Plant character nondestructive monitoring method and device based on deep learning

Abstract

The invention provides a plant character nondestructive monitoring method and device based on deep learning. The problem that an existing plant character monitoring method is low in generalization and does not have universality is effectively solved. The method comprises the following steps: carrying out spectral measurement on a plurality of plant data sets with plant species in a target spectral range, and obtaining various plant character parameters in a measurement process through a spectral radiometer to generate a plurality of groups of spectral data; extracting a plurality of groups of label pairs from reflection simulation spectrums and transmission simulation spectrums generated based on the plurality of groups of spectral data and the plurality of plant character parameters, and encoding the plurality of groups of label pairs to obtain a plurality of spectral features; and controlling the character prediction model to process the spectral feature map to obtain plant character data so as to complete nondestructive monitoring of plant characters in different ecological systems based on the plant character data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of non-destructive monitoring of plant traits, in particular, to a non-destructive monitoring method and device for plant traits based on deep learning. BACKGROUND

[0002] Plant traits such as chlorophyll content (CHL), equivalent water thickness (EWT), and leaf mass per area (LMA) are important indicators for evaluating ecosystem function and vegetation health. However, traditional measurement methods such as destructive sampling and laboratory analysis are labor-intensive, time-consuming, and low-throughput, limiting their applicability in large spatial and temporal scales. Therefore, there is an urgent need for non-destructive and efficient methods to quickly and accurately estimate plant traits in order to enable real-time decision-making in precision agriculture, ecological conservation, and climate restoration programs.

[0003] High spectral remote sensing provides a non-destructive method for trait estimation. Existing models used in high spectral remote sensing mainly focus on single ecosystem applications (forests or farmland). In these applications, the models are trained and validated in homogeneous environments, making it difficult to generalize between heterogeneous ecosystems and sensor types. This limits their universality in heterogeneous ecosystems, such as transitioning from controlled farmland to natural grasslands or mixed forests. Due to differences in species composition, canopy structure, and environmental stressors, spectral characteristics relationships can change significantly. Few studies have systematically evaluated the ability of transfer learning to enhance cross-ecosystem feature prediction, and there is a significant gap in understanding how to optimize adaptation in the field of vegetation monitoring. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a non-destructive monitoring method and device for plant traits based on deep learning, which effectively solves the problem of low generalization and lack of universality in existing methods for monitoring plant traits.

[0005] In a first aspect, the present application provides a non-destructive monitoring method for plant traits based on deep learning, which comprises: Spectrally measuring a plurality of plant data sets with plant species in a target spectral range, and obtaining a plurality of sets of spectral data by a spectral radiometer during the measurement process, the plurality of sets of spectral data being generated by a plurality of plant trait parameters; different plant data sets correspond to different ecosystems; Extracting a plurality of sets of label pairs from reflection simulation spectra and transmission simulation spectra generated based on the plurality of sets of spectral data and the plurality of plant trait parameters, respectively, and encoding the plurality of sets of label pairs to obtain a plurality of spectral features; establishing a spectral feature map based on the plurality of spectral features, and inputting the spectral feature map into a pre-trained trait prediction model; the trait prediction model is established and trained based on deep learning; controlling the trait prediction model to process the spectral feature map to obtain plant trait data, so as to complete nondestructive monitoring of plant traits in different ecosystems based on the plant trait data. With reference to the first aspect, in a first possible implementation manner of the first aspect, the specific steps of establishing and training the trait prediction model based on deep learning include: Based on the spectral range features and spectral patterns, an initial trait prediction model based on deep learning is constructed. The initial trait prediction model is trained and tested multiple times through a training set and a test set generated by a plant data set, so as to obtain the trait prediction model.

[0006] With reference to the first aspect, in a second possible implementation manner of the first aspect, the initial trait prediction model based on deep learning is constructed based on the spectral range features and spectral patterns, and includes: A pre-training encoder based on a residual network of deep learning is constructed based on the spectral range features and spectral patterns in advance. A multilayer perceptron, a gradient reversal layer and a domain discriminator head are connected at an output end of the pre-training encoder, so as to construct the initial trait prediction model.

[0007] With reference to the first aspect, in a third possible implementation manner of the first aspect, the initial trait prediction model is constructed by connecting the multilayer perceptron, the gradient reversal layer and the domain discriminator head at the output end of the pre-training encoder, and includes: A corresponding loss function is set in the multilayer perceptron and the domain discriminator head respectively to optimize the initial trait prediction model for the first time. Based on the positional relationship of the multilayer perceptron and the domain discriminator head, the gradient reversal layer is set to optimize the initial trait prediction model for the second time.

