Deep Learning-Based Prediction of Multi-Species Leaf Pigment Content Using Hyperspectral Reflectance

Highlights

What are the main findings?

• CNN models combined with genetic algorithm–based spectral band selection achieved high-accuracy estimation of leaf pigment content across tree species.
• The 2D CNN outperformed the 1D CNN, with optimal results obtained using 3–4 convolutional layers.

What is the implication of the main finding?

• The study provides a non-destructive and robust approach for monitoring leaf pigments across different tree species.
• The CNN-based approach improved remote sensing applications in vegetation health assessment and forest ecosystem management.

Abstract

Leaf pigment composition and concentration are crucial indicators of plant physiological status, photosynthetic capacity, and overall ecosystem health. While spectroscopy techniques show promise for monitoring vegetation growth, phenology, and stress, accurately estimating leaf pigments remains challenging due to the complex reflectance properties across diverse tree species. This study introduces a novel approach using a two-dimensional convolutional neural network (2D-CNN) coupled with a genetic algorithm (GA) to predict leaf pigment content, including chlorophyll a and b content (C_ab), carotenoid content (C_ar), and anthocyanin content (C_anth). Leaf reflectance and biochemical content measurements taken from 28 tree species were used in this study. The reflectance spectra ranging from 400 nm to 800 nm were encoded as 2D matrices with different sizes to train the 2D-CNN and compared with the one-dimensional convolutional neural network (1D-CNN). The results show that the 2D-CNN model (nRMSE = 11.71–31.58%) achieved higher accuracy than the 1D-CNN model (nRMSE = 12.79–55.34%) in predicting leaf pigment contents. For the 2D-CNN models, C_ab achieved the best estimation accuracy with an nRMSE value of 11.71% (R² = 0.92, RMSE = 6.10 µg/cm²), followed by C_ar (R² = 0.84, RMSE = 1.03 µg/cm², nRMSE = 12.29%) and C_anth (R² = 0.89, RMSE = 0.35 µg/cm², nRMSE = 31.58%). Both 1D-CNN and 2D-CNN models coupled with GA using a subset of the spectrum produced higher prediction accuracy in all pigments than those using the full spectrum. Additionally, the generalization of 2D-CNN is higher than that of 1D-CNN. This study highlights the potential of 2D-CNN approaches for accurate prediction of leaf pigment content from spectral reflectance data, offering a promising tool for advanced vegetation monitoring.

1. Introduction

The production of oxygen and organic matter produced by photosynthesis is of tremendous significance for the global carbon cycle in the biosphere [1]. Photosynthesis is primarily driven by leaf pigments, which serve as excellent indicators of plant growth, stress, function, and phenology from local to global scales [2,3]. The three main families of pigments are chlorophylls, carotenoids, and anthocyanins. Chlorophylls, including chlorophyll-a and chlorophyll-b, are the fundamental sunlight energy-absorbing pigments involved in photosynthetic reactions [4]. Carotenoids, composed of carotenes and xanthophylls, contribute to light harvesting and provide essential photoprotective effects, preventing excess energy from damaging the photosynthetic system [5]. Anthocyanins are strongly associated with foliage coloration and protect leaves from excess light and ultraviolet radiation [6]. Variation in pigment content thus provides information concerning leaf physiological state.

However, the ability to extract pigment content quickly and accurately is still limited. Leaf pigment content estimation is mainly based on chemical and remote sensing methods [7]. Traditional wet chemical pigment analysis requires the destruction of the detached leaves, followed by solvent extraction and spectrophotometric analysis according to a standard procedure [8,9]. These wet laboratory methods are time-consuming and labor-intensive. Furthermore, the destructiveness of traditional techniques limits the ability to monitor the change in pigments over time for the individual leaf [5,10]. In contrast, remote sensing-based approaches, which measure spectral reflectance, are nondestructive, rapid, and applicable across various spatiotemporal scales [11].

Leaf reflectance spectra are directly influenced by the spectral absorbance properties of pigments [10,12]. Therefore, the measurement of leaf reflectance can offer the opportunity to quantify pigment content [3]. Remote sensing-based pigment estimation generally follows two approaches, i.e., data-driven and radiative transfer model (RTM) inversion [13]. Among data-driven approaches, spectral index-based statistical models are widely used to predict leaf pigment content. The vegetation indices (VIs) were calculated from the reflectance of wavelengths related to pigment absorption features located in the visible (VIS) region (400–800 nm) [14]. To date, most studies on leaf pigments have focused on chlorophylls [15,16,17], while fewer have concentrated on the estimation of carotenoids [18,19,20] and anthocyanins [21,22]. More specifically, Croft et al. (2014) comprehensively summarized 47 VIs that are sensitive to chlorophyll at both the leaf and canopy scales [23]. It was concluded that the estimation of leaf chlorophyll content can achieve comparable results using simple statistical models based on the linear relationship between chlorophyll and VIs. However, overlapping absorption coefficients of pigments in the VIS region make it difficult to separate and quantify these pigments using VI-based linear models [12]. Actually, the relationships between the reflectance in the VIS region and pigment contents are inherently nonlinear. Consequently, using a mathematical formulation derived from VIs to calculate the pigment content can sometimes be inaccurate [24]. Secondly, multivariate statistical models such as partial least squares regression (PLSR) [25,26] have also been extensively used to estimate leaf pigment content. In PLSR, the full range of reflectance was transformed into several latent variables, which were then used to estimate leaf pigment content [27]. Furthermore, machine learning (ML) algorithms based on nonparametric approaches, e.g., neural networks (NNs), random forest regression (RFR), and Gaussian processes regression (GPR), perform better on pigment estimation [28,29,30]. However, the performance of these models depended mainly on the trained spectral dataset and pre-processing techniques. ML may lose universality when it is applied to another new dataset with the variation in feature spaces and distributions from different plant species and environmental conditions [31].

Another approach is based on the physical principle through the inversion of RTMs. RTMs such as PROSPECT [32] and LIBERTY [33] are widely used to simulate directional–hemispherical reflectance over the wavelength across 400 to 2500 nm at the leaf scale. The LIBERTY model was developed to represent the optical properties of needles. However, the generalization of the LIBERTY model is limited because measuring the spectrum of narrow needles is challenging. The PROSPECT model has evolved through several versions. PROSPECT-4 and 5 [34] introduced the separation of photosynthetic pigments (i.e., chlorophyll and carotenoid) in leaf optical property models. Subsequently, a new anthocyanin pigment was included in PROSPCT-D [13]. Recently, Féret et al. (2021) developed PROSPECT-PRO [35], which separated proteins from carbon-based constituents. This advancement brings leaf optical model simulations closer to natural leaf spectra. However, increased model input parameters can lead to ill-posed inversion problems because various combinations of model parameters can compensate for each other, resulting in leaf reflectance with very similar optical properties [36].

