Predictions of Spartina alterniflora leaf functional traits based on hyperspectral data and machine learning models

ABSTRACT

Investigating the functional traits of Spartina alterniflora can provide insights towards understanding its invasion mechanism, and developing a method leaves can improve its management in coastal wetlands. Here, we examined the relationship between 11 leaf functional traits of S. alterniflora and hyperspectral data and investigated the feature bands through importance score analysis. Using original spectral and first-order differential conversion data of feature bands, we established four prediction models: random forest (RF), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost), and back propagation neural network (BPNN). The study results showed that: (1) the SVM model based on Random Forest Importance Score is well-suited for S. alterniflora leaf functional trait inversion; (2) the importance score of leaf functional traits differed, and first-order differential spectral data produced more bands with high scores compared with the original hyperspectral reflectance data; (3) first-order differential data modelling effects were slightly better than those of the original spectral data. However, the first-order differential treatment did not show a significant improvement in the validation accuracy compared with the original data, and the accuracy of some traits decreased. Our study provides a new methodological approach for improving the monitoring and management of S. alterniflora in coastal wetlands.

KEYWORDS:

• Coastal wetlands
• functional traits
• hyperspectral data
• machine learning
• spartina alterniflora

Introduction

Coastal wetlands are ecologically dynamic areas where the sea and land intersect, and they play important roles in protecting coastlines, preventing erosion, regulating climate, purifying water, and providing habitats for a diverse range of organisms. Thus, these wetlands are among the most abundant ecosystems in terms of species richness worldwide (Barbier et al., 2011; Cai et al., 2009; Sheaves et al., 2014). However, with the increasing impact of human activities and climate change, these wetlands are facing serious threats, resulting in considerable damage to their ecosystem functions (Blankespoor et al., 2014; Kirwan & Megonigal, 2013; Tian et al., 2016). Spartina alterniflora (S. alterniflora) is a perennial herb commonly found in intertidal wetlands. It is native to the southeastern coast of the United States and provides ecological functions such as soil improvement, air purification, greening, wind and wave protection for soil, beach sand suppression, and wetland area expansion (Liu et al., 2021; Yu et al., 2022). Introduced to China’s coast in 1979, S. alterniflora initially played an important ecological role in resisting typhoons and ameliorating wave-induced damage to the coastline (An et al., 2007). However, recent studies have identified it as a threat to the biodiversity of coastal wetland ecosystems (Chen et al., 2018; Liu et al., 2018). Therefore, studies on S. alterniflora can help provide a theoretical basis for the conservation and monitoring of local biodiversity.

Plant functional traits are defined as a series of plant morphological and physiological attributes that can significantly affect their colonization, survival, and growth (Violle et al., 2007; Westoby & Wright, 2006). Leaf functional traits are among the most important components of plant functional traits, and they can reflect the adaptive strategies of plants to environmental changes (Jankowska & Włodarska-Kowalczuk, 2022; Jiang et al., 2022). For instance, the specific leaf area (SLA) and leaf nutrient content are closely related to the relative growth rate (Reich & Oleksyn, 2004). Additionally, the chlorophyll concentration, leaf nitrogen, SLA, and nitrogen to phosphorus ratio are indicative of the survival strategies of plants to optimize their access to resources (Carlucci et al., 2015; Liu et al., 2017).

Research on the functional traits of wetland plants has increased in recent years. Zhang (2022) analysed the functional traits of seven dominant plants in the lakeside wetland of the Inner Mongolian Plateau, including Phragmites australis and Suaeda glauca. They discovered that environmental changes significantly affect plant height, leaf carbon content, and leaf nitrogen content. Similarly, Zhou et al. (2022) studied the leaf functional traits of Phragmites australis, a representative wetland plant, and observed that the traits leaf thickness (LT) and SLA decreased from wetland to desert habitats and soil bulk density impacted the leaf functional traits in wetland habitats. In another study, Zhu (2023) conducted a statistical comparison of the functional characteristics of 12 wetland plants, including Hemerocallis fulva and Isachne globose, and found significant correlations between certain traits and rich variations in the functional traits of wetland plants. Leaf functional traits of wetland plants are closely related to their resource utilization efficiency, which can reflect the adaptation strategies of wetland plants under different resource conditions, and are of great ecological importance.

Traditional measurement methods for leaf functional traits involve the transportation of leaf samples back to the laboratory for chemical measurement immediately after collection. Although the measurement accuracy of this processing method is high, it is time-consuming and laborious. Moreover, due to the complexity of the geographic conditions of coastal wetlands and high temperature and humidity conditions in summer, leaf traits are difficult to measure in a timely manner, and dried, curly leaves lead to measurement deviations. However, the increasingly mature hyperspectral remote sensing technology has provided new ideas and methods for solving the current problem of measuring plant functional traits. The spectral data acquired by this technology are high-resolution, information-rich, and characterized by rapid, nondestructive, repeatable and large-area monitoring, which has enabled high-throughput and high-precision monitoring of plant functional traits at the leaf scale (Asner et al., 2016; Li et al., 2020; Pacheco-Labrador et al., 2022; Rebelo et al., 2018).

