Spectroscopic assessment of flavor-related chemical compounds in fresh tea shoots using deep learning

9 This study employs a deep-learning method, Y-Net, to estimate 10 tea flavor-related chemical 10 compounds (TFCC), including gallic acid, caffeine and eight catechin isomers, using fresh tea 11 shoot reflectance and transmittance. The unique aspect of Y-Net lies in its utilization of dual 12 inputs, reflectance and transmittance, which are seamlessly integrated within the Y-Net 13 architecture. This architecture harnesses the power of a convolutional neural network-based 14 residual network to fuse tea shoot spectra effectively. This strategic combination enhances the 15 capacity of the model to discern intricate patterns in the optical characteristics of fresh tea shoots, 16 providing a comprehensive framework for TFCC estimation. In this study, we destructively 17 sampled tea shoots from tea farms in Alishan (Ali-Mountain) in Central Taiwan within the 18 elevation range of 879–1552 m a.s.l. Tea shoot reflectance and transmittance data (n = 2032) 19 within the optical region (400–2500 nm) were measured using a portable spectroradiometer and 20 pre-processed using an algorithm; corresponding TFCC were qualified using the high-21 performance liquid chromatography analysis. To enhance the robustness and performance of Y-22 Net, we employed data augmentation techniques for model training. We compared the 23 performances of Y-Net and seven other commonly utilized statistical, machine-/deep-learning 24 models (partial least squared regression, Gaussian process, cubist, random forests and three 25 feedforward neural networks) using root-mean-square error (RMSE). Furthermore, we assessed 26 the prediction accuracies of Y-Net and Y-Net using spectra within the visible and near-infrared 27 (VNIR) regions (for higher energy throughput and low-cost instruments) and reflectance only 28 (for airborne and spaceborne remote sensing applications). The results showed that overall Y-Net 29 (mean RMSE ± standard deviation [SD] = 2.51 ± 2.20 mg g-1) outperformed the other statistical, 30 machine- and deep-learning models (≥ 2.59 ± 2.64 mg g-1), demonstrating its superiority in 31 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 3 predicting TFCC. In addition, this original Y-Net also yielded slightly lower mean RMSE (± SD) 32 compared with VNIR (2.76 ± 2.41 mg g-1) and reflectance-only (2.68 ± 2.74 mg g-1) Y-Nets 33 using validation data. This study highlights the feasibility of using spectroscopy and Y-Net to 34 assess minor biochemical components in fresh tea shoots and sheds light on the potential of the 35 proposed approach for effective regional monitoring of tea shoot quality. 36

caffeine, Camellia sinensis, catechin, deep learning, gallic acid, machine learning, 37 neural network, oolong tea, reflectance, transmittance, Y-Net 38 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 4 1. Introduction 39 Plant metabolism collectively produces many metabolites, crucial in resisting biotic stress and 40 adapting to abiotic pressure. These metabolites also serve as invaluable resources for human 41 health and survival [1-6]. Tea plants are predominantly cultivated in Asia, producing some of the 42 most popular non-alcoholic beverages in the world [7, 8]. The characteristics of tea are primarily 43 composed of its chemical components (tea flavor-related chemical compounds [TFCC] 44 hereafter), including gallic acid (GA), caffeine (CAF) and eight catechin isomers including 45 gallocatechin (GC), epicatechin gallate (EGC), catechin (C), epicatechin (EC), epigallocatechin 46 gallate (EGCG), gallocatechin gallate (GCG), epicatechin gallate (ECG) and catechin gallate 47 (CG) [8]. These chemicals contribute not only to the flavor of a cup of tea but also to active 48 responders to the environments in fresh tea leaves as tea plants grow [9]. The catechins, GA and 49 CAF mainly contribute to the astringency and bitterness of taste, which is the main body of the 50 tea soup, and the catechins oxidation is the essential chemical reaction that determines the tea 51 characteristics during the manufacturing process [10, 11]. The more we understand the TFCC 52 status in tea leaves, the more effectively we can decide on the management of the tea plants and 53 control the quality of tea production. 54

have been developed to quantify TFCC, such as gas chromatography or high-55 performance liquid chromatography (HPLC) [12-14]. However, they require pre-treatment, 56 leading to slow data retrieval, and are infeasible for the real-time TFCC assessment. One 57 practical alternative is non-destructive optical (approximately 400–2500 nm) spectroscopy [15, 58 16]. Previous studies employed optical spectroscopic estimation methods for some TFCC in 59 ground tea leaves [17, 18]. Huang et al. [19] and Wang et al. [20] used leaf reflectance to 60 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 5 estimate pigments and four main catechins and CAF in fresh green leaves using partial least 61 squares regression (PLS). However, these linear models may not be suitable for the prediction 62 due to the complex nature of TFCC. Yamashita et al. [6] uses machine learning, including 63 random forests (RF) and cubist, to analyze leaf reflectance and quantify TFCC in fresh tea 64 shoots. However, information may be lost solely relying on leaf reflectance during optimization. 65 Although reflectance provides valuable insights into interactions between reflected photons and 66 tea leaf surface, some TFCC may absorb light at specific wavelengths, and those features may 67 not be able to be delineated by reflectance. 