With the advancement of modern computational capabilities, deep learning-based algorithms have been successfully applied in the field of computer vision. These methods generally employ neural network architectures with multiple processing layers to extract higher-level features from input data. Convolutional neural networks (CNNs) are among the most widely used deep learning algorithms, achieving high performance in image classification and pattern recognition [37]. Recently, CNNs have also been proven effective in addressing regression problems in the remote sensing community, such as agricultural yield prediction [38] and the estimation of forest structural parameters [39]. Moreover, an emerging trend in remote sensing research highlights the integration of multi-sensor data (e.g., RGB cameras, hyperspectral, and LiDAR) with artificial intelligence methods (e.g., CNNs), which has shown great potential for non-destructive, precise, and efficient quantification of plant physiological and biochemical traits [40,41,42]. Traditionally, CNNs are applied to two-dimensional (2D) image data. By contrast, leaf reflectance spectra are often represented as one-dimensional (1D) sequential signals [43]. Several recent studies have therefore developed 1D-CNNs with leaf reflectance to estimate leaf pigments [44,45,46,47]. However, the contrast between absorption peaks and their surroundings tends to be smoothed when only sequential spectra are considered, particularly across different vegetation species, limiting the model’s ability to capture important spectral features.

This paper aims to develop a CNN-based approach for accurately predicting leaf pigment content across different tree species using reflectance spectra. We propose two CNN-based approaches to simultaneously estimate chlorophyll, carotenoid, and anthocyanin from the leaf reflectance spectra: (1) a 1D-CNN applied directly to spectral signals, and (2) a 2D-CNN applied to spectral signals transformed into 2D matrices. The specific objectives are to (1) investigate leaf spectral diversity along with the relationship between pigment content and reflectance spectra across various tree species; (2) compare the performance of 1D-CNN models and 2D-CNN models; (3) evaluate model accuracy and robustness across different tree species.

The structure of this paper is arranged as follows: leaf reflectance spectra and leaf pigment content measurements, as well as CNN-based methodologies, are described in Section 2; the comparison results between 1D- and 2D-CNN are described in Section 3; the discussion and conclusions are given in Section 4 and Section 5, respectively.

2. Materials and Methods

2.1. Study Area and Field Collection

The study site (see Figure 1) is located in Hangzhou (118°21′–120°30′E, 29°11′–30°33′N), Zhejiang Province, China. The region has a subtropical monsoon climate, with a mean annual temperature of 17.8 °C, a mean relative humidity of 70.3%, and annual precipitation of 1454mm. The elevation ranges from 3 m to 1587 m.

This study was conducted between April and September from 2021 to 2023, during which 2094 leaf samples were collected from 28 broadleaf tree species (e.g., Diospyros oleifera, see

Table 1. Tree species measured in this study (n = 2094).

Tree Species	Number of Samples	Tree Species	Number of Samples
Osmanthus fragrans	74	Liquidambar formosana	74
Diospyros oleifera	80	Acer cinnamomifolium	75
Sapindus mukorossi	68	Eriobotrya japonica	76
Phoebe sheareri	74	Photinia serratifolia	77
Ginkgo biloba	75	Michelia figo	74
Glochidion puberum	77	Lagerstroemia indica	77
Ilex chinensis	74	Prunus serrulata var. lannesiana	70
Carya illinoinensis	72	Gardenia jasminoides	68
Quercus glauca	78	Firmiana simplex	74
Acer buergerianum Miq.	79	Liriodendron chinense	76
Ligustrum lucidum	80	Yulania denudata	77
Cinnamomum camphora	73	Magnolia grandiflora	73
Ailanthus altissima	75	Albizia julibrissin	76
Viburnum macrocephalum	74	Styphnolobium japonicum	74

for the complete list). For each species, several leaf samples were obtained from different individual trees to ensure diversity. During the sample collection process, we only selected healthy leaves to represent the physiological characteristics of the growing season. The sampling of leaves followed consistent standards in different regions, with leaves collected from different heights and orientations within the tree crown. The collected samples were labeled and stored in an ice box to prevent discoloration before pigment analysis in the laboratory.

2.2. Spectra and Leaf Pigment Content Measurements

Leaf hemispherical reflectance spectra were measured using an ASD FieldSpec-4 Wide-Res (Analytical Spectra Devices, Inc., Boulder, CO, USA) handheld spectrometer equipped with an integrating sphere. This instrument measures spectra from 350 to 2500 nm with a spectral resolution of 1 nm. The radiometer was calibrated and optimized for dark current and system offsets using a white reference panel. To reduce instrument noise, each leaf sample was measured ten times, and then the results were averaged to obtain the mean reflectance. Finally, the mean reflectance of each sample was calculated from all collected leaves within a tree crown.

After the spectral measurements, leaf pigments, including chlorophyll a and b content (C_ab), carotenoid content (C_ar), and anthocyanin content (C_anth), were measured immediately in the laboratory. Each sample was cut into eight small leaf disks (approximately 0.5 g) with a diameter of 1.2 cm. Two types of organic solvents were used for pigment extraction, including 80% acetone for C_ab and C_ar, and a mixture of methanol, concentrated hydrochloric acid, and distilled water for C_anth. Leaf disks were placed in tubes containing organic solvents (25 mL) and kept in a dark environment for 48 h to dissolve pigments completely. A dual beam scanning UV-2100 Spectrophotometer was utilized to measure the supernatant absorbance at wavelengths of 470 nm, 537 nm, 647 nm, and 663 nm. The pigments were derived using Equations (1)–(4). Finally, C_ab, C_ar, and C_anth expressed in the same unit (µg/cm²) were determined using a multi-wavelength calculation, according to [10,48].

$$C_a=12.25{\mathrm{A}}_{663}-2.79{\mathrm{A}}_{647}$$

(1)

$$C_b=21.50{\mathrm{A}}_{647}-5.10{\mathrm{A}}_{663}$$

(2)

$$C_{ar}=(1000{\mathrm{A}}_{470}-1.82C_a-85.02C_b)/198$$

(3)

$$C_{anth}=36.71{\mathrm{A}}_{537}-3.13{\mathrm{A}}_{646}-{ \mathrm{A}}_{663}$$

(4)

2.3. Model Development

2.3.1. Data Preparation

The reflectance spectra were processed to reduce noise and enhance model performance. Wavelengths with a low signal-to-noise ratio, from 350 nm to 400 nm, were excluded. Then, the wavelength ranges from 400 nm to 800 nm, which are strongly related to pigment spectral absorption, were selected to estimate leaf pigment content.