Currently, significant progress has been made in the use of hyperspectral data for the inversion of vegetation functional traits in forest and grassland ecosystems (Gitelson et al., 2001; Yendrek et al., 2017). In wetlands, several studies have also explored the inversion of functional traits using hyperspectral techniques. For example, the correcting normalized difference nitrogen index (NDNI) has been proved to have great potential in the prediction of canopy and leaf nitrogen content of aquatic vegetation in wetlands (Wang & Wei, 2016; Zhang et al., 2019). Dou et al. (2019) used linear and partial least squares regressions to establish an inversion model for Suaeda salsa biomass and found that the partial least squares regression model had good accuracy. In addition, random forest (RF) model has been shown to be effective in inverting the ecological chemometric characteristics of Phragmites communis and Kandelia candel (Cui et al., 2020; Tang et al., 2022). And Li et al. (2020) estimated the total nitrogen (TN) and total phosphorus (TP) contents of artificial wetlands using two wetland plant canopy spectral reflectance data, thus confirming the feasibility of using plant canopy spectral reflectance data to estimate water and sediment TN and TP contents. Furthermore, some scholars have also explored the inversion of functional traits of wetland plants using hyperspectral techniques. For instance, Rebelo et al. (2018) modelled the relationship between the canopy spectra and functional traits of 22 dominant plant species in South African wetlands, thus providing an effective tool for the differentiation of plant functional groups and even species. Zuo (2021) conducted an inversion study of the functional traits of S. alterniflora using machine learning models, such as RF, and they achieved good accuracy and provided theoretical support for the inversion of functional traits using hyperspectral techniques.

An increasing number of studies have shown that plant functional traits play an important role in mechanistic analysis and process prediction at numerous ecological scales in organs, species, and ecosystems, and represent a reliable method of addressing important ecological questions at the population, community, and ecosystem scales (He et al., 2022; Levine, 2016; Sack et al., 2013). Systematic studies on the functional traits of S. alterniflora have revealed the growth status of S. alterniflora and analysed its growth strategies and adaptive mechanisms. Although traditional trait measurement methods are difficult to apply for the large-scale measurement of S. alterniflora functional traits, the rapid development of hyperspectral technology provides technical possibilities for realizing the inversion of functional traits of S. alterniflora. Exploring methods for inverting plant functional traits through hyperspectral techniques can pave the way for direct predictions of a wide range of plant functional traits from hyperspectral data and analyses of ecological problems based on the predicted trait values. In this study, we aimed to identify the optimal hyperspectral techniques and machine learning models to invert the functional traits of S. alterniflora. Specifically, we measured the leaf functional traits of S. alterniflora and used hyperspectral data to establish prediction models of the main leaf functional traits. This study provides a theoretical basis for further understanding the invasion mechanism of S. alterniflora and its environmental response strategies and proposes a new approach for monitoring and managing S. alterniflora in coastal wetlands.

Materials and methods

Study area

The Yancheng coastal wetland in Jiangsu Province, China (32°56’–33°36’ N, 120°13’–120°56’ E), is an extensive intertidal system covering a 582 km coastline, and it is located along the lower reaches of the Huaihe River and faces the Yellow Sea to the east. The region experiences northerly winds from high-latitude regions in winter and southerly winds from the Pacific Ocean in summer, resulting in a regulated annual average temperature with limited variations. The area is characterized by various types of coastal wetland ecosystems that include salt marsh wetlands, mudflats, and shallow water areas. The study site was located in the Dafeng Elk Nature Reserve of Yancheng and the Jiangsu Yancheng Wetland National Nature Reserve Rare Birds, which are known for their diverse halophytic vegetation, including S. alterniflora, P. communis, and Imperata cylindrica. However, the rapid expansion of S. alterniflora since its initial invasion has become a major threat to the local ecology and economy, thus highlighting the need for effective management and monitoring of this invasive species.

Acquisition of hyperspectral data

In August 2019 and August 2020, 238 sampling points were randomly arranged and sampled in the Yancheng coastal wetland, with 120 sampling sites located in Dafeng Elk and 118 sampling sites in Rare Birds. Sampling sites were separated by 10 m, and each site consisted of a single species: S. alterniflora. Spectral information was acquired using an FS4 portable ground object spectrometer produced by the Analytical Spectral Device (ASD) company (United States), and it had a spectral range of 350–2500 nm and a sampling interval of 1 nm. Three unfurled, intact, and disease-free healthy leaves were randomly collected from the upper, middle, and lower parts of S. alterniflora at each sampling site. An ASD FS4 with leaf clips was utilized to measure the spectral reflectance from the middle part of the leaves, avoiding the main leaf veins. Ten spectral data points were collected for each leaf, and the spectral reflectance values of the three leaves were averaged to represent the spectral reflectance of the leaves at the sampling point.

