-
Vine NNI	Training	SVR L	0.81	0.09	0.12	0.80	0.65
	Testing	SVR L	0.80	0.12	0.15	0.75	0.57
NNI	Training	SVR L	0.83	0.11	0.15	0.84	0.71
	Testing	SVR L	0.81	0.17	0.22	0.75	0.54

summarizes the performance metrics of the best ML model for each N status indicator prediction using SPAD or Dualex data, with the auxiliary information in the training and testing datasets. Linear SVR models presented the best performance metrics in most cases with smaller testing MAE and RMSE values than the LASSO regression models, demonstrating better ability to balance variance and bias. The leaf sensor types did not make much difference in the performance of these N status indicator prediction models. The testing R² values of the VNC and PNC models were very high (i.e., 0.85–0.90). Both the Vine NNI and NNI demonstrated comparable R² values and higher diagnostic accuracy than the PNNC in the testing dataset (i.e., 0.75–0.80 in R², 0.63–0.64 vs. 0.75–0.77 in accuracy, and 0.42–0.43 vs. 0.54–0.58 in a Kappa statistic).

Figure 6 shows the results of SHAP analysis for the best Vine NNI and NNI prediction models using SPAD or Dualex in the beeswarm plots. The features were ordered by the level of contribution from top to bottom. The distribution of SHAP values and the range of feature values were also visualized in the plots. Regardless of the leaf sensor types, accumulated GDDs, the as-applied N rate, and accumulated total moisture were the top three contributing features, reaffirming the effectiveness of the data fusion approach. Dualex Flav and the NBI appeared to be slightly more conducive for predicting the NNI than SPAD.

4. Discussion

4.1. Comparing the Ability of SPAD and Dualex to Predict Potato N Status Indicators

An improvement in the ability to predict potato N status indicators using Dualex over SPAD across years was not apparent in this study, despite additional functionality provided by the Dualex sensor (Figure 2 and Table 5). Padilla et al. [33] found that the leaf Chl reading was more useful on a standardized growth stage basis, while Flav or the NBI was more useful within each growing season because of its sensitivity to variable environmental factors, including solar radiation and air temperature. If the Dualex NBI is superior to SPAD within each growing season, the N sufficiency index approach can utilize this advantage. The slow-release N fertilizer may have been released more quickly and potentially leached more by higher average air temperature and early-season precipitation in 2018 than 2019, inducing a different plant N status in response to similar management (Figure 1 and Table 4) [42]. Because SPAD and Dualex Chl readings have exponential and linear relationships with the Chl concentration, relatively higher sensitivity to lower and higher ranges of Chl concentration can be expected and, thus, might explain the lower and higher R² values of SPAD and Dualex Chl in 2019, respectively [27]. Meanwhile, the same rationale does not hold for the comparison between 2021 and 2023. SPAD and Dualex use red and red-edge spectral bands for the transmittance-based mechanism to obtain the Chl reading [27]. Due to the combined spectral characteristics of red and near-infrared, the red-edge band is more influenced by leaf structure (e.g., mesophyll) than the red band. Different types of plant stress (e.g., N, heat, water, pest) can alter the leaf structure, potentially making the Dualex Chl reading less N specific in some conditions than SPAD.

The Dualex NBI demonstrated improvements over Dualex Chl, while the degree of improvement was influenced by the N stress levels introduced by the experimental treatments (e.g., cultivars with different N use efficiency, N fertilizer application rate, and method). The benefit of using the Dualex NBI within a single year over multiple years was not observed in this study. Neither was there enough evidence to claim that the Dualex NBI outperformed SPAD within each growing season. Ultimately, rather than demonstrating the superiority of one leaf sensor over the other, the results revealed the unsatisfactory and inconsistent performance of both leaf sensors when only sensor data were used—potentially caused by low sensitivity or specificity—as previously summarized for SPAD by Goffart et al. [24]. Such weaknesses of proximal and remote-sensing technologies can be overcome using the multi-source data fusion approach [35,36,37,54].