In addition, all leaf pigment contents were normalized due to the significant differences in the amounts of three pigments, with C_ab typically being more than ten times higher than C_ar and C_anth. This normalization ensures that each neuron (filter) in the CNN has an equal opportunity to learn the pattern for each pigment [44]. The raw content of three pigments was converted into a 0–1 range by

$$P_i^{*}={\displaystyle \frac{P_i-\min (P_i)}{\max (P_i)-\min (P_i)}}$$

(5)

where P_i and $P_i^{*}$ are the original and normalized values of each leaf pigment, respectively.

2.3.2. Design CNN Architecture

In this study, two CNN-based approaches were developed to estimate leaf pigment contents. CNN-based strategies have proven effective in the processing of both one-dimensional (1D) signals (e.g., time series) and two-dimensional (2D) images, providing accurate estimation through a hierarchical architecture [49,50]. Recent research has shown that AlexNet performs well in pigment prediction tasks [44,45,46,47]. Therefore, the architecture of AlexNet was referenced in this study. The original architecture of AlexNet included five convolutional layers and three fully connected layers. It also employs rectified Linear Unit activations (ReLU), dropout regularization, and overlapping max-pooling to achieve high efficiency and good generalization ability.

(1) 1D-CNN-based approach

For the 1D-CNN, the reflectance spectra (wavelength range: 400–800 nm) with the size of 1 × 400 were considered as a waveform input, and the output layer with the size of 1 × 3 was the prediction of three pigment contents in the 1D-CNN architecture (see Figure 2). In each convolution operation, the feature map of spectral data was extracted from the convolutional filter, with feature sizes determined by the filter (kernel) sizes, stride, and padding. Considering the narrow wavelength range and data length used in this study, the kernel size was set to 1 × 5 and 1 × 3 for the first and other convolutional layers, respectively. To avoid feature dimensions and to minimize network overfitting, the dropout method was used instead of max-pooling in the 1D-CNN architecture [43]. Each convolutional layer employed the same padding with a stride of 1. After all the convolution operations, the first two fully connected layers combined all the features. The last fully connected layer performed multi-output regression with a linear activation function to predict pigment variables. In addition, the number of filters and neurons in each layer was set according to AlexNet and other similar research [39,43], which follows the general principles applied in the design of convolutional neural networks [51]. The parameters of the 1D-CNN are summarized in

Table 2. The parameters of the one-dimensional convolutional neural network (1D-CNN).

Layer Type	Size	Filters	Activation Function
Input layer	1 × 400	-	-
Conv1	1 × 5	96	ReLU
Dropout (0.5)	-	-	-
Conv2	1 × 3	96	ReLU
Dropout (0.5)	-	-	-
Conv3	1 × 3	192	ReLU
Conv4	1 × 3	192	ReLU
Conv5	1 × 3	256	ReLU
Dropout (0.5)	-	-	-
Fully1	1 × 4096	-	ReLU
Fully2	1 × 4096	-	ReLU
Fully3	1 × 3	-	Linear

.

(2) 2D-CNN-based approach

To analyze the spectral data using 2D-CNN, each waveform spectrum must be encoded as a 2D matrix format for feeding as input to the 2D-CNN. Several studies have addressed the transformation of 1D reflectance spectra to 2D representations using spectrograms [50,52]. However, these transformations did not achieve better accuracy in parameter estimation because the frequency count with the short-time Fourier transform altered the original spectra arrangement. Here, we proposed a simple and effective 2D transformation for reflectance spectra data.

The spectral reflectance was transformed into a 2D representation by rearranging the elements of the 1D reflectance spectrum in a 2D matrix. The 2D matrix dimensions were set to 8 × 50, 10 × 40, 20 × 20, 40 × 10, and 50 × 8. The rearrangement process is performed by dividing the reflectance data into vectors of equal horizontal size and then stacking this information together vertically. For comparison purposes, the 2D-CNN architecture (Figure 3) was designed to be similar to that of the 1D-CNN model. Batch normalization was applied to the convolutional layer features to accelerate convergence speed and improve model stability [53]. Additionally, the nonlinear activation function ReLU and the Dropout method were used for 2D-CNN. The first two fully connected layers flattened and combined all features extracted from the convolutional kernels. Finally, similarly to 1D-CNN architectures, the last fully connected layer performed multi-output regressions with three neurons representing C_ab, C_ar, and C_anth, respectively. The parameters of the 2D-CNN are summarized in

Table 3. The parameters of the two-dimensional convolutional neural network (2D-CNN) model.

Layer Type	Size	Filters	Activation Function
Layers	20 × 20
Conv1	5 × 5	96	ReLU
BatchNormalization ()	-	-	-
Dropout (0.5)	-	-	-
Conv2	3 × 3	96	ReLU
BatchNormalization ()	-	-	-
Dropout (0.5)	-	-	-
Conv3	3 × 3	192	ReLU
BatchNormalization ()			-
Conv4	3 × 3	192	ReLU
BatchNormalization ()
Conv5	3 × 3	256	ReLU
BatchNormalization ()	-	-	-
Dropout (0.5)	-	-	-
Fully1	1 × 4096	-	ReLU
Fully2	1 × 4096	-	ReLU
Fully3	1 × 3	-	Linear

.

(3) Optimization methods

In CNN-based methods, features are extracted from spectral reflectance using convolutional layers, which strongly influence leaf pigment estimation performance [46]. Therefore, the number of convolutional layers (N_conv) is an important factor in assessing prediction accuracy. Five strategies with an increasing number of convolutional layers were implemented in 1D- and 2D-CNN to optimize feature extraction for pigment prediction. The architectures of these five strategies are shown in

Table 4. Five strategies of the CNN model with different numbers of convolutional layers for leaf pigment estimation.

Strategies	A	B	C	D	E
Layers	Input layer	Input layer	Input layer	Input layer	Input layer
	Conv1	Conv1	Conv1	Conv1	Conv1
	Dropout	Dropout	Dropout	Dropout	Dropout
	-	Conv2	Conv2	Conv2	Conv2
	-	Dropout	Dropout	Dropout	Dropout
	-	-	Conv3	Conv3	Conv3
	-	-	-	Conv4	Conv4
	-	-	-	-	Conv5
	-	-	-	-	Dropout
	Fully1	Fully1	Fully1	Fully1	Fully1
	Fully2	Fully2	Fully2	Fully2	Fully2
	Fully3	Fully3	Fully3	Fully3	Fully3

.