Determination of leaf functional traits

The freshly collected leaves were sealed in aluminum envelopes after collection, stored in a portable cooler, and measured for their fresh weight (FW), LT, soil plant analysis development (SPAD) index, and leaf area within 4 h of collection. Subsequently, the leaves were blanched at 105°C for 15 min and then dried to a constant weight at 85°C for 48 h to obtain the dry weight. Other indicators were determined after crushing and sieving the dried leaves. The TN and total carbon (TC) contents were simultaneously determined using an elemental analyser (vario PYRO cube; Elemental, Germany). The plant samples were decomposed via high-temperature combustion, and the mixed gas was automatically measured using a thermal conductivity detection system. The organic phosphorus in the plants was converted into phosphate by digestion with H₂SO₄–H₂O₂ and measured using a spectrophotometer. To determine the TP content, a standard curve was generated. Table 1 lists the methods used for determining the various leaf functional traits of S. alterniflora.

Table 1. Methods for determining the main leaf functional traits of Spartina alterniflora.

Leaf functional trait	Measurement method
Moisture content (MC)	[(Wet weight − Dry weight)/Dry weight] × 100%
Soil plant analysis development (SPAD)	Chlorophyll meter (SPAD-502 Plus; Konica Minolta, Japan)
Specific leaf area (SLA)	Leaves and extracts the leaf area are scanned through ImageJ software. Specific leaf area: single leaf area/single leaf dry weight
Total nitrogen (TN)	Element analyser (Vario PYRO cube; Elemental, Germany)
Total phosphorus (TP)	Molybdenum Antimony Colorimetry (NY/T 2421-2013)
Total carbon (TC)	Element analyser (Vario PYRO cube; Elemental, Germany)

Spectral data conversion

Viewspec Pro software (ASD, Inc., Boulder, Colorado, USA) was used to process the collected hyperspectral data of the S. alterniflora leaves. The spectral curve had to be corrected because the acquisition of the spectrum may cause pulsation of the connection points.

In this experiment, two spectral conversion methods, namely, the original spectral and first-order differential (FD) reflectance methods, were used to study the relationship between spectral reflectance and leaf functional traits. The FD processing of the spectrum can amplify the difference in the spectral characteristics among the samples (Chen et al., 2019), and its applicability for functional trait inversion has been demonstrated (Zuo, 2021). The FD conversion formula is as follows:

(1) $\mathrm{F}\mathrm{D}\mathrm{R}\left({\lambda }_i\right)=\frac{R\left({\lambda }_{i+1}\right)-R\left({\lambda }_{i-1}\right)}{\mathrm{\Delta }\lambda }$(1)

where ${\lambda }_i$ is the wavelength at band i, FDR(${\lambda }_i$) is the first-order differential spectral value at wavelength ${\lambda }_i$, and $\lambda$ is the wavelength difference between band i +1 and band i.

Feature selection

Utilizing hyperspectral data directly can pose challenges in optimization and lead to extended computational times due to the characteristics of the data, including a large number of bands, high correlation, and data redundancy. Therefore, we used the special importance calculation methods that come with both Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) algorithms to calculate the importance scores of the input variables for each leaf functional trait of S. alterniflora, respectively. We arranged them in descending order and filtered out the bands with importance scores exceeding the standard deviation, identifying them as the characteristic bands. The computation of RF feature importance scores was implemented in MATLAB, while the computation of XGBoost feature importance scores was implemented in Python using the xgboost package.

Model establishment

Based on the principle that the ratio between the number of validation and training sets is 1/3, 238 samples were randomly divided into 179 modelling and 59 verification samples. We chose four models, RF, Back Propagation Neural Network (BPNN), Support Vector Machine (SVM) and XGBoost, to realize the inversion of functional traits of S. alterniflora for hyperspectral data. SVM is a kernel-based learning method that uses a kernel function to map input variables to a high-dimensional feature space and extracts a linear hyperplane from the feature space as a decision function to solve the regression problem (Beltrami & da Silva, 2020). The learning rule of BPNN is to minimize the sum of squared errors of the network by using the most rapid descent method, which continuously adjusts the weights and winks of the network through backpropagation (Wan et al., 2018). RF is an integrated algorithm based on decision trees. Its fundamental concept involves the random selection of multiple samples, with replacement, from the training set to construct a set of entirely independent decision trees. The final prediction result is independent of each other and the final prediction result is decided by the majority voting principle (Breiman, 2001; Liaw & Wiener, 2002). XGBoost is a machine learning algorithm for regression, classification, and ranking. It is an efficient implementation of Gradient Boosting Decision Trees (GBDT) that can run on large-scale datasets with strong generalization capabilities (Chen & Guestrin, 2016). The three models RF, BPNN, and SVM are implemented relying on the Weka 3.8.0 software, while the XGBoost model is implemented based on the xgboost package for Python. Default values were used for the parameters of all four models. Weka is an open-source data mining software in the field of machine learning (Hall et al., 2009; Suykens, 2014).