68 Leaf transmittance, another measurable leaf optical attribute, indicates the proportion of the 69 incident energy passing through a substance without being reflected and absorbed; transmittance 70 variation may also be related to TFCC. By combining both leaf reflectance and transmittance, we 71 could optimize TFCC modeling in fresh tea shoots. Due to the potentially complex relationship 72 between green vegetation optics and TFCC (e.g., [6]), we employed deep learning to derive 73 TFCC from fresh tea shoots in this study. To our knowledge, this has yet to be carried out in 74 previous literature. Deep learning is more effective in handling large and high-dimensional 75 datasets than machine learning, which frequently requires manual feature engineering and 76 domain-specific knowledge to extract relevant features. The key strength of deep learning lies in 77 its ability to automatically learn and extract important features, enabling it to recognize complex 78 patterns and relationships in the data. Hence, the main objective of this study is to use an 79 advanced deep-learning algorithm (Y-Net, a convolutional neural network [CNN] based residual 80 network [ResNet] design approach) to unravel the complex relationship between tea shoot leaf 81 reflectance/transmittance and TFCC. 82 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 6 2. Materials and Methods 83 2.1. Tea shoot spectra collection and analysis 84 We conducted the field campaign in the Greater Alishan (or Ali-Mountain) tea plantation region 85 (23.47 N, 120.69 E) located in Central Taiwan within the elevation range of 800–1600 m a.s.l. 86 (Fig. 1). Fresh tea shoots (mainly the Chin-Hsin-Oolong cultivar [Camellia sinensis var. 87 sinensis]) were collected from 15 tea farms with the mean (± standard deviation [SD]) elevation 88 of 1225 ± 178 m a.s.l. ranging from 879 to 1552 m a.s.l (Table 1). The data collection period was 89 from April 12 to June 30, 2022, and May 11–24, 2023, encompassing the spring and summer tea 90 harvesting seasons. For each farm, we randomly collected two airtight polyethylene bags of tea 91 shoots, and each bag contained approximately 50–60 samples (an apical bud and three leaves). 92 Tea shoots were placed in a 0 ºC cooler right after being destructively sampled and were 93 transferred to the laboratory within the same day (mostly within six hours). 94 We randomly selected 30 samples from each bag to measure tea shoot spectra. Tea leaf 95 reflectance and transmittance were assessed using a portable spectroradiometer (FieldSpec 3, 96 Analytical Spectral Devices, Inc., Boulder, Colorado, USA) with a single integrating sphere 97 (ASD RTS-3ZC). The spectral resolutions were 3 nm (between 350 and 1000 nm) and 10 nm 98 (1000–2500 nm), with sampling intervals set at 1.4 nm and 2 nm, respectively. Due to the sizes 99 of the sample ports of the integrating sphere (diameters ≥ 13 mm), we only measured the spectra 100 of the third leaf (the largest one) for each tea shoot. To ensure the freshness of the tea leaf 101 samples for subsequent biochemical analysis, each spectral reading was an average of 10 102 measurements. To assess the reflectance (Rs) and transmittance (Ts) of tea leaves, a non-103 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 7 polarized measurement method was employed by referring to the manufacturer-recommended 104 procedure and [21]: 105 /g1844 /g3046/g3404 /g4666/g3010/g3294/g3267/g2879/g3010/g3279 /g3267/g4667/g3019/g3293 /g4666/g3010/g3278/g3267/g2879/g3010/g3279 /g3267/g4667, (1) 106 /g1846 /g3046/g3404 /g4666/g3010/g3294/g3269/g2879/g3010/g3279 /g3269/g4667/g3019/g3293 /g4666/g3010/g3296/g3269/g2879/g3010/g3279 /g3269/g4667, (2) 107 where Is, Ic, Iu and Id represent spectroradiometer readings taken from the tea leaf sample, 108 calibrated and un-calibrated white references, and stray light (measured using a light trap), 109 respectively. Superscripts R and T denote reflectance and transmittance, respectively, and Rr 110 indicates the reflectance factor of the calibrated white reference. The procedure effectively 111 measures the directional reflectance and transmittance factors [22]; the terms "leaf reflectance" 112 and "leaf transmittance" are used for simplicity. One major issue with spectral data collection in 113 a moist region like Taiwan is that high humility may introduce erroneous signals [23, 24]. Hence, 114 we employed a statistical/mathematical spectral reconstruction approach to retrieve noise-free 115 fresh tea leaf spectra involving spectral database matching, multivariable linear regression, linear 116 parameter multiplication and spectral reversion. The procedure is beyond the scope of this study, 117 and details of the spectral pre-processing and reconstruction were described in [25]. There were 118 1016 pairs (total n = 2032) of reflectance and transmittance data collected in this study. 119 2.2. TFCC analysis 120 Leaf samples (n = 30) were returned to the polyethylene bags (total 36 bags) after leaf spectral 121 measurement, and the samples were then immediately placed in a -20 ºC cold storage unit until 122 TFCC analysis. We note that instead of using the spectrally measured leaf [19], we processed an 123 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 8 entire tea shoot (an apical bud and three leaves) since tea makers use the former one to 124 manufacture dried tea products. These samples were freeze-dried, ground into a fine powder, and 125 sifted through a 4 mm sieve. Then, 0.5 grams of this homogeneous tea powder was immersed in 126 50 ml of boiling deionized water for 20 minutes in a water bath held at 90/i1 . The obtained tea 127 infusions were subsequently filtered via 0.45 μ m syringe filters in preparation for HPLC 128 analysis. The HPLC analysis was set up using a Shimadzu system (PU-2089, AS-2057, UV-2075 129 detector; Shimadzu, Kyoto, Japan) with an equipped C18 column (5 μ m × 4.6 mm × 250 mm; 130 Waters, Milford, MA, USA). Solvents A (0.1 % formic acid aqueous solution) and B (the 131 acetonitrile) were introduced following a specific gradient, which involved a linear increment of 132 solvent B from 1% to 10% over 15 minutes, subsequently escalating from 10% to 20% in the 133 next 14 minutes, then further increasing from 20% to 22% within 6 minutes, and lastly 134 maintaining the 22% concentration for an additional 5 minutes. A flow rate of 1.0 ml per minute 135 was maintained throughout the separation process. Absorbance was monitored at 280 nm for 136 real-time tracking of peak intensities. The TFCC components GA, CAF and eight catechin 137 isomers (GC, EGC, C, EC, EGCG, GCG, ECG and CG) were quantified utilizing a calibration 138 curve for each compound, respectively. The calibration curve was formulated using peak areas 139 from HPLC chromatograms at pre-determined concentrations with commercially available 140 chemical compounds (Sigma-Aldrich, St. Louis, MO, USA). All samples were analyzed in 141 triplicate to ensure consistency and accuracy (n = 108). 