We trained the aforementioned CNN architectures using stochastic gradient descent with momentum (SGDM), while internal neuron weights were updated with the Adam optimizer. During backpropagation, the mean absolute error (MAE) was used as the cost function. Therefore, the loss function was defined as follows:

$$\mathrm{Loss} ={\displaystyle \frac{1}{N}}{\sum }_{j=1}^N{\displaystyle \frac{1}{R}}{\sum }_{i=1}^R\left|y_i-\widehat{y_i}\right|$$

(6)

where $y_i$and $\widehat{y_i}$represent the measured pigment content and the predicted pigment content, respectively; N is the number of observations, and R is the number of responses.

In addition, the standard hyperparameters like learning rate and batch size were set to 0.001 and 32, respectively. An epoch of 2000 was used to train for both 1D- and 2D-CNN models.

2.3.3. Spectral Bands Selection Using Genetic Algorithm

To reduce computational time and avoid under- and overfitting, the Genetic algorithm (GA) was coupled with CNN models for spectral feature selection [54,55]. We employed a pre-optimized 1D-CNN model for fitness evaluation of GA (named as 1D-CNN_GA) to select the important spectral bands in the prediction of leaf pigment contents. In this study, CNN models adopt the multi-output regression for three pigment estimations. Therefore, the sensitive bands selected by GA represent the combined influence of three pigments on leaf spectral reflectance. For the GA coupled with a 2D-CNN model (named as 2D-CNN_GA), the rearrangement of the 2D matrix for spectral reflectance was the same as in the original 2D-CNN with full spectrum (referred to as 2D-CNN_f). Pixel values corresponding to selected importance bands were filled with measured reflectance, while pixels for non-selected bands were set to zero.

The standard GA procedures included encoding (waveband reflectance was translated into binary strings (named as chromosomes)), initial population (n = 100), calculation of fitness, selection, crossover, mutation, and fitness evaluation. After evaluation, a 500 iteration search for GA was implemented to obtain the best performing chromosomes over several generations.

2.4. Accuracy Assessment

In this study, the measured leaf dataset (n = 2094) was used to evaluate the performance of four variants of CNN models, which are mentioned above. To assess the performance of the four versions of CNNs and five strategies, the dataset was divided into a training set (70%) and a validation set (30%) using random systematic sampling based on tree species and collection month to ensure model generalization. For comparison, three benchmark estimation models, including PLSR, RFR, and GPR, were employed as standard competitors to further evaluate the effectiveness of CNN models. During training, the optimal number of latent variables used in the PLSR was determined by identifying the number that resulted in the smallest root mean squared error (RMSE) using the cross-validation approach (5-fold) to create the linear models. Hyperparameter optimization for RFR and GPR was implemented using a grid search method.

All models (see

Table 5. List of tested models.

Models	Data Used
1D-CNNf	full spectrum (400–800 nm)
1D-CNNGA	subset spectrum (400–800 nm)
2D-CNNf	full spectrum (400–800 nm) in 2D representation
2D-CNNGA	subset spectrum (400–800 nm) in 2D representation
PLSRf	full spectrum (400–800 nm)
PLSRGA	subset spectrum (400–800 nm)
RFRf	full spectrum (400–800 nm)
RFRGA	subset spectrum (400–800 nm)
GPRf	full spectrum (400–800 nm)
GPRGA	subset spectrum (400–800 nm)

for details) were trained and tested using both full-spectrum reflectance and the subset of important bands selected by GA. The accuracy of leaf pigment content prediction was evaluated using the coefficient of determination (R²), RMSE, and normalized root mean square error percentage (nRMSE). Finally, a t-test was conducted to assess whether there were significant differences in prediction accuracy between the CNN models and the three benchmark models.

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

(7)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}-{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}$$

(8)

$$\mathrm{nRMSE}=100 \times {\displaystyle \frac{RMSE}{\stackrel{\mathrm{-}}{y}}}$$

(9)

where $\stackrel{\mathrm{-}}{y}$ is the mean of n measured pigment contents. The nRMSE was used to compare the performances across four variants of CNN models.

For further validation of the effectiveness and generalizability of the optimal 1D- and 2D-CNN architectures (with the best number of convolutional layers), we incorporated the internationally available LOPEX93 leaf spectral dataset, which provides leaf optical properties for 331 samples from 45 species across wavelengths of 400–2500 nm [56]. Since C_anth measurements are not included in this dataset, only C_ab and C_ar were used to validate the optimal 1D- and 2D-CNN model. The LOPEX93 dataset was randomly divided by plant type into training (70%) and validation (30%) subsets. The LOPEX93 training subset was then combined with the training dataset from this study to train models, while the remaining LOPEX93 validation subsets, together with our study’s validation dataset, were used to evaluate model performance in estimating C_ab and C_ar.

3. Results

3.1. The Distribution of Leaf Pigment Content and Spectral Variation

The statistics of the three pigment contents from 2094 samples across 28 tree species are given in

Table 6. Statistics of measured leaf pigment contents (n = 2094).

Pigments	Mean	SD	Max	Min	Range
Cab (µg/cm2)	52.51	21.78	97.73	15.50	82.23
Car (µg/cm2)	8.33	2.49	14.36	4.15	10.21
Canth (µg/cm2)	1.18	1.15	7.18	0.02	7.16

Note: standard deviation (SD), maximum (Max), and minimum (Min).

. C_ab values showed a wide range (15.50–97.73 µg/cm²) with a mean of 52.51 µg/cm² and a standard deviation (SD) of 21.78 µg/cm². In contrast, the variation in C_ar and C_anth was relatively small, with a range value of 10.21 µg/cm² (SD: 2.49 µg/cm²) and 7.16 µg/cm² (SD: 1.15 µg/cm²), respectively. The frequency distribution of C_ab, C_ar, and C_anth is shown in Figure 4. The distribution of C_ab and C_ar was normal. However, as for C_anth, most of the samples are distributed in the low value region (<2 µg/cm²).

Figure 5 shows the average reflectance of 28 tree species. Due to differences in tree species and pigment content, substantial spectral variation was observed across the 400–800 nm wavelength range. The significant differences in reflectance among different tree species occurred in the blue light region (400–500 nm), green light region (500–600 nm), and near-infrared region (760–800 nm). However, a relatively low variation was found in red valley (650–700 nm) and red-edge (700–760 nm).