The accuracy of the models was assessed by calculating the determination coefficient (R²) of the regression model and the root mean square error (RMSE) of the model simulation value. The model with the highest accuracy was selected as the optimal model. The calculation formulas are as follows:

(2) ${\mathrm{R}}^2=\frac{{\sum }_{i=1}^n(y-y_i{)}^2}{{\sum }_{i=1}^n(\bar{y}-y_i{)}^2}$(2)

(3) $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{{\sum }_{i=1}^n(y-y_i{)}^2}{n}}$(3)

where $y$ represents the measured values of leaf functional traits, $y_i$ represents the model predictive value of leaf functional traits, $\bar{y}$ represents the measured average value of the leaf functional traits, and n represents the number of samples.

Results

Leaf functional traits

gThe leaf functional traits of S. alterniflora were analysed, and the results are presented in Table 2. The FW of a single leaf ranged from 0.68–5.84 g, with a standard deviation of 0.84. The largest LT value was 0.74 mm, and the standard deviation was 0.09. The minimum, maximum, and standard deviation values of the SPAD index were 14.97, 54.00, and 5.38, respectively. Among the nutrient indicators, the TC content had the highest average value of 360.96 g/kg and the TP content had the lowest average value of 1.90 g/kg. The mean N/P, C/N, and C/P values were 8.00, 25.31, and 298.63, respectively, with N/P being the lowest. The average MC of the leaves was 68.18, with a standard deviation of 5.75. The maximum observed SLA value was 87.42 cm²/g, with a standard deviation of 28.33. SLA data showed the largest dispersion, with a coefficient of variation of 32.40%, while MC data showed the smallest, with a coefficient of variation of 8.43%. Overall, the range of functional traits and inclusion of the standard deviation indicate the differences among the leaves of S. alterniflora, thus providing conditions for the inversion of functional traits based on hyperspectral data.

Table 2. Leaf functional traits of Spartina alterniflora.

Leaf functional trait	No. of Samples	Minimum	Maximum	Mean	Standard Deviation	Coefficient of Variation
FW (g)	238	0.68	5.84	2.64	0.84	31.92%
LT (mm)	238	0.29	0.74	0.52	0.09	17.79%
SPAD	238	14.97	54.00	42.32	5.38	12.72%
TN (g/kg)	238	9.23	36.17	14.60	2.90	19.85%
TP (g/kg)	238	0.90	3.19	1.90	0.38	20.05%
TC (g/kg)	238	258.74	545.84	360.96	42.10	11.66%
MC	238	21.55	83.82	68.18	5.75	8.43%
SLA (cm^2/g)	238	24.02	258.24	87.42	28.33	32.40%
N/P	238	4.45	21.14	8.00	2.29	28.65%
C/N	238	9.23	37.83	25.31	3.89	15.35%
C/P	238	111.69	424.71	198.63	50.48	25.36%

Note: FW: fresh weight; LT: leaf thickness; SPAD: soil plant analysis development; TN: total nitrogen; TP: total phosphorus; TC: total carbon; MC: moisture content; SLA: specific leaf area; N/P: TN/TP; C/N: TC/TN; C/P: TC/TP.

Extraction of spectral features

In our study, we used two importance score methods to extract characteristic bands for the functional traits of S. alterniflora leaves. The importance scores of the original and FD spectra for each functional trait based on RF are shown in Figure 1, while Figure 2 shows the importance scores of OR and FD spectra for each functional trait based on the XGBoost method. Each functional trait had specific importance score bands.

Figure 1. Analysis of random forest (RF) importance scores for the original (OR) and first-order differential (FD) hyperspectral data: (a) FW, (b) LT, (c) SPAD, (d) MC, (e) SLA, (f) TN, (g) TP, (h) TC, (i) C/N, (j) C/P, and (k) N/P.

Figure 2. Analysis of eXtreme Gradient Boosting (XGBoost) importance scores for the or and FD hyperspectral data: (a) C/N, (b) C/P, (c) N/P, (d) TC, (e) TN, (f) TP, (g) FW, (h) LT, (i) MC, (j) SLA, and (k) SPAD.