142 2.3. Deep learning (Y-Net) 143 2.3.1. Data augmentation 144 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 9 We divided 36 bags of samples into training (70% of the data for developing models) and testing 145 (30% for validation) groups. It is necessary to obtain sizable spectral and TFCC datasets to 146 develop a robust deep-learning model for learning and optimization. However, it is not feasible 147 to produce a large amount of samples of TFCC for deep learning using HPLC due to the nature 148 of the analytical procedure. To mitigate this technical challenge, a simplified approach was 149 adopted using an augmentation process. The rationale for employing the augmentation process 150 stems from the typically distributed correlation between spectral and biochemical characteristics, 151 similar to the principles applied in Gaussian processes [26]. Hence, the augmentation involves 152 using a normal distribution and requires the mean and SD from each bag of TFCC sample to 153 simulate 30 TFCC data to match the sample size of reflectance and transmittance data (illustrated 154 in Fig. 3a). This augmentation technique facilitates expanding the dataset and increases the 155 variation of the available data, thereby enhancing the training process and improving the ability 156 of the model to generalize new data (also see Chen et al. [27] and Wieland et al. [28]). 157 2.3.2. Y-Net structure 158 Y-Net is a specific architecture that utilizes feedforward neural networks (FNN) or CNN-based 159 ResNet approaches. The algorithm can take both reflectance and spectrally corresponding 160 transmittance as input (Fig. 3b). The Y-Net architecture consists of a one-dimensional (1D) 161 convolution block (Fig. 3c), 1D max-pooling, fusion and fully connected layers. The process can 162 be categorized into three steps, including feature extraction, fusion extraction and TFCC 163 prediction: 164 Feature extraction: This step comprises the encoder, which consists of six processes 165 responsible for extracting meaningful features from the input spectral data. The input dimension 166 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 10 was ~×2101×1, indicating one channel and a 1D spectral array with 2101 data points. The 167 symbol ~ represents the number of sample data corresponding to 36 datasets here. The processes 168 numbered 1, 3 and 5 involved convolutional blocks which are part of ResNet processing design, 169 and those with 2, 4 and 6 entailed max-pooling operations (Fig. 3c). The convolutional block 170 operation uses six layers including convolution, batch normalization, activation, and residual 171 layer added to the last layer (Fig. 3d). The kernel size was 1×5, indicating that the convolution 172 operation performed in the spectral windows with a stride of one and using padding to preserve 173 the data dimension. Each convolutional block comprised several kernel filters, and an increased 174 number of kernel filters in each block reduced the spectral dimension. This convolutional block 175 was responsible for feature extraction, mitigating the vanishing gradient problem and preserving 176 information across layers. The max-pooling process occurs between convolutional blocks and 177 was applied to a 1D array using a window size of two. It found the maximum value within each 178 window and applied a stride of two, meaning it moved by two positions at a time. A rectified 179 linear unit function was employed as the activation function in this architecture. It introduced 180 non-linearity by setting negative values to zero and leaving positive values unchanged. The 181 feature extraction step has six convolutional blocks (three reflectance and three transmittance 182 branches). Each corresponding convolutional block (reflectance and transmittance) has a various 183 number of parameters, namely [96, 64, 1296, 64, 96], [2592, 128, 5152, 128, 2592] and [10304, 184 256, 20544, 256, 10304], in which the total parameters are 10744, and these parameters 185 corresponded to the weights, normalization and biases in the filter masks; where each 186 convolutional block stands for [conv, norm, conv, norm, conv]. These parameters were 187 responsible for capturing and learning the patterns within the data. 188 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 11 Fusion extraction: The fusion extraction consisted of three processes numbered 7, 8 and 9 189 (Fig. 3c). The fusion process combined the outputs of the right (reflectance) and left 190 (transmittance) processes. It merged the information from these two sources using convolutional 191 block and max-pooling operation, resulting in a fused single processing operation. In the fusion 192 extraction step, one convolutional block fuses the reflectance and transmittance branches. It 193 consists of five parameters [82048, 512, 82048, 512, 82048], which are [conv, norm, conv, norm, 194 conv], respectively. The total number of unknown parameters in the fusion extraction step was 195 247168. These parameters include normalization, weights and biases associated with the 196 convolutional and max-pooling operations. 197 TFCC prediction: The process involved in transitioning from the convolutional to the 198 dense processes. This transition included flattening and dropping the data, which converted it 199 from a 2D representation (0-axis for spectral and 1-axis for kernel filters) to 1D dense vectors 200 and passed it through fully connected layers. The reshaped vector had a length of 33290 (Fig. 3c, 201 step 10), which was determined by the size of the output from the last max-pooling in the 202 previous step, which has the dimension ~×26×128. The vector was then connected to a hidden 203 layer with 10 neurons (Fig. 3c, step 11), representing the 10 TFCC components of the first 204 dataset. In the TFCC estimation step, there were 33290 unknown parameters with weights and 205 biases. 206 A total of 388202 parameters were utilized for optimizing Y-Net, encompassing both the 207 convolutional block and fully connected layers. Studies employed FNNs to determine the 208 optimal neural network structure and parameters, such as filter sizes, number of filters, layers, 209 neurons and activation functions [29, 30]. However, this process often involves trial and error 210 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 12 and can be time-consuming. Moreover, due to limited in-situ samples and the potential for 211 overfitting, finding the optimal neural structure with the best parameters may not always be 212 feasible. Therefore, this study adopts an alternative strategy by emphasizing the design of the 213 steps within the neural structures rather than exclusively pursuing the optimization of 214 parameters. 