[0008] With reference to the first aspect, in a fourth possible implementation manner of the first aspect, the initial trait prediction model is trained and tested multiple times through a training set and a test set generated by a plant data set, so as to obtain the trait prediction model, and includes: A plurality of generation modes for the training set and the test set are set to perform corresponding verification and generate corresponding verification results. The verification results corresponding to the plurality of generation modes are compared to evaluate the trait prediction model based on the comparison results.

[0009] In combination with the first aspect, the present application provides a fifth possible implementation of the first aspect, wherein extracting multiple sets of label pairs from the reflectance simulation spectra and the transmittance simulation spectra generated based on the multiple sets of spectral data and multiple plant trait parameters, respectively, includes: Obtaining a reflection spectrum curve and a transmission spectrum curve based on the reflectance values ​​of the reflection simulation spectrum and the transmission simulation spectrum at different wavelengths; The reflectance spectrum curve and the transmission spectrum curve are integrated with corresponding multiple plant trait parameters to form the label pair.

[0010] In combination with the first aspect, the embodiment of the present application provides a sixth possible implementation of the first aspect, wherein, after controlling the trait prediction model to process the spectral feature map to obtain plant trait data, the method includes: Evaluate the plant trait data based on multiple evaluation dimensions to obtain corresponding evaluation results; The evaluation results corresponding to the multiple evaluation dimensions are fused to obtain a fused evaluation result, so as to evaluate the trait prediction model based on the fused evaluation result.

[0011] In a second aspect, an embodiment of the present application provides a non-destructive monitoring device for plant traits based on deep learning, the device comprising: The measurement module is used to perform spectral measurements on multiple plant datasets containing plant species within the target spectral range, and to generate multiple sets of spectral data by obtaining various plant trait parameters during the measurement process through the spectroradiometer; different plant datasets correspond to different ecosystems; A generating module, configured to extract a plurality of tag pairs from the reflection simulated spectra and the transmission simulated spectra generated based on the plurality of spectral data sets and a plurality of plant trait parameters, and encode the plurality of tag pairs to obtain a plurality of spectral features; An establishment module is used to establish a spectral feature map based on the multiple spectral features, and input the spectral feature map into a pre-trained trait prediction model; the trait prediction model is established and trained based on deep learning; The monitoring module is used to control the trait prediction model to process the spectral feature map to obtain plant trait data, so as to complete non-destructive monitoring of plant traits in different ecosystems based on the plant trait data. In a third aspect, an embodiment of the present application provides an electronic device comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the steps of any one of the methods for non-destructive monitoring of plant traits based on deep learning are performed.

[0012] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium, which stores a computer program. When the computer program is run by a processor, the steps of any one of the plant trait nondestructive monitoring methods based on deep learning are performed.

[0013] The embodiments of the present application provide a plant trait nondestructive monitoring method based on deep learning. The method first performs spectral measurement on multiple plant data sets with plant species in a target spectral range, and obtains multiple sets of spectral data by a spectral radiometer in the measurement process of multiple plant trait parameters. Different plant data sets correspond to different ecosystems. Second, multiple sets of label pairs are extracted from the reflection simulation spectrum and the transmission simulation spectrum generated based on the multiple sets of spectral data and the multiple plant trait parameters, and the multiple sets of label pairs are encoded to obtain multiple spectral features. Then, a spectral feature map is established based on the multiple spectral features, and the spectral feature map is input into a pre-trained trait prediction model. The trait prediction model is established and trained based on deep learning. Finally, the trait prediction model processes the spectral feature map to obtain plant trait data, so as to complete nondestructive monitoring of plant traits in different ecosystems based on the plant trait data. Based on the above method, nondestructive monitoring of plant traits is realized on the deep learning architecture, and nondestructive monitoring of plant traits in different ecosystems is also realized, which improves the generalization and universality of nondestructive monitoring, and solves the problem of low generalization and universality of existing plant trait monitoring methods. BRIEF DESCRIPTION OF DRAWINGS In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0014] Figure 1 A flowchart of a plant trait nondestructive monitoring method based on deep learning provided by the embodiments of the present application is shown; Figure 2 Another flowchart of a plant trait nondestructive monitoring method based on deep learning provided by the embodiments of the present application is shown; Figure 3 Distribution characteristics of three plant trait parameters in different data sets (D1-D5) and their mutual relationship provided by the embodiments of the present application are shown; Figure 4 An architecture diagram of ResNet provided by the embodiments of the present application is shown; Figure 5A schematic diagram of generation of the test set and the training set provided by the embodiments of the present application is shown. Figure 6 A t-SNE visualization diagram of the features extracted from different data sets provided by the embodiments of the present application is shown. Figure 7 A performance and evaluation schematic diagram of the trait prediction model constituted by different models provided by the embodiments of the present application under various plant shape parameters is shown. Figure 8 A performance schematic diagram of the estimation of CHL of different data sets using PLSR and PPADA network models provided by the embodiments of the present application is shown. Figure 9 A performance schematic diagram of the estimation of EWT of different data sets using PLSR and PPADA network models provided by the embodiments of the present application is shown. Figure 10 A performance schematic diagram of the estimation of LMA of different data sets using PLSR and PPADA network models provided by the embodiments of the present application is shown. Figure 11 (a) in the figure shows the average value, standard deviation and coefficient of variation of spectral reflectance, (b) shows the Pearson correlation coefficient relationship diagram of spectral reflectance and traits. Figure 12 A structural block diagram of a plant trait nondestructive monitoring device based on deep learning provided by the embodiments of the present application is shown. Figure 13 A structural diagram of an electronic device provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0015] To make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of illustration and description, and do not serve to limit the protection scope of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowchart shows the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowchart can not be implemented in sequence, and the steps without logical contextual relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowchart or one or more operations can be removed from the flowchart under the guidance of the content of the present application.