4.2. Improving Potato N Status Indicator Prediction Using Multi-Source Data Fusion

Quantifying environmental and management factors is important to predict N status indicators because fertilization, irrigation, and environmental conditions largely influence the demand and accessibility of N to plants [1]. Previous research has also found that the N fertilizer application rate helped greatly improve the plant N status or yield prediction for corn, rice, and winter wheat, while parameters related to air temperature, precipitation, and irrigation had lesser and varying levels of contribution [15,30,36]. Air temperature information was parameterized using GDDs for simplicity in this study and was considered most conducive in Figure 6, but other types of parameters such as day–night temperature difference might provide additional insight because of the effects on the allometric dynamics of above and below-ground biomass [55]. Other potentially useful environmental and management information includes solar radiation and planting density [13]. One of the roles Flav and Anth play in plants is photoprotection, creating a high positive correlation between Dualex Flav/Anth and solar radiation and possibly accounting for some of the importance of Dualex Flav/Anth observed in this study [33]. The minimal differences in planting density in our experiments between and within cultivars did not justify using planting density as one of the input variables. However, planting density will be more useful when data are pooled across various potato production systems, which use different planting densities between and within cultivars. The genetic information is the other important factor to be considered because the allometric dynamics also greatly vary among varieties [13].

Both LASSO and linear SVR are regularized linear models. LASSO minimizes the ordinary least squares loss with an L1 regularization term, while linear SVR uses support vectors with the epsilon-insensitive loss function and an L2 regularization term. The linear SVR model was superior to the LASSO model because of its better ability to generalize and handle multicollinearity through the more elaborate design (e.g., loss function, error margin, sparsity). Nevertheless, the linear relationships between the selected features and N status indicators were found, coinciding with our previous findings [19]. As a result, the linear SVR models accurately and computationally efficiently combined leaf sensor data with the auxiliary information in predicting most of the N status indicators. It is worth noting that the relationships may become slightly less linear as more input features are added, favoring RF and XGBoost models over linear SVR [19].

4.3. Implications for In-Season Potato N Status Diagnosis

PNNC has been used as an industry standard N status indicator for in-season potato N status diagnosis and is therefore used as a reference for comparison here [5]. The effectiveness of VNC and PNC in potato N status diagnosis must be investigated in comparison with PNNC, and if validated, the determination of VNC and PNC sufficiency ranges is also necessary to be used in place of PNNC. The decreasing trend of VNC and PNC sufficiency ranges across growth stages must be characterized with reference to yield or biomass. The PNU (i.e., PNC × biomass) was predicted less accurately, despite the accurate PNC prediction, indicating the difficulty in predicting biomass using a leaf sensor. The biomass has been more successfully predicted using canopy sensors for rice and winter wheat [15,56,57]. For potato crops, below-ground biomass (e.g., tubers) makes the prediction of whole-plant biomass using canopy sensors more challenging. The improved diagnostic accuracy of the Vine NNI and NNI over PNNC may be attributed to the less variable and more robust nature of the NNI resulting from derivation based on more holistic (e.g., vines) or whole-plant parts. The Vine NNI and NNI prediction and diagnostic accuracy were comparable, despite the difficulty of sensing below-ground biomass, as explained above. Leaf sensor data and plant samples were collected mostly between 60 and 90 days after planting, during which tuber growth is still minimal, warranting the high performance of Vine NNI [14].

The NNI calculation (4) can be modified using the biomass:

$$\mathrm{N}\mathrm{N}\mathrm{I}={\displaystyle \frac{PNC\times AGB}{N_c\times AGB}}={\displaystyle \frac{\mathrm{P}\mathrm{U}\mathrm{N}}{{\mathrm{P}\mathrm{N}\mathrm{U}}_{\mathrm{c}}}}={\displaystyle \frac{{\mathrm{P}\mathrm{N}\mathrm{U}}_{\mathrm{c}}-\Delta \mathrm{P}\mathrm{N}\mathrm{U}}{{\mathrm{P}\mathrm{N}\mathrm{U}}_{\mathrm{c}}}},$$