215 2.3.3. Y-Net Transfer Learning 216 The training of Y-Net was performed in two stages. In the first stage, we applied data 217 augmentation to 1080 samples from 36 batches. In the second stage, the data was transformed by 218 taking the average of three observed data points, resulting in 36 averaged data points 219 corresponding to the 36 batches. Subsequently, we split the data into training (70% of the data) 220 and testing (the rest of the data) sets. The number of the observed points was less than the 221 augmented points to train Y-Net. During Y-Net training, observed data were less than the 222 augmented ones. Hence, there were fewer actual data points available for training Y-Net. 223 Utilizing the augmented data points made it possible to train Y-Net effectively despite the 224 limited number of observed data points. In the initial stage of the two-stage training process (Fig. 225 3d), the focus was on reaching a local optimum by adjusting the weights, biases and dense 226 parameters from their initial values. The second stage was designed to achieve the global 227 maximum while also reducing the overall training time required to reach the optimal model. In 228 other words, the initial parameters used in the second stage were derived from the pre-trained 229 model obtained during the first stage of training. In the optimization process, we applied the 230 built-in function Adam optimizer. It incorporates adaptive learning rate and moment techniques 231 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 13 during training to reach the optimal model. We utilized the mean squared error (MSE) as the loss 232 function (eq. 3): 233 /g1838/g3404 /g2869 /g3041∑ /g4666 predicted /g3036/g3398 observed /g3036/g4667 /g2870,/g3036/g2880/g3041 /g3036/g2880/g2868 (3) 234 where the predicted /g3036/g3404 /g4668 /g1855/g1855/g3549 /g4669 /g3037/g2880/g2869 /g3011 and observed /g3036/g3404 /g4668 /g1855/g1855 /g4669 /g3037/g2880/g2869 /g3011; j is the index of TFCC J, n is the 235 number of data, and /g1855/g1855 and /g1855/g1855/g3549 are observed and predicted TFCC, respectively. During the 236 training process, a significant number of epochs were executed and the number of epochs was set 237 to 50. Throughout the training, the parameters of the epoch with the minimal loss value were 238 stored. These optimal parameter values and the network structure were implemented to estimate 239 TCFF. 240 2.4. Performance assessment 241 To assess the effectiveness and viability of Y-Net, we conducted a comparative analysis by 242 contrasting the performance of the two-stage training with that of the conventional one-stage 243 training. We employed the two-stage training approach to determine whether the additional 244 training stage would enhance model performance by referring to metrics losses. Furthermore, we 245 also selected an advanced multivariate statistical (PLS) approach [31] and other commonly used 246 machine-/deep-learning methods, including cubist [32] with de-trend pre-processing [6], 247 feedforward neural networks (FNNs) [33], Gaussian process [34] and RF [35]. Since these 248

have been commonly used in scientific literature, we only provide key references here 249 for expedience. To assess the performances of these methods and Y-Net, we use the root-mean-250 square error (RMSE) as the evaluation metric, as it measures the average deviation between the 251 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 14 predicted and actual observed values. To further evaluate the consistency and stability of the 252 models, we also took the mean and SD into account. 253 2.5. Y-Net spectral analysis 254 It is important to understand the contribution of different tea shoot spectral regions to estimate 255 TFCC in Y-Net. Therefore, we carried out a spectral feature importance analysis. The feature 256 important analysis took the features of each hidden layer in reflectance and transmittance 257 branches. From the first layer of each branch to the fusion layers, we resampled, averaged and 258 normalized those to the scale of 0–1 to assess their importance to TFCC prediction. The feature 259 importance analysis can be categorized into the local and global aspects to identify the pattern. 260 The global pattern constitutes a significant contribution that can be visually identified, whereas 261 the local pattern represents a minor contribution that requires a tool to unveil subtle information. 262 Therefore, the high-pass filter was also adopted using a fast Fourier transform to enhance the 263 minor contribution by capturing the high frequency in the feature importance [36]. In addition, 264 we carried out a spectral subset analysis. We investigated if visible and near-infrared spectral 265 regions (VNIR, 400–1100 nm) would be sufficient to estimate TFCC using Y-Net since they are 266 responsive to photosynthesis/non-photosynthesis pigments (mainly chlorophyll a and b, 267 carotenoids, xanthophylls and anthocyanins) [37] and cell structure (e.g., the mesophyll cells and 268 the fraction of air space) [38, 39] of fresh green leaves. Therefore, VNIR could be much more 269 sensitive than leaf water content dominant shortwave-infrared bands (SWIR, about 1300–2500 270 nm) [40]. Moreover, the spectroscopic measurement may be preferable within VNIR due to the 271 relatively higher energy throughput by referring to Planck's radiation law and more cost-effective 272 for the instruments (silicon/InGaAs for VNIR vs. InSb). Finally, we used tea shoot reflectance of 273 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 15 the entire optical region only (without transmittance) for Y-Net since the physical property may 274 be measured by airborne or spaceborne sensors permitting regional mapping of TFCC. 275 3. Results 276 3.1. Input data assessment 277 To evaluate the performance of Y-Net, it is important to compare the data distribution of the 278 observed data and augmented data (Fig. 4). We found that they have similar characteristics by 279 referring to statistics indicating the central tendency and spread of the data (Table S1). Therefore, 280 the observed and augmented data have similar distributions, which justifies using the latter for 281 further analysis. In addition, to investigate the data distribution of observed and augmented data, 282 a loss value (MSE) comparison for a hundred epochs was conducted in the optimization process 283 (Fig. 5). The optimization process provided two stages of training, with the first stage using 284 augmented data and the second stage using observed data. There was an apparent improvement 285 in the second stage, especially at the initial epoch; the difference in MSE became relatively 286 negligible after 25 epochs. The performance of the two-stage training through this optimization 287 was satisfactory by comparing the observed and predicted training and validation data (Fig. 6 288 and Table S2). Overall, the data distributions of those groups were comparable except for some 289 outliers of C, ECG and CAF. 290 3.2. Model performance comparison 291 Y-Net underwent pre-training and acquired knowledge through a novel two-stage training 292 approach (Fig. 5), utilizing TFCC from observed and augmented data (Fig. 4 and Table S1). 