Feature bands based on RF importance scores

Figure 1 shows that the number of bands with high-importance scores increased significantly for the FD data compared with that for the original spectral data, and the largest importance scores were observed. Specifically, TN, TP, C/P, C/N, MC, and SPAD had more high-importance bands for the original spectral data while other functional traits had fewer. However, for the FD processed spectral data, each functional trait index had bands with higher importance scores. It should be noted that TC and SLA had relatively fewer bands with high importance scores for the FD processed spectral data.

Bands with values larger than the standard deviation of the importance score were selected as characteristic bands (Supplementary table S1). The high-importance score bands of TN, TP, C/N, C/P, and N/P were distributed across all ranges; nevertheless, it mainly concentrated in the 300–700 nm band. The high-importance score bands of the FD processing data of TC were mainly distributed between 2200 and 2500 nm, while those of the original spectral data of TC were mainly distributed between 300 and 500 nm. The high-importance score bands of FW and SLA were mainly distributed between 300 and 500 nm, while those of SPAD were distributed between 400 and 800 nm. For LT, the high-importance score bands of the FD processing data were mainly distributed between 900 and 1300 nm while those of the original spectral data were between 300 and 400 nm and 1400 and 1500 nm. For MC, the high-importance score bands of the FD processing data were distributed in the 1600–1900 nm range, while those of the original spectral data were in the 300–500 nm range.

Feature bands based on XGBoost importance scores

Combining Figure 2 and Supplementary table S2, we find that FW screens the least number of feature bands, which is due to the fact that it has bands with importance scores greater than 0.3 in both OR spectral data and FD spectral data. The feature bands of the FD spectra of LT are mainly distributed in the 400–1400 band range, while the feature bands of the OR spectra, although distributed in all ranges, have fewer feature bands after 2000 nm. The feature bands of SLA and SPAD are mainly concentrated in the region of the visible wavelength. In terms of the distribution of the characteristic bands of chemical element content, TC and TN have fewer characteristic bands, while C/N has more characteristic bands. In addition, the importance scores of the characteristic bands of TP, C/N, C/P and N/P are generally low, which leads to the high number of feature bands.

Evaluation of leaf functional trait modelling

Model accuracy based on random forest band importance scores

The remaining 59 validation data samples were used to evaluate the performance of the four models in predicting the functional traits of S. alterniflora leaves. The validation results are shown in Table 3, which shows that the RF model had the best inversion accuracy for TC, with R² values of 0.56 and 0.69 and RMSEs of 27.46 and 26.38 for the original and FD data, respectively. However, N/P showed the lowest prediction accuracy. The SVM model had the best inversion accuracy for MA, with R² values of 0.92 and 0.90 for the original and FD data, respectively. The BPNN model had the best inversion accuracy for LT using the original data, with an R² of 0.88; nevertheless, it had the best inversion accuracy for FW using the FD data, with an R² of 0.76. The XGBoost model had the best inversion accuracy for TC using the FD data, with an R² of 0.66; however, it showed but poor prediction accuracy for eco-chemometric features, such as TP and N/P, with R² values of less than 0.1, which is similar to the validation by RF.

Table 3. Modelling accuracy verification.

Model	Leaf function trait	Origin		First-order differential
		R^2	RMSE	R^2	RMSE
RF	FW	0.53	0.63	0.65	0.62
	LT	0.24	0.08	0.39	0.08
	SPAD	0.20	4.40	0.44	3.81
	MC	0.53	3.32	0.62	3.08
	SLA	0.27	23.80	0.29	25.10
	TN	0.13	2.51	0.17	2.43
	TP	0.05	0.38	0.18	0.36
	TC	0.56	27.46	0.69	26.38
	C/N	0.06	3.61	0.20	3.32
	C/P	0.01	51.45	0.14	48.29
	N/P	0.01	2.37	0.02	2.26
SVM	FW	0.88	1.06	0.89	1.06
	LT	0.71	0.07	0.76	0.07
	SPAD	0.87	6.04	0.85	6.04
	MC	0.92	6.04	0.90	6.03
	SLA	0.47	33.19	0.28	33.06
	TN	0.82	3.31	0.81	3.30
	TP	0.85	0.42	0.86	0.42
	TC	0.21	48.23	0.41	44.55
	C/N	0.87	4.40	0.87	4.41
	C/P	0.30	59.72	0.26	59.73
	N/P	0.56	2.64	0.55	2.65
BPNN	FW	0.76	1.13	0.76	1.22
	LT	0.88	1.12	0.63	0.14
	SPAD	0.54	6.91	0.62	6.45
	MC	0.53	5.70	0.38	6.62
	SLA	0.23	41.07	0.03	52.15
	TN	0.47	4.56	0.28	4.68
	TP	0.60	0.56	0.49	0.51
	TC	0.69	56.21	0.72	56.88
	C/N	0.45	3.96	0.23	5.95
	C/P	0.51	66.20	0.48	58.70
	N/P	0.28	3.36	0.41	2.73
XGBoost	FW	0.57	0.57	0.65	0.51
	LT	0.20	0.08	0.30	0.08
	SPAD	0.16	4.56	0.32	4.08
	MC	0.45	3.72	0.59	3.02
	SLA	0.16	25.33	0.29	23.69
	TN	0.12	2.68	0.17	2.51
	TP	0.01	0.41	0.07	0.38
	TC	0.44	31.03	0.66	24.11
	C/N	0.01	4.04	0.03	3.84
	C/P	0.01	58.72	0.02	52.18
	N/P	0.01	2.55	0.02	2.42

