a. Reference data of chlorophyll and macronutrient content
Different P treatments caused major variations in pigment content in all the studied plants. Figure 1 illustrates the contents of Chlorophyll a, Chlorophyll b, Total chlorophyll and Carotenoids in sugar beet, celery and strawberry plants in response to different P fertilizations. The results show that, in the case of sugar beet and strawberry plants, exceeding the recommended dose of phosphorus in the nutrient solution (yellow bars in Fig. 1 depicting a 33% increase of P) caused a decrease in the chlorophyll content, which refers to all the measured kinds of chlorophyll (Chlorophyll a, Chlorophyll b and Total chlorophyll). The same trend was observed in celery plants, but to a smaller extent. For the three studied species, the maximum chlorophyll concentrations were recorded for various P doses: for sugar beet, the dose was under 33% of the recommended dose, for celery, under 67% of the recommended dose and, for strawberry, under the recommended dose. These differences speak to the varying impacts of P on the chlorophyll activity of various species. The deficiency or excessive application of P into the growing pots caused very high decreases in chlorophyll and carotenoid concentrations only in strawberry leaves compared to the control group (those with the recommended dose), which can be related to leaf chlorosis. This result indicates a high sensitivity of strawberry plants to imbalanced phosphorus dosing, in agreement with observations of Trejo-Téllez and Gómez-Merino [43], who noticed a considerable decrease of chlorophyll content in P-deficient strawberry leaves, which became uniformly yellow under P stress. Additionally, Estrada-Ortiz et al. [44] confirmed a strong relationship between P content in strawberry plants and the accumulation chlorophylls in its leaves. Moreover, these authors indicated that excess P application causes a decrease in the contents of photosynthetic pigments and also influences serious soil and environmental degradation. Fig. 1 shows that the concentrations of chlorophyll a in the studied species were much higher than the concentrations of chlorophyll b. This result is not in agreement with Costa et al. [45], who observed that Chlorophyll b concentrations were higher than Chlorophyll a in 2 cultivars of strawberry under varied lightening conditions. However, it was previously indicated that the relationship between chlorophyll a and b depends on many factors, including source of light, shading, ambient conditions of plant growth [46-48] and the specific role of these two pigments in plant physico-chemistry. Chlorophyll a is responsible for the collection of photons and plays an essential role in photosynthesis, while chlorophyll b additionally participates in the transference of light radioactive energy [49, 50].
In celery leaves, the chlorophyll concentration (for all 3 types of chlorophyll) at the lowest P treatment (P-33) was much lower than in other P treatments, indicating that such low P supply has a stress effect on celery. For celery leaves, the highest values of chlorophyll concentration occurred for the variant P-67, not P-100, which speaks to the overestimation of the recommended P dose in the fertilizer in this case. The same was noticed for the carotenoid content in celery leaves (the highest value was noticed for the P-67 variant). It was also observed in celery plants that the highest dose of P in fertilizer (P-133 variant) did not lead to such high decreases in chlorophyll a, b or total concentrations, as was the case for the lowest fertilizer dose (variant P-33), which suggests that celery is more sensitive to the scarcity of P than to its excess.
The total N, P, K, Mg and Ca contents, measured by reference methods at 49, 51 and 45 DAT for celery, sugar beet and strawberry plants, are shown in Figure 2. Generally, considerable and statistically significant differences in macronutrient contents were observed between various P treatments in the studied species. However, neither foliar P concentrations in the sugar beet and celery plants nor Mg concentrations in celery and strawberry were significantly affected by P treatment. It was difficult to find strict tendency in macronutrient content changes in the leaves of the three studied species with changing P fertilization. For example, in sugar beet and celery, the lowest dose of P in fertilizer (P-33) led to the highest values of N, which could be due to specific interactions between nutrient elements in the substrate, as explained in research conducted by Y. Li et al. [3]. Similarly, the lowest dose of P in fertilizer (P-33) was reflected in the highest concentrations of K for each of the species. The highest contents of Ca were observed in celery leaves; however, increasing trends with rising P concentrations in the treatments were not confirmed in sugar beet or strawberry. These results confirm the complicated relationships between the contents of macronutrients in leaves and P treatments in the soil, which was also suggested in other sources [3, 51, 52].
b. Correlation between leaf biochemical constituents, phosphorus fertilization and mass of the leaf/roots of the plants
Pearson correlation coefficients (PCC) between the leaf macronutrients (N, P, K, Ca, Mg), total chlorophyll (Chltot), carotenoids (Car), phosphorus fertilization level (Psuppl), mass of the leaf (mleaf) and mass of root (mroot) for the studied species are presented in Fig. 3 as correlation matrices. The leaf pigments and nutritional elements were evaluated through laboratory analysis at the end of the experiment. All plants showed a negative correlation between the level of P supply and the concentration of nitrogen (N) and potassium (K) macronutrients. The results obtained for celery showed a strong positive correlation (PCC=0.77) between the applied dosed of phosphorus fertilizer and the calcium content in plant leaves, whereas the other plants indicated a negative correlation (PCC=-0.81 for sugar beet and PCC=-0.69 for strawberry). Earlier reports also indicated a strong phosphorus fertilization effect on other macronutrient accumulation in plants [53, 54].