293 Overall, the performance of Y-Net (mean ± SD of RMSE = 2.51 ± 2.20) to model the variation 294 of TFCC was superior to (yielding the lowest mean and SD) selected statistical (PLS) (7.86 ± 295 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 16 7.58) and machine learning (cubist, FNNs, Gaussian, RF) methods (2.59 ± 2.64). The TFCC 296 assessment of Y-Net was relatively consistent by referring to SD demonstrating its effectiveness 297 across various compounds. However, the performance for modeling each TFCC compound 298 differed (Table 2). 299 3.3. Y-Net spectral analysis 300 According to our spectral feature importance analysis, several spectral regions of reflectance and 301 transmittance were pivotal for modeling TFCC; reflectance (55% of the optical spectral bands) 302 played a more critical role than transmittance (45%) by referring to feature important values 303 (Fig. 7). Specifically, after applied the high-pass filter (adopted with fast Fourier transform), 304 spectral bands of reflectance and transmittance located around 400 nm, 700 nm, 1400 nm, 1900 305 nm and 2500 nm were identified to be crucial for indirectly estimating TFCC. The spectral 306 subset analysis depicts that, overall, the Y-Net using both reflectance and transmittance of the 307 entire optical region (400–2500 nm) (mean ± SD of RMSE = 1.29 ± 1.08 for training and 2.51 ± 308 2.20 for validation datasets) (Table 3) performed slightly better and relatively more consistent 309 (with the lowest mean and SD) than the VNIR (2.72 ± 2.49; 2.76 ± 2.41) and the reflectance-310 only (2.07 ± 2.11; 2.68 ± 2.74) Y-Nets. However, the performance of estimating each TFCC 311 compound was different. 312 4. Discussion 313 4.1. Y-Net performance 314 The Y-Net architecture features a dual-branch input, integrating a ResNet-based CNN [41, 42]. 315 The dual input of Y-Net involves data augmentation or enrichment and pre-training through a 316 novel two-stage training approach (Fig. 3a and 3d), including utilizing TFCC obtained from both 317 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 17 augmented and observed data measurements. This unique process markedly improves the 318 performance in predicting TFCC, as evidenced by the initial epoch of the second-stage training 319 involving pre-training using augmented data. Compared to conventional machine-learning 320 methods, the enhancement in performance is demonstrated (see Table 2). The reduction in loss 321 during this phase can be attributed to the utilization of pre-trained parameters from the first-stage 322 training in the second stage. This process serves a dual purpose: Preventing overfitting arising 323 from limited observational data and augmenting the capacity for model generalization. 324 Furthermore, this approach reinforces the adaptability and robustness of the model. It facilitates 325 effective generalization of model performance across diverse datasets, utilizing data 326 augmentation and chemical compounds (TFCC) through feature extraction, fusion and 327 estimation (Fig. 3b). Therefore, Y-Net becomes adept at capturing intricate data patterns with a 328 specific focus on optimizing both local and global aspects during training. The local aspect of Y-329 Net involves capturing details or specific patterns of the input data. In contrast, the global aspect 330 recognizes the overarching trend or general pattern spanning the entire dataset. Additionally, 331 integrating ResNet into the architecture addresses challenges such as vanishing gradients and 332 ensuring information preservation across layers (Fig. 3c and 3d). 333 4.2. Spectral importance to TFCC retrieval using Y-Net 334 To assess the performance of Y-Net, we aim to compare the spectral effectiveness of the optical 335 region (400–2500 nm) using feature important values by calculating the average feature 336 extraction values in each branch, namely reflectance and transmittance. The average importance 337 was derived from each convolutional block in the deep CNN (total of 37 layers). Each layer can 338 yield results for feature importance, referred to as spectral contribution to predict TFCC. In the 339 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 18 hidden layers, higher spectral feature values contribute significantly to the prediction of TFCC, 340 whereas lower values indicate a comparatively less influential role in the predictive process. The 341 analysis of spectral contributions across layers aids in understanding the significance of various 342 features and their impact on accurately estimating TFCC concentrations by Y-Net throughout the 343 optical range, particularly at 400 nm, 700 nm, 1400 nm, 1900 nm and 2500 nm (Fig. 7). To 344 qualitatively identify patterns, encompassing both local and global aspects, and to isolate local 345 patterns within the branch layers of reflectance and transmittance, a high-pass filter was 346 employed in this study [43, 44]. The high-pass filter adopts a fast Fourier transform that can 347 enhance the minor contribution or local aspect by capturing the high frequency in the feature 348 importance [35]. It is acknowledged that global aspects include visually apparent patterns such as 349 trends and curves of the feature importance. Those global patterns are a significant contribution 350 to TFCC prediction. On the other hand, interpreting local aspects may be challenging due to their 351 minor contribution despite utilizing a substantial influence on TFCC prediction within specific 352 wavelength regions. Using the high-pass filter, we can extract local patterns instead of 353 considering them as interference, especially the feature importance in the deep layer networks. 354 The local pattern was exhibited and appeared in the wavelength at around 400, 500, 700, 1150, 355 1400, 1900 and 2500 mm (Fig. 7). These significant patterns are evident at the two ends of the 356 optical region (400 nm, 2500 nm) and at 1900 nm, while lower significances were observed 357 around 1150 nm and 1400 nm. 358 We also compared the spectral effectiveness of the optical region (reflectance and 359 transmittance) with the models ingesting the VNIR range of 400–1100 nm of reflectance and 360 transmittance (due to high signal-to-noise ratio and cost-effective instrument), and reflectance 361 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 19 only (for airborne and spaceborne remote sensing applications) of the entire optical region (Table 362 3). Since the spectral patterns of green vegetation reflectance and transmittance are similar [38] 363 (Fig. 2), they yielded almost identical feature important values (Fig. 7). Thus, the reflectance 364