Overall, the SVM model had better validation accuracy than the other three models, however, it showed the lowest modelling accuracy. In addition, the difference in RMSE between the predicted values of each functional trait after FD processing was not significant compared with that for the original data. However, in the modelling, the difference in RMSE between the FD and original data for each functional trait was significant.

In addition, the FD treatment improved the accuracy of the RF and XGBoost models, and the inversion accuracy of the FD treatment data for each functional trait was higher than that of the original data, although the effect was not apparent for the SVM and BPNN models. In the SVM model, the inversion accuracy of the original data for MC and SLA was significantly higher than that of the FD processing data, while the inversion accuracy of the original data for TC was significantly lower. Similarly, in the BPNN model, the inversion accuracy of the original data for LT, SLA, TN, and C/N was also greater than that of the FD processing data. The accuracy of the four models for most functional traits decreased in the validation phase to varying degrees, with the RF and XGBoost model being the most affected and showing high modelling accuracy but poor validation results. Nonetheless, some traits did show improvements in validation accuracy, such as the MC prediction based on the SVM model, which presented R² values of 0.92 and 0.90 for the original and FD processed data, respectively.

Model accuracy based on xgboost band importance scores

RF model

The accuracy of the RF model is shown in Figure 3, where a larger R² and smaller RMSE indicate better model prediction ability and higher model ability. Among the six traits of carbon, nitrogen and phosphorus and their ecological stoichiometric characteristics, TC exhibited the highest validation accuracy, whose R² of the spectral data of OR and FD was about 0.6, and the R² of the other traits was below 0.2. Meanwhile, the validation accuracies of FW and MC were relatively good, and the R² of the two data treatments were above 0.5. Overall, the data of each trait after FD processing showed better validation effects, except for SLA. The R² of the FD data of SLA was 0.26 and the RMSE was 23.81, whereas the R² of the OR data was 0.41 and the RMSE was 20.98. The validation accuracy of the FD data showed a decrease in the R² and an increase in the RMSE compared with that of the original data.

Figure 3. Scatter plots of RF model accuracy.

SVM model

The results of the SVM validation are shown in Figure 4. The highest accuracy of OR data validation is FW with R² of 0.57, RMSE of 0.59, and the worst validation of C/P with R² of only 0.07 and RMSE of 50.82. The best validation of FD data is MC with R² of 0.65, RMSE of 2.85, and the lowest accuracy is TP with R² of 0.13, RMSE of 0.36. The SVM validation results show that the validation accuracy of hyperspectral data after first-order differentiation is significantly improved compared with that of the original spectral data. Overall, the validation accuracy of the SVM model is lower than that of the RF model for both the OR data and the FD data.

Figure 4. Scatter plots of SVM modelling accuracy.

BPNN model

The effect of the BPNN model is illustrated in Figure 5, which shows that the accuracy of the neural network model was relatively different between the original and FD processing spectra. The validation accuracy of the BPNN model is poor, with R² mostly below 0.3. Overall, the validation accuracy has not improved much after first-order differentiation, and even C/N and FW have the phenomenon of smaller R². Combined with the scatter distribution in the figure, only FW and TC have slightly better model validation accuracies after FD processing, with R² of 0.53 and 0.54, respectively. In addition, traits such as TC and C/P exhibit a situation where both R² and RMSE increase after the FD processing. This contradicts the principle that the larger R² and the smaller RMSE indicate higher model accuracy, suggesting lower model reliability. Comparing the validation accuracies of different models for the same trait, the RMSE of the BPNN model is larger than that of RF and SVM, indicating that the model is less stable.

Figure 5. Scatter plots of BPNN modelling accuracy.

Xgboost model

The accuracy of the XGBoost model is shown in Figure 6. After the FD processing, the XGBoost model demonstrates improved predictions for the four indicators: TC, FW, MC, and SPAD, with R² values exceeding 0.50, while the R² and RMSE values had a slightly lower accuracy than that of the RF model; however, they outperform SVM and BPNN. C/N, C/P, and N/P had the lowest prediction accuracy, and the R² of both data processing methods is lower than 0.2. After the first-order differential transform, the accuracy of each model built by the XGBoost algorithm improved, among which the LT had the highest improvement, and the R² and RMSE of the original spectral data are 0.18 and 0.09, respectively, while those of the FD spectra are 0.40 and 0.07, and the R² is apparently more improved. Overall, the modelling accuracy of XGBoost is slightly lower than that of RF, and more accurate and stable than SVM and BPNN.