A negative and highly significant correlation was observed between the level of P supplementation and chlorophyll concentration in sugar beet, while the other plants showed non-significant correlations. The applied fertilization had no strong effect on the concentrations of carotenoids in plant leaves. Phosphorus fertilization was positively correlated with the concentration of this element in plant leaves, especially celery (PCC=0.70) and strawberry plants (PCC=0.70). Sugar beet showed a smaller correlation between the level of P supplementation and P content in plant leaves. This dependence is consistent with some observations in previous studies indicating that P fertilization increases the phosphate content of sugar beet roots [55]. Most of the plant nutritional elements were highly correlated with each other. Numerous significant correlations existed between nutrients in sugar beet, especially between K and Ca (PCC=0.88) and N and Mg (PCC=0.82). The strong correlations between chlorophyll content and carotenoids were observed for celery (PCC=0.85) and strawberry (PCC=0.95) plants. This is because chlorophylls and carotenoids are co-varying in nature (as a components of photosynthetic antenna complexes) and statistically dependent, as observed in previous studies [16, 56].
Negative correlations between P supplementation in soil and mass of the plant roots suggest that P deficiency promotes a reduction in the mass and length of roots in root vegetables and causes the reduction in yield. In the case of strawberry, this correlation was positive. The concentration of chlorophylls and carotenoids in the above-ground parts of the tested plants significantly affected the mass of their roots. Sugar beet had a positive correlation between the carotenoid content and root mass (PCC = 0.6), whereas a strong negative correlation was observed between the concentration of chlorophyll and carotenoid content in leaves and root mass for celery plants.
c. Spectral features of plants
Figure 4 represents the general scheme of the procedure to obtain spectral characteristics from the leaf surfaces of the three studied plants. The average reflectance spectra of ROIs, covering the spectral range of 400-2500 nm for leaf samples of the three studied species of plants with different P treatments and for five development stages, are shown in Fig. 5. The spectral curves of the leaves of the three studied species exhibited similar shapes, although differences are visible between the spectra belonging to specific variants. In the visible spectral region, a characteristic peak was observed at 550 nm with some differences between variants, especially in sugar beet and strawberry. This peak is characteristic of chlorophyll absorption. In the region of rapid change in the reflectance of vegetation in the near infrared range of 650-750 nm of the electromagnetic spectrum (so called red edge), high increases of reflectance occur, which enabled us to distinguish differences between some variants of the experiments. The highest differentiation between the spectral curves of the plants belonging to specific variants was observed in the range of 750-1300 nm, in which reflectance patterns are strongly connected with the internal cellular structure of plants [57]. Unfortunately, in this range, there was a break (discontinuity) in the registered reflected radiation, which is connected to low sensitivity of the two spectral cameras used in the part of this range. Because of this, the raw spectra of the leaves in this range were not good at distinguishing between variants. Another part of the spectrum that seems to be appropriate for distinguishing differences between variants is absorption at approximately 1400 and 1950 nm, which are highly related to the absorption by water. The results presented in Fig. 5 indicate quantitative relationships between the amount of reflected light and P treatment at the succeeding growing stages. In plants of all three species, the highest changes in reflectance values were observed in the SWIR region (2200 – 2400 nm). It suggests that the SWIR region is useful for distinguishing levels of P fertilization. Wavelengths in the SWIR region are mainly associated with light absorption by proteins, nitrogen, cellulose, starch and sugar. It is known that P plays an important role in protein synthesis, which may explain these differences, as suggested by Knox et al. [58].