was adopted as the comparison in this study. The comparison was visually addressed 365 using the feature importance and their high-pass filter, specifically for reflectance and 366 transmittance, in each branch, while the transmittance was not used in the reflectance method. 367 This implies that by employing VNIR in the experiment, we neglect the significant important 368 features and the less significant features in the spectral range of 1101 nm to 2500 nm (Table 3, 369 VNIR), namely that evident in the spectral edges (400 nm, 2500 nm) and at 1900 nm. Therefore, 370 the proposed Y-Net using entire optical region reflectance/transmittance outperformed the other 371 two Y-Nets models. This is because, although they share almost identical important features and 372 local patterns, they complement each other, with specific patterns present in spectral reflectance 373 that may not be found in spectral transmittance. We also note that the performance of the Y-Net 374 was merely slightly better than the VNIR and reflectance-only Y-Nets. The spectral analysis 375 demonstrates that we could re-produce similar outcomes with Y-Net using low-cost instruments. 376 In addition, this study also underscores the potential of using airborne or spaceborne 377 hyperspectral remote sensing for regional monitoring of TFCC. 378 4.3. Uncertainties and limitations 379 While Y-Net demonstrates satisfactory performance (Table 2), uncertainties exist, primarily 380 attributed to the complexity of TFCC. The intricate interactions and dependencies among various 381 TFCC pose challenges in achieving absolute precision. The performance of Y-Net is contingent 382 on the quality and diversity of the training data, such as using additional augmented data (Fig. 383 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 20 3a). Limited representation of specific chemical profiles in the training set may lead to 384 challenges in accurately predicting less-represented compounds. For instance, the range of 385 chemical compound values in augmented data may not be sufficient for effective Y-Net training. 386 Expanding the range of compound values in Y-Net involves optimizing sensitivity and precision. 387 Additionally, the generalization of Y-Net to diverse tea cultivars and physical environments 388 could pose a limitation. The inherent variability in TFCC across different varieties and growth 389 environments could ramify the predictive accuracy. Continuous refinement and expansion of the 390 training dataset with actual data (without data augmentation) are crucial measures that could 391 address some of these limitations over time. This ongoing effort aims to enhance the adaptability 392 and robustness of Y-Net to a broader spectrum of TFCC profiles, ultimately improving its 393 applicability across diverse environmental conditions and TFCC. 394 5. Conclusions 395 In this study, we proposed Y-Net, a deep-learning approach, to estimate TFCC from fresh tea 396 shoots indirectly using their optical reflectance and transmittance. We found that the 397 performance of Y-Net was superior to some commonly employed advanced statistical (PLS) and 398 machine-/deep-learning (cubist, FNNs, Gaussian and RF) methods by referring to RMSE. The 399 dual-input approach of Y-Net, considering both reflectance and transmittance simultaneously, 400 outperformed those selected single-input models, providing a more comprehensive 401 representation of TFCC and improving predictive accuracy. Our spectral analyses showed that 402 some specific bands located at 400 nm, 500 nm, 700 nm, 1150 nm, 1400 nm, 1900 nm and 2500 403 nm were important for retrieving TFCC. In addition, the reduction of the spectral range (VNIR 404 Y-Net) or exclusion of transmittance (reflectance only Y-Net) would also retard the overall 405 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 21 performance. These results emphasize the importance of advanced optimization techniques and 406 leveraging multiple sources of spectral data for precise chemical compound prediction in fresh 407 tea shoots. 408 Acknowledgments 409 We appreciate the field and lab assistance provided by Chi-Ching Huang, Yi-Ching Hung and 410 Zih-Yu Shen. This work was supported by the National Science and Technology Council (NSTC 411 112-2321-B-002-016), National Taiwan University (NTU-AS-112L104303), and the Research 412 Center for Future Earth, the Featured Areas Research Center Program, the Higher Education 413 Sprout Project, Ministry of Education (Taiwan). 414 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 22

415 [1] R.A. Dixon, D. Strack, Phytochemistry meets genome analysis, and beyond ......... 416 Phytochemistry, 62 (2003) 815–816. 417 [2] C. Fang, A.R. Fernie, J. Luo, Exploring the diversity of plant metabolism, Trends in Plant 418 Science, 24 (2019) 83–98. 419 [3] A. Rai, K. Saito, M. Yamazaki, Integrated omics analysis of specialized metabolism in 420 medicinal plants, The Plant Journal, 90 (2017) 764–787. 421 [4] J.K. Weng, The evolutionary paths towards complexity: A metabolic perspective, New 422 Phytologist, 201 (2014) 1141–1149. 423 [5] Q. Yang, Y. He, M. Kabahuma, T. Chaya, A. Kelly, E. Borrego, Y. Bian, F. El Kasmi, L. 424 Yang, P. Teixeira, J. Kolkman, R. Nelson, M. Kolomiets, J.L. Dangl, R. Wisser, J. Caplan, X. Li, 425 N. Lauter, P. Balint-Kurti, A gene encoding maize caffeoyl-CoA O-methyltransferase confers 426 quantitative resistance to multiple pathogens, Nature Genetics, 49 (2017) 1364–1372. 427 [6] H. Yamashita, R. Sonobe, Y. Hirono, A. Morita, T. Ikka, Potential of spectroscopic analyses 428 for non-destructive estimation of tea quality-related metabolites in fresh new leaves, Scientific 429 Reports, 2021, pp. 4169. 430 [7] Y. Hilal, U. Engelhardt, Characterisation of white tea - Comparison to green and black tea, 431 Journal of Consumer Protection and Food Safety, 2 (2007) 414–421. 432 [8] S.Y. Pan, Q. Nie, H.C. Tai, X.L. Song, Y.F. Tong, L.J.F. Zhang, X.W. Wu, Z.H. Lin, Y.Y. 433 Zhang, D.Y. Ye, Y. Zhang, X.Y. Wang, P.L. Zhu, Z.S. Chu, Z.L. Yu, C. Liang, Tea and tea 434 drinking: China’s outstanding contributions to the mankind, Chinese Medicine, 17 (2022) 27. 435 [9] S. Ahmed, T.S. Griffin, D. Kraner, M.K. Schaffner, D. Sharma, M. Hazel, A.R. Leitch, C.M. 436 Orians, W. Han, J.R. Stepp, A. Robbat, C. Matyas, C. Long, D. Xue, R.F. Houser, SB Cash, 437 Environmental factors variably impact tea secondary metabolites in the context of climate 438 change, Frontiers in Plant Science, 10 (2019) 939. 