3.2. Importance Bands for Estimating Leaf Pigments

The importance bands (Figure 6) were derived using the 1D-CNN model coupled with GA for leaf pigment content estimation. A total of 218 bands was selected by GA from the original 400 bands across the full wavelength (400–800 nm) region. Most of the sensitive spectral bands (accounting for 75%) are distributed within the absorption wavelength ranges of the three pigments, located in the blue (430–500 nm), green (520–600 nm), and red regions (640–680 nm). Specifically, the broad and continuous important bands were located around 430–440 nm, 465–485 nm, 510–515 nm, 525–540 nm, 545–555 nm, 575–580 nm, 620–630 nm, 655–660 nm, 680–700 nm, and 705–715 nm. Additionally, some important bands (about 10%) were found outside the pigments’ absorption ranges, at wavelengths above 750 nm.

3.3. Model Performance for the Prediction of Leaf Pigment Content

The performance of 1D-CNN models for leaf pigment content prediction is summarized in

Table 7. Accuracy of pigment estimation using 1D-CNN models with different numbers of convolutional layers.

Models	Strategies (Nconv)	Cab			Car			Canth
		R2	RMSE (µg/cm2)	nRMSE (%)	R2	RMSE (µg/cm2)	nRMSE (%)	R2	RMSE (µg/cm2)	nRMSE (%)
1D-CNNf	A (1)	0.88	7.67	14.72	0.70	1.47	17.52	0.05	1.05	93.51
	B (2)	0.89	7.38	14.16	0.73	1.37	16.31	0.46	0.86	76.88
	C (3)	0.90	7.41	14.22	0.75	1.29	15.36	0.54	0.78	69.48
	D (4)	0.91	6.75	12.94	0.77	1.25	14.88	0.68	0.63	56.13
	E (5)	0.90	14.25	27.34	0.58	1.96	23.33	0.23	1.00	89.13
1D-CNNGA	A (1)	0.91	6.82	13.08	0.78	1.28	15.28	0.11	1.02	90.87
	B (2)	0.91	6.67	12.81	0.78	1.25	14.89	0.55	0.79	71.22
	C (3)	0.91	6.74	12.93	0.80	1.17	14.02	0.59	0.72	64.53
	D (4)	0.92	6.66	12.79	0.81	1.13	13.53	0.72	0.62	55.34
	E (5)	0.91	6.70	12.85	0.80	1.19	14.26	0.51	0.78	69.63

Note The bold represents the best strategy for each model.

. This table shows three assessment indicators, including R², RMSE, and nRMSE of 1D-CNN_f and 1D-CNN_GA for the prediction of C_ab, C_ar, and C_anth. For 1D-CNN_f, the prediction accuracy improved as the number of convolutional layers increased from 1 to 4. Specifically, C_ab achieved nRMSE = 12.94–14.72% (R² = 0.88–0.91, RMSE = 6.75–7.67 µg/cm²), C_ar showed nRMSE = 14.88–17.52% (R² = 0.70–0.77, RMSE = 1.25–1.47 µg/cm²), and C_anth achieved nRMSE = 56.13–93.51% (R² = 0.05–0.68, RMSE = 0.63–1.05 µg/cm²). Compared with 1D-CNN_f, 1D-CNN_GA exhibited more stable prediction accuracy, particularly for C_ab and C_ar. For both 1D-CNN_f and 1D-CNN_GA, strategy D with four convolutional layers provided the best estimate for all pigments. Figure 7 illustrates the 1:1 relationship between measured and predicted values of the three pigments using the optimal 1D-CNN models with four convolutional layers. In terms of nRMSE, the 1D-CNN_GA model (using the subset spectrum) slightly outperformed the 1D-CNNf model (using the full spectrum), with reductions in nRMSE ranging from 0.15% to 1.35% for all pigments. For both 1D-CNN_f and 1D-CNN_GA models, C_ab (nRMSE = 12.94% or 12.79%) was predicted more accurately than C_ar (nRMSE = 14.88% or 13.53%) and C_anth (nRMSE = 56.13% or 55.34%), with C_anth values above 2 µg/cm² mostly underestimated. Additionally, the nRMSE (%) of pigment estimation using the optimal 1D-CNN models with four convolutional layers across tree species is presented in

Table A1. nRMSE (%) of pigment estimation using optimal 1D- and 2D-CNN models with four convolutional layers for each tree species.

Tree Species	1D-CNNf			1D-CNNGA			2D-CNNf			2D-CNNGA
	Cab	Car	Canth	Cab	Car	Canth	Cab	Car	Canth	Cab	Car	Canth
Osmanthus fragrans	9.77	12.83	56.55	12.71	12.21	55.95	9.71	13.56	36.43	11.67	11.89	26.42
Diospyros oleifera	13.91	15.00	77.57	11.37	14.71	71.85	14.53	11.10	69.20	11.96	11.11	50.70
Sapindus mukorossi	12.29	11.45	40.07	10.87	11.82	45.35	11.69	9.70	46.03	12.79	9.06	36.12
Phoebe sheareri	15.43	11.63	55.61	18.90	12.14	61.06	13.79	9.13	45.56	15.44	9.47	44.67
Ginkgo biloba	12.22	15.07	45.48	11.84	13.92	47.29	9.84	10.89	27.21	10.38	12.94	25.87
Glochidion puberum	10.79	18.10	47.67	12.45	15.90	59.52	11.02	15.05	41.07	11.92	13.65	28.81
Ilex chinensis	12.02	14.02	66.66	11.96	12.73	56.75	12.45	9.75	54.75	9.91	13.90	47.11
Carya illinoinensis	18.71	14.61	79.00	14.42	13.40	91.82	14.93	13.31	64.85	13.66	11.29	43.03
Quercus glauca	11.56	18.85	39.89	12.16	18.55	51.23	12.07	8.74	45.12	10.27	15.79	29.62
Acer buergerianum Miq.	11.11	16.36	52.12	10.23	15.97	52.79	9.88	14.32	35.03	8.88	14.15	24.80
Ligustrum lucidum	10.35	13.08	40.86	8.89	13.07	50.21	12.68	12.17	23.70	9.57	9.90	17.53
Cinnamomum camphora	11.79	14.45	53.66	12.75	14.18	52.80	10.65	14.85	70.31	10.64	9.44	52.19
Ailanthus altissima	13.44	16.00	63.12	12.97	14.64	65.46	15.68	18.19	48.20	12.99	11.42	48.58
Viburnum macrocephalum	11.00	17.94	83.19	9.77	15.62	52.77	12.50	11.29	29.35	9.89	12.20	26.78
Liquidambar formosana	12.20	18.10	66.15	14.42	16.86	72.38	10.64	16.82	56.04	9.97	13.77	33.91
Acer cinnamomifolium	9.84	12.46	50.17	11.76	12.28	52.72	10.72	13.85	87.68	9.58	10.89	39.35
Eriobotrya japonica	12.55	12.43	67.32	14.95	11.41	74.86	13.99	9.07	54.47	12.15	9.91	41.20
Photinia serratifolia	15.23	14.22	39.22	11.76	11.53	37.40	15.80	13.14	43.97	12.75	11.13	27.60
Michelia figo	13.51	14.27	86.07	14.52	13.24	45.56	18.05	11.12	37.84	11.70	13.78	35.73
Lagerstroemia indica	10.90	14.91	40.17	11.10	13.00	46.01	9.42	12.64	44.84	10.19	11.12	31.09
Prunus serrulata var. lannesiana	17.68	15.57	57.43	9.87	12.86	54.80	13.91	9.04	32.40	12.21	10.70	33.67
Gardenia jasminoides	10.11	13.45	36.39	10.50	13.56	42.89	9.93	13.68	36.94	10.44	10.16	21.91
Firmiana simplex	14.62	13.87	77.83	13.86	13.37	76.80	12.85	14.49	71.57	12.44	12.05	38.60
Liriodendron chinense	15.72	12.68	38.69	16.21	13.59	47.55	15.17	12.61	22.07	16.03	13.76	15.56
Yulania denudata	10.68	12.83	45.47	11.38	11.50	48.49	11.13	14.46	41.10	10.31	12.01	25.47
Magnolia grandiflora	12.91	16.17	29.67	14.29	15.14	28.52	10.71	13.95	31.20	12.46	12.96	26.88
Albizia julibrissin	16.43	14.38	42.27	18.03	12.76	50.23	12.21	11.46	55.66	14.71	19.09	21.07
Styphnolobium japonicum	9.95	14.04	35.31	7.50	12.56	36.87	10.51	12.80	38.27	8.98	11.68	32.15
Mean	12.74	14.60	54.06	12.55	13.66	54.64	12.37	12.54	46.10	11.57	12.12	33.09
SD	2.39	1.94	15.84	2.53	1.71	13.32	2.19	2.36	15.59	1.85	2.12	9.79