d. Effective wavelength selection
To reduce the high dimensionality of the extracted spectral data and to make the classification models more robust, the most appropriate wavelengths that give the highest discrimination among different levels of P-treatment were selected based on 2nd derivative transformation of raw spectra. The 2nd derivative averaged spectra are shown in Fig. 6. Based on the second derivative transformation of the original spectra and by applying the CFS algorithm with greedy stepwise selection method, 10, 7 and 4 wavelengths were selected for classification according to the P treatment of sugar beet, celery and strawberry plants, respectively (Table 1). The wavelengths used to distinguish between levels of P fertilization were localized in the blue spectral band (400-480 nm), NIR (760-900 nm) and SWIR (1000-2500 nm) regions of the spectrum. In all studied plants, the level of P supply did not significantly affect the reflectance in the green (500-560 nm) region. In the case of strawberry plants, the wavelengths from the red region (715 and 723 nm) and SWIR region (2301 and 2332 nm) had particular importance for the separation of the levels of P treatment. The wavelengths in the red and far-red regions of the electromagnetic spectrum (723, 754, 715 and 723 nm) in plants are mainly associated with the absorption of Chlorophyll a. It was shown that varied P rates cause changes in the concentration of Chlorophyll a in plant leaves (Fig. 1).
Table 1. Wavelengths selected based on the second derivative transformed spectra and CFS algorithm with greedy-stepwise selection methods
Plant species
|
Number of selected wavelengths
|
Selected wavelengths [nm]
|
|
|
Sugar beet
|
10
|
422, 569, 723, 850, 1250, 2227, 2276, 2314, 2345, 2351
|
|
Celery
|
7
|
414, 419, 429, 564, 754, 1395, 2264
|
|
Strawberry
|
4
|
715, 723, 2301, 2332
|
|
The previous study performed by Osborne et al. [9] also indicated that NIR (730 nm and 930 nm) and blue (440 and 445 nm) regions of the spectrum are useful for the prediction of P concentrations in corn canopy. Differences in the selected wavelengths among the three studied species might be due to differences in plant structure or changes in the chemical concentration.
e. Results of discrimination analysis
The prediction accuracies of the models created to distinguish between plants grown under different levels of P treatment at different development stages obtained for the three studied plant species are presented in Table 2. It results from the analysis of the supervised classification algorithms that very similar and relatively high prediction accuracies in the majority of cases (ranging from 40% to 100% for validation sets) were obtained for all four methods of machine learning model creation methods (i.e., backpropagation neural network, random forest, naive Bayes and support vector machines). In all cases, despite very limited numbers of wavelengths selected for the classification (from 4 to 10), the prediction accuracies for training sets were very high in all variants of the experiment. This confirms a good performance of the CFS wavelength selection algorithm and is in agreement with other studies on plant material classification with the use of this algorithm [26, 59]. The performance of the validation sets was considerably lower than that of the training sets, but the accuracy at distinguishing between various levels of P treatment were equal or higher than 80% in 11 variants among 15 variants of species/stages of plant development. This result is very good although difficult to compare with other studies that used different experimental setups and limited numbers of P treatment variants [38, 42]. The lowest percentages of correctly classified instances were obtained for the first stage of plant development; however, with progress in the development of plants, this accuracy was higher. This result comes from the fact that, in the first period of plant development, the changes in leaf spectral properties are considerably minimal between various P treatments and misclassification, especially with one level or higher of P fertilization. Although all four methods of supervised classification model creation were highly effective, the highest overall classification performance was obtained for RF models. The validation results indicated that this model correctly classified more than 70% of all instances in the case of strawberry plants and more than 80% (except the second term) for celery across five development stages. The average accuracy of RF classification for sugar beet was lower compared to other plants (65%). This result might be explained by the specific nutrient requirements of the sugar beet [55, 60] and its lower sensitivity of imbalanced P-fertilization than strawberry and celery plants.