439 [10] L. Zhang, Q.Q. Cao, D. Granato, Y.Q. Xu, C.T. Ho, Association between chemistry and 440 taste of tea: A review, Trends in Food Science & Technology, 101 (2020) 139–149. 441 [11] D. Arumai Selvan, D. Mahendiran, R. Senthil Kumar, A. Kalilur Rahiman, Garlic, green tea 442 and turmeric extracts-mediated green synthesis of silver nanoparticles: Phytochemical, 443 antioxidant and in vitro cytotoxicity studies, Journal of Photochemistry and Photobiology B: 444 Biology, 180 (2018) 243–252. 445 [12] D.A. Dias, O.A.H. Jones, D.J. Beale, B.A. Boughton, D. Benheim, K.A. Kouremenos, J.L. 446 Wolfender, D.S. Wishart, Current and future perspectives on the structural identification of small 447 molecules in biological systems, Metabolites, 6 (2016) 46. 448 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 23 [13] AR Fernie, R.N. Trethewey, A.J. Krotzky, L. Willmitzer I N N Ovat I O N, Metabolite 449 profiling: From diagnostics to systems biology, Nature Reviews Molecular Cell Biology, (2004) 450 763–769. 451 [14] J.L. Wolfender, J.M. Nuzillard, J.J.J. Van Der Hooft, J.H. Renault, S. Bertrand, 452 Accelerating metabolite identification in natural product research: Toward an ideal combination 453 of liquid chromatography-high-resolution tandem mass spectrometry and NMR profiling, in 454 silico databases, and chemometrics, American Chemical Society Publications, 91 (2019) 704–455 742. 456 [15] L. Li, J. Huang, Y. Wang, S. Jin, M. Li, Y. Sun, J. Ning, Q. Chen, Z. Zhang, Intelligent 457 evaluation of storage period of green tea based on VNIR hyperspectral imaging combined with 458 chemometric analysis, Infrared Physics & Technology, 110 (2020) 103450. 459 [16] M.N. Merzlyak, T.B. Melo, K. Razi Naqvi, Estimation of leaf transmittance in the near 460 infrared region through reflectance measurements, Journal of Photochemistry and Photobiology 461 B: Biology, 74 (2004) 145–150. 462 [17] M.S. Lee, Y.S. Hwang, J. Lee, M.G. Choung, The characterization of caffeine and nine 463 individual catechins in the leaves of green tea (Camellia sinensis L.) by near-infrared reflectance 464 spectroscopy, Food Chemistry, 158 (2014) 351–357. 465 [18] H. Schulz, U.H. Engelhardt, A. Wegent, H.H. Drews, S. Lapczynski, Application of near-466 infrared reflectance spectroscopy to the simultaneous prediction of alkaloids and phenolic 467 substances in green tea leaves, Journal of Agricultural and Food Chemistry, 47 (1999) 5064–468 5067. 469 [19] Y. Huang, W. Dong, A. Sanaeifar, X. Wang, W. Luo, B. Zhan, X. Liu, R. Li, H. Zhang, X. 470 Li, Development of simple identification models for four main catechins and caffeine in fresh 471 green tea leaf based on visible and near-infrared spectroscopy, Computers and Electronics in 472 Agriculture, 173 (2020) 105388. 473 [20] Y.J. Wang, S.S. Jin, M.H. Li, Y. Liu, L.Q. Li, J.M. Ning, Z.Z. Zhang, Onsite nutritional 474 diagnosis of tea plants using micro near-infrared spectrometer coupled with chemometrics, 475 Computers and Electronics in Agriculture, 175 (2020) 105538. 476 [21] A. Hovi, P. Forsström, M. Mõttus, M. Rautiainen, Evaluation of accuracy and practical 477 applicability of methods for measuring leaf reflectance and transmittance spectra, Remote 478 Sensing, 10 (2018) 25. 479 [22] M.A. Greiner, BD. Duncan, M.P. Dierking, Bidirectional scattering distribution functions of 480 maple and cottonwood leaves, Applied Optics, 46 (2007) 6485–6494. 481 [23] A. Simic, J.M. Chen, S.G. Leblanc, A. Dyk, H. Croft, T. Han, Testing the top-down model 482 inversion method of estimating leaf reflectance used to retrieve vegetation biochemical content 483 within empirical approaches, IEEE Journal of Selected Topics in Applied Earth Observations 484 and Remote Sensing, 7 (2014) 92–104. 485 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 24 [24] H.M. Noda, T. Motohka, K. Murakami, H. Muraoka, K.N. Nasahara, Reflectance and 486 transmittance spectra of leaves and shoots of 22 vascular plant species and reflectance spectra of 487 trunks and branches of 12 tree species in Japan, Ecological Research, 29 (2014) 111. 488 [25] L. Denaro, H. Li, J. Chong, and C. Huang, Reconstructing fresh green leaf spectra in the 489 SWIR-2 region (2001-2500 nm) collected in a humid environment by referring to publicly 490 available green leaf spectral databases, bioRxiv, doi: https://doi.org/10.1101/2024.01.04.574267. 491 [26] S. Van Wittenberghe, J. Verrelst, J.P. Rivera, L. Alonso, J. Moreno, R. Samson, Gaussian 492 processes retrieval of leaf parameters from a multi-species reflectance, absorbance and 493 fluorescence dataset, Journal of Photochemistry and Photobiology B: Biology, 134 (2014) 37–494 48. 495 [27] L. Chen, Y. Wei, Z. Yao, E. Chen, X. Zhang, Data augmentation in prototypical networks 496 for forest tree species classification using airborne hyperspectral images, IEEE Transactions on 497 Geoscience and Remote Sensing, 60 (2022) 4410116. 498 [28] M. Wieland, Y. Li, S. Martinis, Multi-sensor cloud and cloud shadow segmentation with a 499 convolutional neural network, Remote Sensing of Environment, 230 (2019) 111203. 500 [29] V.K. Ojha, A. Abraham, V. Snášel, Metaheuristic design of feedforward neural networks: A 501 review of two decades of research, Engineering Applications of Artificial Intelligence, 60 (2017) 502 97–116. 503 [30] M. Paliwal, U.A. Kumar, Assessing the contribution of variables in feed forward neural 504 network, Applied Soft Computing, 11 (2011) 3690–3696. 505 [31] GLAA dos Santos, A.S. Reis, M.R. Besen, R.H. Furlanetto, M. Rodrigues, L.G.T. Crusiol, 506 K.M. de Oliveira, R. Falcioni, RB de Oliveira, M.A. Batista, M.R. Nanni, Spectral method for 507 macro and micronutrient prediction in soybean leaves using interval partial least squares 508 regression, European Journal of Agronomy, 143 (2023) 126717. 509 [32] P.C. Agyeman, J. Kingsley, N.M. Kebonye, V. Khosravi, L. Borů vka, R. Vašát, Prediction 510 of the concentration of antimony in agricultural soil using data fusion, terrain attributes 511 combined with regression kriging, Environmental Pollution, 316 (2023) 120697. 512 [33] P. Li, Classification of black tea leaf water content based on hyperspectral imaging, in: 513 2023 IEEE International conference on control, electronics and computer technology (ICCECT), 514 Institute of Electrical and Electronics Engineers Inc., 2023, pp. 621–624. 515 [34] W.S. Holmes, M.P.-L. Ooi, S. Abeysekera, Y.C. Kuang, R. Simpkin, M. Caddie, J. Nowak, 516 S. Demidenko, On machine learning methods to estimate cannabidiolic acid content of Cannabis 517 sativa L. from near-infrared hyperspectral imaging, in: 2023 IEEE International instrumentation 518 and measurement technology conference (I2MTC), 2023, pp. 01–06. 