. Both 1D-CNN_f and 1D-CNN_GA exhibited good generalizability and robustness for C_ab (SD = 2.53–2.54%) and C_ar (SD = 1.71–1.94%) estimation. However, the 1D-CNN models for C_anth estimation showed relatively higher uncertainty (SD = 13.32–15.84%).

The prediction accuracy of leaf pigments using 2D-CNN models based on five strategies and different matrix dimensions input is shown in

Table 8. Accuracy of pigment estimation using 2D-CNN models with different numbers of convolutional layers.

Models	Strategies (Nconv)	Cab			Car			Canth
		R2	RMSE (µg/cm2)	nRMSE (%)	R2	RMSE (µg/cm2)	nRMSE (%)	R2	RMSE (µg/cm2)	nRMSE (%)
2D-CNNf (20 × 20)	A (1)	0.91	6.92	13.28	0.78	1.07	12.70	0.82	0.49	43.72
	B (2)	0.92	7.08	13.58	0.78	1.10	13.04	0.76	0.56	49.90
	C (3)	0.92	6.52	12.50	0.80	1.12	13.34	0.82	0.47	41.89
	D (4)	0.92	6.43	12.35	0.81	1.07	12.77	0.84	0.46	41.43
	E (5)	0.92	7.30	14.01	0.80	1.12	13.38	0.82	0.53	46.90
2D-CNNGA (20 × 20)	A (1)	0.91	6.98	13.40	0.83	1.05	12.53	0.88	0.37	32.99
	B (2)	0.92	6.38	12.24	0.82	1.09	12.98	0.87	0.40	35.29
	C (3)	0.92	6.23	11.95	0.83	1.16	13.80	0.89	0.37	33.14
	D (4)	0.92	6.10	11.71	0.84	1.03	12.29	0.89	0.35	31.58
	E (5)	0.92	6.81	13.07	0.82	1.11	13.29	0.89	0.36	31.84
2D-CNNf (40 × 10)	A (1)	0.92	6.75	12.96	0.82	1.12	13.31	0.80	0.50	44.54
	B (2)	0.92	6.67	12.79	0.83	1.07	12.78	0.81	0.48	42.46
	C (3)	0.92	6.78	13.01	0.82	1.10	13.13	0.78	0.61	54.02
	D (4)	0.91	7.16	13.75	0.81	1.11	13.23	0.78	0.54	47.79
	E (5)	0.92	6.88	13.20	0.84	1.08	12.85	0.79	0.52	46.36
2D-CNNGA (40 × 10)	A (1)	0.92	6.46	12.39	0.82	1.11	13.24	0.86	0.41	37.05
	B (2)	0.92	6.36	12.20	0.84	1.05	12.49	0.88	0.37	33.25
	C (3)	0.92	6.38	12.24	0.83	1.07	12.76	0.85	0.42	37.80
	D (4)	0.92	6.47	12.41	0.81	1.12	13.39	0.84	0.43	38.38
	E (5)	0.92	7.13	13.68	0.82	1.08	12.83	0.91	0.43	38.61
2D-CNNf (10 × 40)	A (1)	0.91	6.84	13.12	0.81	1.13	13.42	0.82	0.52	46.06
	B (2)	0.91	7.31	14.03	0.81	1.13	13.45	0.86	0.53	47.75
	C (3)	0.92	6.63	12.72	0.82	1.10	13.14	0.83	0.47	41.82
	D (4)	0.91	6.79	13.04	0.81	1.16	13.78	0.78	0.51	45.26
	E (5)	0.91	6.71	12.88	0.82	1.13	13.43	0.83	0.54	48.36
2D-CNNGA (10 × 40)	A (1)	0.91	6.65	12.76	0.80	1.18	14.02	0.89	0.42	37.65
	B (2)	0.92	6.75	12.96	0.81	1.15	13.68	0.88	0.40	35.48
	C (3)	0.92	6.49	12.45	0.82	1.13	13.39	0.88	0.38	33.98
	D (4)	0.92	6.68	12.81	0.79	1.20	14.23	0.84	0.43	38.30
	E (5)	0.92	6.66	12.77	0.80	1.14	13.54	0.86	0.40	35.90
2D-CNNf (50 × 8)	A (1)	0.91	6.81	13.06	0.83	1.16	13.85	0.79	0.52	46.16
	B (2)	0.92	6.72	12.89	0.84	1.08	12.83	0.77	0.56	49.96
	C (3)	0.92	6.64	12.74	0.84	1.06	12.60	0.80	0.48	42.74
	D (4)	0.90	7.31	14.02	0.83	1.08	12.90	0.77	0.57	50.61
	E (5)	0.92	6.65	12.76	0.83	1.07	12.72	0.80	0.49	43.62
2D-CNNGA (50 × 8)	A (1)	0.92	6.41	12.30	0.83	1.09	12.99	0.81	0.46	41.36
	B (2)	0.92	6.53	12.54	0.82	1.11	13.25	0.82	0.47	42.35
	C (3)	0.92	6.24	11.96	0.83	1.08	12.85	0.85	0.41	36.66
	D (4)	0.91	6.84	13.12	0.82	1.10	13.10	0.84	0.46	40.67
	E (5)	0.92	6.94	13.31	0.82	1.08	12.87	0.84	0.45	40.07
2D-CNNf (8 × 50)	A (1)	0.91	7.11	13.64	0.81	1.23	14.70	0.83	0.52	45.99
	B (2)	0.91	6.82	13.09	0.82	1.13	13.46	0.81	0.47	41.76
	C (3)	0.92	6.77	12.98	0.82	1.12	13.30	0.86	0.46	41.22
	D (4)	0.90	7.04	13.51	0.80	1.21	14.43	0.81	0.46	41.27
	E (5)	0.90	7.11	13.63	0.86	1.28	15.28	0.86	0.48	42.43
2D-CNNGA (8 × 50)	A (1)	0.91	6.92	13.28	0.82	1.12	13.29	0.86	0.48	42.67
	B (2)	0.92	6.49	12.45	0.82	1.09	12.97	0.85	0.43	38.62
	C (3)	0.92	6.41	12.29	0.84	1.08	12.91	0.87	0.42	38.22
	D (4)	0.91	6.95	13.33	0.81	1.11	13.24	0.83	0.46	40.97
	E (5)	0.92	7.30	14.00	0.80	1.20	14.26	0.83	0.44	39.57