Figure 6. Scatter plots of XGBoost modelling accuracy.

Discussion

Analysis of the feature bands extracted using the importance score method

The high correlation between consecutive bands of hyperspectral data causes an inevitable dimensionality disaster. Therefore, dimensionality reduction of the data is required before constructing the model, which reduces the complexity of the model, improves computational efficiency, and reduces the risk of overfitting (Ding et al., 2020). The importance score method is used to rank the input features of a prediction model according to their relative importance in making predictions. In this study, we used this method to extract feature bands that were strongly correlated to the target parameters from the hyperspectral data, which reduced the dimensionality of the data and eliminated redundancy. We found that the importance score of different leaf functional traits and hyperspectral data differed, which led to variations in the extracted characteristic bands. IIn general, the results of the feature bands screened by the importance value ranking based on the two methods of Random Forest Fand XGBoost exhibited similarities. Both methods screened a small number of bands with high importance scores for FW, TC, and SLA. For SPAD, the selected feature bands screened are concentrated in the visible light zone, while chemical elements such as TN, TP, and ecological stoichiometric characteristics are distributed uniformly throughout the zone. This indicates that the two methods, RF and XGBoost, are consistent in the screening of feature bands and show certain regularity for different features. This provides important guidance and reference for further research and application. While the characteristic bands extracted from the original spectrum were significantly different from those extracted from the FD processing data, the importance score after FD processing was greatly improved compared to that for the original reflectance. This may be because the original hyperspectral data presented certain issues that hindered the extraction of relevant information, such as lack of linear correlation between reflectivity and measured traits and a low signal-to-noise ratio. The original reflectivity information is easily obscured by redundant information; thus, the dimension must be reduced by data conversion. The FD treatment eliminated noise and enhanced the subtle changes in spectral reflectivity (Cui et al., 2021; Liu et al., 2020), which facilitation the extraction of characteristic bands with high-importance scores. However, we used two spectral downscaling methods, XGBoost and RF, respectively, for the subsequent analysis, which may have an impact on the accuracy of other models. In future studies, we will further explore the fusion of multiple spectral dimensionality reduction methods and synthesize the advantages of each method to obtain more reliable and robust feature selection results.

Spectral modelling of functional traits in the leaves of S. alterniflora

In this study, the data of 11 leaf traits of S. alterniflora were modelled uniformly with the corresponding spectra using four methods, namely RF, SVM, BPNN and XGBoost. In the final trait-spectrum models obtained, the accuracy of TC, FW, and MC improved, followed by LT, SLA, and SPAD, and the inversion accuracy of TN, TP, and three ecological chemometric traits was poor. In this study, the leaf water content model was fitted better, and the R² of only two of the eight models established by the two characteristic band screening methods after FD processing was lower than 0.50, while the others were higher than 0.55, which was better than the previous model fit for leaf water content of wetland plants (R² = 0.49; Yao et al., 2022), indicating that the model had some predictive ability. The inverse modelling effect for SPAD of S. alterniflora leaves was poor, only the SVM and BPNN models based on the high importance score bands of random forests had better accuracy, with R² higher than 0.60, and the R² of the rest of the models was below 0.50, which was poorer than the inverse study of SPAD of mangrove leaves (R² >0.70; Dou et al., 2019). This may be due to the different absorption, scattering and transmission characteristics of different plant leaves, coupled with the differences in the environmental conditions in which S. alterniflora grows, which may lead to a more complex and unstable relationship between leaf SPAD values and other environmental factors, thus reducing the effectiveness of the inversion model.

In this study, we also conducted an inversion study on the chemical element content of S. alterniflora leaves, and the results showed that only the TC had a better inversion effect, while the inversion effect of the other indexes was poor. This may be due to the failure to consider proper model selection and algorithm optimization in this inversion study. Previous studies have shown that the PLS model has a better performance in inversion of elemental content (Tang et al., 2022; Wold et al., 2001). In our study, the four regression models were less effective in predicting chemical elemental content, but it does not mean that the chemical elemental content of S. alterniflora leaves could not be predicted by hyperspectral data, and it is necessary to continue to explore more model combinations and parameter optimization. Therefore, we will further explore the performance of different models in the inversion of chemical elemental content and select more appropriate algorithms and models for experiments to improve the inversion effect and accuracy in future studies.

As a whole, when we divided the dataset, we divided it into training and testing sets separately, which may lead to a large difference in data distribution between the training and testing sets. Thus, we obtained a model with a better accuracy during training and not achieved the expected results during testing due to the inability of the testing dataset to adapt to the model, which is one of the larger decreases in the validation accuracy of this study compared to the modelling accuracy possible reason (Große‐Stoltenberg et al., 2018).