(13)

where PNU_c is the critical PNU and ΔPNU is the difference between the PNU_c and PNU at the particular above-ground biomass (AGB) in Mg DM ha⁻¹. This modification clarifies the potential to implement variable rate N application according to the NNI, where ΔPNU can be converted to the N fertilizer application rate using the estimated N recovery rate. The capability to implement variable rate N application favors the NNI calculation through predicted PNU and biomass, as long as the prediction of PNU and biomass is equal to or more accurate than the direct prediction of the NNI. Our results did not support this approach this time, and, even if both PNC and PNU were predicted as accurately as PNNC and the NNI, their cumulative error effect in the NNI calculation process must also be carefully evaluated. Therefore, the best strategy for in-season potato N status diagnosis using a leaf sensor is to directly predict the NNI using the data fusion approach. It is important to highlight the priority of the data fusion approach over the leaf sensor upgrade for accurate diagnosis of the potato N status, as shown in our findings. The careful selection of easily accessible auxiliary data makes this approach practical. Vegetation indices (e.g., Normalized Difference Vegetation Index, Normalized Difference Red-Edge Index) calculated using canopy sensors may predict biomass or PNU more accurately, potentially enabling the indirect NNI approach and variable rate N application. The sensor fusion approach (e.g., leaf and canopy sensor) could also further improve the accuracy and scalability of NNI prediction.

This study used the N_c dilution curves developed by Giletto et al. [14] because they assessed a number of cultivars and both Vine and whole-plant NNI. Nevertheless, their study was conducted in Canada and Argentina. The exact cultivar matches only included Russet Burbank and Umatilla Russet, resorting to substitution for other cultivars based on maturity class. The effect of the reduced irrigation treatments in the 2023 experiment was not taken into account when using the NNI framework either. Bohman et al. [13] found the significant genetic environment management effect on the determination of N_c dilution curves. The robustness of the NNI framework across major environmental factors was reported by Gastal and Lemaire [58], but the contrasting effects of water stress on N_c have also been reported [59,60].

The NNI sufficiency thresholds have been commonly set at 0.95–1.05, but this convention also requires reconsideration. The 95% confidence interval of the posterior distribution of N_c was suggested to be used directly or parametrically [13]. Another potential approach is to adjust the sufficiency threshold for each genetic environment management condition empirically. These limitations do not affect the findings of this study, but have implications for decision-making based on the predicted NNI values.

Finally, the use of the developed ML model in the field can be facilitated through the development of an application (App), which should access publicly available weather information via user location and weather station application program interface (API) and asks for a minimum set of data from the user including SPAD meter data. More studies are also needed to develop effective and practical in-season site-specific N recommendation strategies based on the sensor-predicted N status indicators to support potato growers in improving their N management.

5. Conclusions

The SPAD meter and Dualex sensor predicted various in-season potato N status indicators with similar accuracy across site-years, cultivars, and N rates, although one sensor might perform better than the other under a specific condition. The multi-parametric functionality of Dualex improved prediction over its single parameters but not consistently over the SPAD meter, regardless of model complexity. The multi-source data fusion approach using a leaf sensor and genetic, environmental, and management data in sophisticated ML models significantly improved prediction and diagnosis accuracy compared to using either leaf sensor alone. The linear SVR model demonstrated the most consistent and accurate performance using leaf sensor data, cultivar information, accumulated GDDs, accumulated total moisture, and the as-applied N rate. Directly predicting the NNI using the linear SVR model presented the highest prediction and diagnosis accuracy in the testing datasets with R² of 0.80–0.82, accuracy of 0.75–0.77, and Kappa statistic of 0.54–0.58 (near-substantial). The leaf sensors did not perform well for predicting biomass or PNU. Further research is needed to develop an application to facilitate practical in-field applications and develop effective and practical in-season site-specific N recommendation strategies using the sensor-predicted N status indicators to support potato growers in improving their N management.