Note: The bold represents the best strategy for each model.

. For both C_ab and C_ar, all 2D-CNN models with different strategies have good and stable prediction accuracy. In particular, when the number of convolutional layers was 3 or 4, C_ab estimation achieved R² > 0.92 and RMSE < 7 µg/cm² for both 2D-CNN_f and 2D-CNN_GA. Similarly, across different input matrix sizes, the 2D-CNN_GA model consistently yielded slightly higher estimation accuracy for all three pigments compared with the 2D-CNN_f model. Figure 8 shows the best-performing 2D-CNN models for different matrix dimensions, with the highest prediction accuracy obtained using a 20 × 20 matrix input under strategy D. Overall, the 2D-CNN-based approach, including 2D-CNN_f and 2D-CNN_GA, produced higher prediction accuracy than the 1D-CNN approach, particularly for C_anth (nRMSE value change from 42.74% to 31.58%). Furthermore, the optimal 2D-CNN models with four convolutional layers for estimating the three pigments exhibited better generalizability across the tree species compared with the 1D-CNN models (shown in Table A1).

Three benchmark estimation models (i.e., PLSR, RFR, and GPR) were trained and tested using the same training and validation dataset as the CNN models. The results of the three pigment estimations are shown in

Table 9. Accuracy comparison of pigment estimation between the optimal CNN models with four convolutional layers and three benchmark models (PLSR, RFR, and GPR).

Models	Cab			Car			Canth
	R2	RMSE (µg/cm2)	nRMSE (%)	R2	RMSE (µg/cm2)	nRMSE (%)	R2	RMSE (µg/cm2)	nRMSE (%)
PLSRf	0.75	11.07	21.24	0.53	1.93	22.96	0.39	0.92	82.13
PLSRGA	0.80	9.91	19.01	0.61	1.67	19.84	0.44	0.86	77.03
RFRf	0.84	9.05	17.36	0.70	1.53	18.17	0.51	0.81	72.15
RFRGA	0.85	8.81	16.90	0.70	1.51	18.00	0.52	0.79	70.50
GPRf	0.86	8.92	17.12	0.72	1.48	17.67	0.62	0.78	69.75
GPRGA	0.87	8.53	16.36	0.74	1.39	16.59	0.63	0.75	66.85

. The t-test results indicate a significant difference (p < 0.05) in prediction accuracy between three benchmark estimation models and CNN models. Compared with the best 1D- and 2D-CNN models, three benchmark estimation models have higher nRMSE, with a range of 16.36–21.24% for C_ab, 16.59–22.96% for C_ar, and 66.85–82.13% for C_anth. Similarly, three benchmark estimation models coupled with GA have higher accuracy than those using full-spectrum reflectance, particularly for the PLSR model. Among the three benchmark models, GPR provided the best estimates for C_ab and C_ar (nRMSE close to 16%), followed by RFR and PLSR. However, all models performed poorly in C_anth estimation (nRMSE > 65%).

The prediction accuracy of C_ab and C_ar using the optimal 1D- and 2D-CNN models with four convolutional layers is summarized in

Table 10. Estimation accuracy of C_ab and C_ar from the combined dataset of this study and LOPEX93 using optimal 1D- and 2D-CNN models with four convolutional layers.

Models	Cab			Car
	R2	RMSE (µg/cm2)	nRMSE (%)	R2	RMSE (µg/cm2)	nRMSE (%)
1D-CNNf	0.86	8.33	17.65	0.70	1.54	17.89
1D-CNNGA	0.87	7.94	16.83	0.74	1.47	17.15
2D-CNNf	0.87	7.92	16.77	0.75	1.46	16.96
2D-CNNGA	0.88	7.81	16.54	0.75	1.43	16.62

. As the number of tree species and sample sizes increased, both models maintained robust performance in estimating C_ab and C_ar (nRMSE < 18%), particularly when coupled with GA.

4. Discussion

4.1. The CNN Model Performance on Leaf Pigment Content

In this study, we developed a CNN-based approach using spectral reflectance to predict leaf pigment content across 28 tree species. The results obtained an acceptable accuracy of leaf pigment estimates, with nRMSE showing a range of 11.71–12.98% for C_ab, 12.29–14.88% for C_ar, and 31.58–56.13% for C_anth. These different accuracies caused by the variant of CNNs (1D-CNN_f or 2D-CNN_f trained with full spectrum (400–800 nm), and 1D-CNN_GA or 2D-CNN_GA trained with subset spectrum) were used. Different tree species differ in growth, phenological development, and plant composition, resulting in large differences in spectral characteristics. These differences reflect the complexity of spectral–pigment relationships, making it difficult to model with a small number of features. This study shows that the proposed CNN can be successfully applied to predict the pigmentation content of leaves of different tree species. Compared with PLSR, RFR, and GPR (nRMSE for C_ab and C_ar >16%), our proposed CNN models with large filters can extract more comprehensive spectral features and achieve more accurate pigment prediction, particularly for C_anth. Overall, these results confirmed that the CNN-based approaches can automatically extract the photosynthetic pigment information with satisfactory accuracy, especially for C_ab and C_ar.