Analysis and comparison of inverse studies on functional traits in S. alterniflora

Currently, numerous studies have inverted the functional characteristics of S. alterniflora, as shown in Table 4. For example, Chen et al. (2022) used a generative adversarial network with a constrained factor (GAN-CF) model and hyperspectral data to invert the ground biomass of S. alterniflora and obtained a good inversion effect. Similarly, González Trilla et al. (2013) predicted the biomass, leaf area index, percent canopy cover, and other biophysical indicators of S. alterniflora in the Bahia Blanca estuary, Argentina, based on the relationship between spectral reflectance data and the normalized vegetation index and achieved good prediction results. Zhu et al. (2020) used vegetation indices to predict the crude protein, fibre, and fat contents of S. alterniflora leaves based on hyperspectral data, and the findings confirmed the applicability of these data in the inversion of S. alterniflora characteristics. Furthermore, Ai et al. (2015) used the correlation between the spectral and vegetation indices to invert the photosynthetic pigment content of S. alterniflora leaves in the Minjiang estuary wetland and found that FD spectral data had a good prediction effect on the photosynthetic pigment content of S. alterniflora leaves.

Table 4. Inversion of the functional traits.

Medium	Index	Accuracy assessment	Reference
Generative adversarial network with a constrained factor (GAN-CF)	Aboveground biomass (AGB)	RMSE = 2.07 MAE = 1.85	Chen et al. (2022)
Modified chlorophyll absorption in reflectance index (MCARI)	Biomass	R^2 = 0.63	González Trilla et al. (2013)
	Percentage canopy cover (PCC)	R^2 = 0.75
	LAI	R^2 = 0.74
Vegetation index	Crude protein	RMSEP = 1.464 PRD = 2.690	Zhu (2023)
	Crude fibre	RMSEP = 2.322 PRD = 1.857
	Crude fat	RMSEP = 0.608 PRD = 1.794
FDNDVI [723, 525]	Chla	R^2 = 0.6037 RMSEP = 0.3425	Ai et al. (2015)
	Chlb	R^2 = 0.6246 RMSEP = 0.1225
	Cars	R^2 = 0.6058 RMSEP = 0.0692

In this study, we obtained the importance scores of 11 leaf functional traits from hyperspectral data of S. alterniflora leaves. Using these scores, we established inversion models for the leaf functional traits using the RF importance score based on the original and FD spectral data and the RF, SVM, XGboost, and BPNN models. Our results showed that the SVM model had good prediction ability for the functional traits of S. alterniflora leaves. Previous studies on S. alterniflora have focused more on component content traits such as chlorophyll and protein, paying comparatively less attention on elemental content and morphological structural traits, specifically at the leaf scale. While the accuracy of the present study is slightly lower than theirs, it distinguishes itself by employing a systematic approach that relies on four machine learning models to study morphological structural traits (e.g. FW, LT, SLA) and component content traits (e.g. C, N, P), which provides ideas for studying the functional traits of S. alterniflora. However, we acknowledge some limitations in this study, such as the lack of analysis of the best inversion model for some functional traits, which should be further explored in future research.

Conclusion

Our study focused on the hyperspectral inversion of functional traits of S. alterniflora leaves from the Yancheng coastal wetland, Jiangsu Province, China. We identified the characteristic bands between original spectrum and FD hyperspectral data and the functional traits of each leaf using the RF and XGBoost importance score. Inversion models for FW, LT, SPAD, TN, TP, TC, MC, SLA, N/P, C/N, and C/P of S. alterniflora were established, and the prediction results of XGBoost, BPNN, SVM, and RF were compared to identify the optimal data type and model. The following conclusions were drawn.

1. The validation results demonstrated that the SVM model built based on the feature bands screened by RF importance score outperformed the remaining three models in terms of accuracy and stability for inverting the functional traits of S. alterniflora leaves.

2. In terms of carbon, nitrogen, and phosphorus and their ecological chemometric traits, only the model of TC had predictability, and the model built based on the XGBoost importance score method showed better performance.

3. Using the extracted characteristic bands from the original spectral and FD hyperspectral data, prediction models for different functional traits were developed and compared. The results showed that the FD hyperspectral data-based models generally outperformed the models based on the original data for most functional traits.

Acknowledgments

We would like to express our gratitude to Deputy Director Chen Hao, Section Chief Zhang Yanan and Mr. Li Zhenghao of the reserve for their valuable assistance in conducting our field experiments.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Hyperspectral data and functional trait data used in this study are available from the corresponding author upon request.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/22797254.2023.2294951

Additional information

Funding

This work was supported by the National Key R&D Program of China under Grant [2017YFC0506200].