Table 2. Model performance on selected wavelengths for classification of the level of P treatment at five developmental stages obtained for the three studied species of plants
Plant species
|
Sugar beet
|
Celery
|
Strawberry
|
Model
|
BNN
|
LIBSVM
|
LOG
|
RF
|
BNN
|
LIBSVM
|
LOG
|
RF
|
BNN
|
LIBSVM
|
LOG
|
RF
|
I
|
Training set
|
%
|
98
|
84
|
91
|
100
|
98
|
91
|
100
|
100
|
86
|
79
|
68
|
100
|
RMSE
|
0.12
|
0.28
|
0.18
|
0.15
|
0.12
|
0.21
|
0
|
0.12
|
0.19
|
0.32
|
0.31
|
0.1
|
Validation set
|
%
|
55
|
45
|
50
|
65
|
75
|
60
|
80
|
80
|
80
|
60
|
55
|
70
|
RMSE
|
0.44
|
0.52
|
0.49
|
0.35
|
0.3
|
0.45
|
0.32
|
0.32
|
0.29
|
0.45
|
0.39
|
0.29
|
II
|
Training set
|
%
|
98
|
89
|
100
|
100
|
98
|
93
|
95
|
100
|
97
|
97
|
100
|
100
|
RMSE
|
0.11
|
0.24
|
0
|
0.11
|
0.12
|
0.18
|
0.17
|
0.13
|
0.09
|
0.11
|
0
|
0.05
|
Validation set
|
%
|
70
|
70
|
65
|
70
|
55
|
60
|
50
|
55
|
80
|
90
|
80
|
95
|
RMSE
|
0.37
|
0.39
|
0.42
|
0.29
|
0.39
|
0.45
|
0.48
|
0.34
|
0.26
|
0.22
|
0.31
|
0.16
|
III
|
Training set
|
%
|
95
|
86
|
100
|
100
|
100
|
91
|
100
|
100
|
98
|
77
|
82
|
100
|
RMSE
|
0.13
|
0.26
|
0
|
0.12
|
0.06
|
0.22
|
0.01
|
0.11
|
0.15
|
0.34
|
0.22
|
0.09
|
Validation set
|
%
|
65
|
60
|
40
|
75
|
80
|
95
|
55
|
80
|
100
|
85
|
80
|
95
|
RMSE
|
0.23
|
0.45
|
0.55
|
0.33
|
0.29
|
0.16
|
0.47
|
0.28
|
0.17
|
0.27
|
0.24
|
0.15
|
IV
|
Training set
|
%
|
100
|
93
|
100
|
100
|
100
|
97
|
100
|
100
|
100
|
100
|
100
|
100
|
RMSE
|
0.07
|
0.19
|
0
|
0.11
|
0.05
|
0.11
|
0
|
0
|
0.02
|
0
|
0
|
0.01
|
Validation set
|
%
|
55
|
85
|
55
|
90
|
80
|
75
|
65
|
85
|
100
|
100
|
95
|
100
|
RMSE
|
0.38
|
0.27
|
0.47
|
0.27
|
0.27
|
0.35
|
0.41
|
0.27
|
0.03
|
0
|
0.14
|
0.05
|
V
|
Training set
|
%
|
87
|
86
|
100
|
100
|
100
|
95
|
100
|
100
|
91
|
86
|
100
|
100
|
RMSE
|
0.21
|
0.26
|
0
|
0.13
|
0.03
|
0.15
|
0
|
0
|
0.2
|
0.26
|
0
|
0.11
|
Validation set
|
%
|
60
|
45
|
65
|
70
|
95
|
95
|
90
|
95
|
65
|
50
|
80
|
80
|
RMSE
|
0.37
|
0.52
|
0.41
|
0.35
|
0.15
|
0.16
|
0.22
|
0.16
|
0.34
|
0.5
|
0.32
|
0.3
|
To assess the performance of the analysed models for specific P levels in five developmental stages, confusion matrices were created, which enabled us to identify misclassification percentages for analysed variants of the experiment. The summary of this analysis is presented in Table 3 for RF models, which gave the best overall results in the performed experiments. The grey cells in this table represent variants with 100% accuracy (all cases classified correctly), yellow cells show misclassified variants in which misclassification refers to one level up or down with respect to the analysed P fertilization level (e.g., P-33 level classified as P-67 level), and red cells indicate misclassification higher than one P fertilization level (e.g., variant P-133 classified as P-33). The confusion matrices for all models divided to 5 growth stages are presented in Table S1 in Supplement 1. For each developmental stage, the percentages of misclassified cases are given, and it is possible to see how misclassification occurred (second column in this table indicates the analysed variants, and separate rows show with which variants they were misclassified and what percent of misclassification occurred). From this table, 100% accuracy was achieved (all cases classified correctly) for 26 variants of the experiment (P level vs development stage), misclassification was one level up or down with respect to the analysed P fertilization level in 27 variants, and misclassification was higher than one P fertilization level in only 13 variants. In the majority of misclassified variants (26), improperly classified cases reached only 20%, there were only 9 variants with misclassified cases of 40%, 3 variants with misclassified cases of 60% and 1 variant with misclassified cases of 80%. Table 3 also shows that there were only 6 variants for which two different levels of P treatment were assigned for a given level, five of which occurred for the first and second plant developmental stages. In Fig. 7, the numbers of misclassified cases in the validation dataset for random forest (RF) models of P content in plant treatment for 5 stages of plant growth and 3 studied species are presented, and these are based on the confusion matrices presented in Table S1 in Supplement 1. This figure shows that the highest number of misclassified cases for sugar beet and strawberry occurred during the first stage of plant growth, whereas this occurred during the second stage of plant growth for celery. This confirms that, in the early stages of plant growth, spectral properties of the affected plant leaves do not always distinguish differences in P content. Despite this, the overall classification performance of the chosen models (and especially RF models) was very good.