The CNN architecture with four convolutional layers achieved the best performance in pigment prediction. Similarly, Prilianti et al. (2021) reported comparable accuracy using a network architecture with only three convolution layers [44]. It showed that, to a certain extent, adding more convolution layers resulted in larger errors in the training process. Therefore, it can be concluded that a relatively simple architecture is suitable for leaf pigment prediction. Another factor that cannot be ignored is the selected optimization method, which influences the training performance to a large extent [57]. SGDM can reduce overfitting in CNN model training, which has been systematically demonstrated by Prilianti et al. (2021) [44].

Four variants of CNNs for the prediction of C_ab, C_ar, and C_anth were compared in this study. The difference between these models was the input of the spectrum representation. The 1D-CNN_f and 1D-CNN_GA models used 1D spectral representations, while the 2D-CNN_f and 2D-CNN_GA models used 2D encoded matrices. The results showed that 2D-CNN-based approaches achieved higher prediction accuracy than 1D-CNN models, particularly when the importance bands selected by GA were used to transform 2D matrices. This improvement can be attributed to the reflectance spectrum being a stationary signal with similar peak–valley patterns. When transformed into a 2D matrix, these spectral features are emphasized, enabling convolutional filters to better capture the relationships between reflection peaks and valleys. Furthermore, the size of the 2D input image influenced performance, with the 20 × 20 format yielding the highest accuracy (Figure 8). This is likely because square matrices provide uniform coverage during convolution and preserve complete structural features. The GA-selected important bands (450 ± 20 nm, 510 ± 5 nm, 550 ± 20 nm, 700 ± 20 nm, and >750 nm) were consistent with previously reported sensitive ranges [7,21]. Among all models, the 2D-CNN_GA achieved the best performance, especially for C_anth. The ability of 2D-CNN to highlight reflectance characteristics within pigment absorption regions not only enhances interpretability but also improves model robustness and generalization.

4.2. Influence of Pigment Distribution

The distribution of the measured leaf pigment content is critical for developing a rubout deep learning-based regression model based on parameter estimates. The wide range of C_ab and C_ar obtained good accuracy (nRMSE around <15%), while a larger error (nRMSE > 30%) appeared on C_anth with a narrow range. Further, the results showed that the normal distribution of C_ab and C_ar performed better than a decreased distribution of C_anth. These results indicate that non-uniform sample distributions, particularly extreme cases, hinder the model’s ability to capture features across the full range, as neuronal weights become biased toward densely sampled regions. Furthermore, the distribution style influence on 1D-CNN models was greater than that of 2D-CNN models based on the prediction accuracy of pigment in this work. Because the features of 2D encoded matrices of reflectance spectra are easier to capture than 1D reflectance signals, this results in accelerated model convergence.

4.3. Limitations and Future Work

For C_ab and C_ar, both multi-output 1D-CNN and 2D-CNN models achieved good prediction accuracy. By contrast, C_anth showed poorer performance, with an nRMSE value around 55% for the 1D-CNN-based approach and around 32% for the 2D-CNN-based approach, mainly because its estimation can be affected by high levels of C_ab and C_ar [58]. Sensitive band selection was a key factor influencing the performance of CNN models. Féret et al. (2017) noted that chlorophyll exhibits two major absorption peaks in the blue (near 430 nm) and red (660–680 nm) regions, carotenoids strongly absorb in the blue region (400–500 nm), while anthocyanins show an absorption peak mainly in the blue–green region (500–550 nm) [13]. These absorption features partly overlap in the VIS region, especially in the blue–green region, which is critical for decoupling the effects of these pigments on leaf reflectance [7,59]. However, the GA-selected bands (Figure 6) in this study were located in low-absorption regions, introducing spectral noise. To mitigate these effects, spectral processing methods should be considered to improve pigment prediction. For example, the widely used derivative transformation could help determine the location of the spectral absorption features and reduce the finer-scale noise [60]. Another method worth considering in the future is wavelet analysis. Wavelet transformations are functions that decompose a complex signal into component sub-signals. Several studies have demonstrated that continuous wavelet decomposition can more effectively characterize spectral variability at appropriate frequencies, thereby improving pigment prediction accuracy [61,62].

The goal of model transferability is to develop a model that can be applied across different locations and times. Thus, the model training for pigment prediction requires sufficient training datasets and covers diverse geographical areas, tree species, and seasons. In this experiment, we obtained a robust CNN model with a dataset of 2094 across 28 broadleaf tree species. Fayad et al. (2021) similarly suggested that the optimal number of training samples was observed to be greater than 1500 for the retrieval of biochemical parameters [39]. However, since all samples were collected only during the growing season (April–September), the models cannot capture the spectral characteristics of autumn leaves with high anthocyanin content, limiting their accuracy in anthocyanin prediction and year-round monitoring. Incorporating samples from different seasons will therefore improve model transferability. In addition, the spectral interpretation is complex as it is influenced by diverse species and different geographic areas, which have different leaf optical properties [63,64]. A promising alternative is a hybrid method that exploits deep learning algorithms trained on RTM simulation, which can combine their advantages of the spectral response mechanism of pigment changes and the hyper-level learning ability. In addition, leaf samples were collected destructively to obtain pigment reference data in this study. While this approach ensures accurate ground truth, it constrains long-term, large-scale, and non-destructive monitoring. Advances in remote sensing offer a way forward, particularly through integrating data from multiple platforms and sensors (e.g., ground-based and UAV hyperspectral, RGB, and LiDAR) with artificial intelligence [40,41,42]. Such strategies are key to developing highly precise and transferable models for predicting plant physiological and biochemical traits across scales and environments.

5. Conclusions

Accurate estimates of leaf pigments are significant for providing valuable information about plant physiology. This study developed two types of convolutional neural networks (1D and 2D) coupled with GA to predict leaf pigment content across different tree species using the leaf reflectance spectrum. The results indicate that 2D-CNN-based approaches performed better than 1D-CNN-based approaches for predicting C_ab, C_ar, and C_anth, with the nRMSE value reduced by between 0.6% and 14.7%. The 2D-CNN approach produced more robust and lightweight results than the 1D-CNN model in this study. Furthermore, the selection of importance bands using the GA method improves the CNN’s performance on pigment prediction accuracy, especially for C_anth. estimation (improved by 9.85%) using the 2D-CNN_GA model. Additionally, the transferability and generalization of the model are important for applying it to predict a new dataset. In the future, larger, more diverse, and distributed databases across different tree species, seasons, and geographical areas will be considered to further enhance model performance.