519 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 25 [35] S. Ye, H. Weng, L. Xiang, L. Jia, J. Xu, Synchronously predicting tea polyphenol and 520 epigallocatechin gallate in tea leaves using Fourier transform–near-infrared spectroscopy and 521 machine learning, Molecules, 28 (2023) 5379. 522 [36] X. Qiao, W. Huang, Spectral-spatial-frequency transformer network for hyperspectral image 523 classification, 2023 IEEE Sensors Applications Symposium (SAS), Institute of Electrical and 524 Electronics Engineers Inc., 2023. 525 [37] S.L. Ustin, J.A. Gamon, Remote sensing of plant functional types, New Phytologist, 186 526 (2010) 795–816. 527 [38] S. Jacquemoud, S. Ustin, Leaf optical properties in different wavelength domains, 528 Cambridge University Press, 2019. 529 [39] K. Klancnik, K. Vogel-Mikus, M. Kelemen, P. Vavpetic, P. Pelicon, P. Kump, D. Jezersek, 530 A. Gianoncelli, A. Gaberscik, Leaf optical properties are affected by the location and type of 531 deposited biominerals, Journal of Photochemistry and Photobiology B: Biology, 140 (2014) 532 276–285. 533 [40] C.y. Huang, W.R.L. Anderegg, G.P. Asner, Remote sensing of forest die-off in the 534 Anthropocene: From plant ecophysiology to canopy structure, Remote Sensing of Environment, 535 231 (2019) 111233. 536 [41] O. Andrew, A. Apan, D.R. Paudyal, K. Perera, Convolutional neural network-based deep 537 learning approach for automatic flood mapping using NovaSAR-1 and Sentinel-1 data, ISPRS 538 International Journal of Geo-information, 12 (2023) 194. 539 [42] C.P. Lee, K.M. Lim, Y.X. Song, A. Alqahtani, Plant-CNN-ViT: Plant classification with 540 ensemble of convolutional neural networks and vision transformer, Plants, 12 (2023) 2642. 541 [43] S.H. Yoo, G. Huang, K.S. Hong, Physiological noise filtering in functional near-infrared 542 spectroscopy signals using wavelet transform and long-short term memory networks, 543 Bioengineering, 10 (2023) 685. 544 [44] G. Blaney, C. Fernandez, A. Sassaroli, S. Fantini, Dual-slope imaging of cerebral 545 hemodynamics with frequency-domain near-infrared spectroscopy, Neurophotonics, 10 (2023) 546 013508. 547 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 26 Table 1. The sample sites and the corresponding elevations (from low to high) and sampling 548 dates. Long dashes (—) indicate no data (dates) available. 549 Site name Elevation (m) 2022 2023 Date 1 Date 2 Date 1 ZMG 879 04/19 — 05/24 LM 985 05/24 — — LT 1012 05/03 — — BHS 1080 04/22 — — FS 1116 05/24 — 05/16 SL 1212 06/17 — 05/11 CST 1221 04/22 — 05/24 YCL 1226 04/19 — — JP 1271 04/22 — — YL 1307 04/19 06/30 — HTD 1344 04/19 — — MF 1346 04/12 — 05/11 LY 1362 04/12 — — SY 1463 05/03 06/30 05/16 SYU 1552 05/03 — — 550 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 27 Table 2. Performance comparison of selected statistics (partial linear least squared [PLS]) and 551 machine-/deep-learning methods (cubist, three feedforward neural networks [FNN-1, FNN-2 and 552 FNN-3], Gaussian process, random forests [RF] and Y-Net based convolutional neural network) 553 to estimate TFCC using root-mean-square error (RMSE). The underscored numbers indicate the 554 lowest values for each TFCC and the mean and standard deviation (SD). 555 TFCC PLS Cubist FNN-1 FNN-2 FNN-3 Gaussian RF Y-Net GC 15.03 8.33 3.38 3.35 3.31 1.24 3.74 3.23 ECG 9.02 4.33 7.93 7.92 8.57 21.27 5.71 8.03 C 4.34 0.63 0.74 0.73 0.83 1.55 0.57 0.86 EC 2.68 1.29 1.53 1.52 1.59 2.63 1.14 1.49 EGCG 25.28 6.28 8.13 8.05 6.88 16.12 9.41 4.09 GCG 1.63 0.51 0.28 0.28 0.28 3.31 0.48 0.40 ECG 4.52 1.13 1.52 1.5 1.24 2.86 1.68 2.24 CG 0.12 0.3 0.06 0.05 0.05 4.37 0.05 0.36 GA 2.10 0.8 0.95 0.94 0.98 3.60 0.94 1.24 CAF 13.92 2.37 2.79 2.83 3.31 19.31 2.77 3.19 Mean 7.86 2.59 2.73 2.72 2.70 7.63 2.65 2.51 SD 7.58 2.64 2.82 2.81 2.75 7.52 2.80 2.20 556 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 28 Table 3. Y-Net performance comparison (RMSE) using spectra (reflectance and transmittance) 557 from the whole optical (400–2500 nm) and visible-near infrared (400–1100 nm) regions and 558 reflectance of the optical region only to estimate TFCC. The underscored numbers indicate the 559 lowest values for each TFCC, mean and SD, separated by training and validation. 560 TFCC Optical region VNIR Reflectance Training Validation Training Validation Training Validation GC 1.57 3.23 2.88 2.00 2.75 2.98 ECG 2.69 8.03 4.64 5.63 4.16 7.94 C 0.47 0.86 0.96 1.19 0.64 0.70 EC 0.66 1.49 1.43 2.02 0.92 1.70 EGCG 3.57 4.09 6.64 7.90 7.12 7.70 GCG 0.30 0.40 0.39 0.40 0.37 0.34 ECG 0.63 2.24 1.61 2.42 1.12 1.72 CG 0.38 0.36 0.25 0.29 0.38 0.36 GA 0.54 1.24 0.95 0.88 0.74 1.21 CAF 2.03 3.19 7.43 4.87 2.90 2.79 Mean 1.29 2.51 2.72 2.76 2.07 2.68 SD 1.08 2.20 2.49 2.41 2.11 2.74 561 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint 29 Figure legends 562 Figure 1. (a) The study area, the Greater Alishan tea plantations (green pixels) within the 563 elevation range of 800–1600 m a.s.l. (b) located in Central Taiwan (the yellow polygon). We 564 collected tea shoot spectral and tea flavor-related chemical compounds data in tea farms (Table 565 1) illustrated as yellow dots. 566 Figure 2. Pre-processing of tea leaf reflectance ((a) before and (c) after the spectral 567 reconstruction) and transmittance ((b) before and (d) after the spectral reconstruction) data. 568 Figure 3. The workflow of using Y-Net to estimate tea flavor-related chemical compounds 569 (TFCC) with tea shoot reflectance and transmittance. (a) Illustration of augmented from 570 observed data. (b) Y-Net-based ResNet (residual network) structure for TFCC prediction. (c) 571 Convolutional block of ResNet, Y-Net. (d) Two-stage training: Top (the first stage) and bottom 572 (the second stage). 573 Figure 4. The (a) observed (n = 114) and (b) augmented (n = 1135) TFCC including gallic acids 574 (GA), caffeine (CAF) and eight catechin isomers including gallocatechin (GC), epicatechin 575 gallate (EGC), catechin (C), epicatechin (EC), epigallocatechin gallate (EGCG), gallocatechin 576 gallate (GCG), epicatechin gallate (ECG) and catechin gallate (CG). 577 Fig. 5. Epoch comparison for 1st and 2nd stage learning; MSE represents the mean squared error. 578 Fig. 6. Estimation of TFCC using Y-Net with tea shoot spectra: Observed (a) training and (b) 579 validation; predicted (a) training and (b) validation datasets. 580 Figure 7. Feature important values of tea shoot optical spectra in Y-Net. 581 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 8, 2024. ; https://doi.org/10.1101/2024.03.05.583504doi: bioRxiv preprint