[0016] In addition, the described embodiments are only some embodiments of the present application, not all embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0018] The existing models used in hyperspectral remote sensing mainly focus on single ecosystem applications (forests or croplands), in which the models are trained and validated in homogeneous environments, making it difficult to generalize between heterogeneous ecosystems and sensor types, which limits their universality in heterogeneous ecosystems, such as the transition from controlled croplands to natural grasslands or mixed forests, where spectral feature relationships can change significantly due to differences in species composition, canopy structure, and environmental stressors. There is a serious gap in understanding how to optimize adaptation in the field of vegetation monitoring, as there are few studies that systematically evaluate the ability of transfer learning to enhance cross-ecosystem feature prediction.

[0019] Based on this, the embodiments of the present application provide a deep learning-based non-destructive monitoring method and device for plant traits, which are described below through embodiments.

[0020] Embodiment 1 In order to facilitate the understanding of the present embodiment, first of all, a deep learning-based non-destructive monitoring method for plant traits disclosed in the embodiments of the present application is introduced in detail. As shown in a flowchart of a deep learning-based non-destructive monitoring method for plant traits, Figure 1 As shown in a flowchart of a deep learning-based non-destructive monitoring method for plant traits, Figure 2 As shown in a flowchart of a deep learning-based non-destructive monitoring method for plant traits, the method provided by the present application comprises: S101, performing spectral measurement on a plurality of plant data sets with plant species in a target spectral range, and obtaining a plurality of groups of spectral data by a spectral radiometer in the measurement process, the plurality of plant trait parameters; different plant data sets correspond to different ecosystems; S102, extracting a plurality of groups of label pairs from the reflection simulation spectrum and the transmission simulation spectrum based on the plurality of groups of spectral data and the plurality of plant trait parameters, and encoding the plurality of groups of label pairs to obtain a plurality of spectral features; S103, establishing a spectral feature map based on the plurality of spectral features, and inputting the spectral feature map into a pre-trained trait prediction model; the trait prediction model is established and trained based on deep learning; S104, controlling the trait prediction model to process the spectral feature map to obtain plant trait data, so as to complete non-destructive monitoring of plant traits in different ecosystems based on the plant trait data. In step S101, a plurality of plant data sets with plant species are set, which are data set 1 (D1), data set 2 (D2), data set 3 (D3), data set 4 (D4) and data set 5 (D5). The data set 1 (D1) contains data of five forest plant species mainly distributed in temperate regions; the data set 2 (D2) contains six tropical plant species across subtropical and tropical ecosystems; the data set 3 (D3) includes two herbaceous species cultured under controlled laboratory conditions; the data set 4 (D4) contains two xerophyte species adapted to arid environments and water-limited growth conditions; and the data set 5 (D5) focuses on three crop species planted in agricultural research areas. The above different plant data sets correspond to different ecosystems, thereby realizing cross-ecosystem research. Fresh leaves in the above five data sets are measured by spectrum according to standard collection protocols. The spectral range of four data sets is 350-2500 nm, and the spectral range of one data set is 400-2400 nm. In order to ensure the consistency and comparability of all data sets, the main spectral range (400-2400 nm) of leaf reflectivity is selected as the reference for modeling LMA, CHL and EWT, and a plurality of spectral data sets are generated by obtaining a plurality of plant trait parameters in the measurement process by a spectral radiometer. The spectral radiometer is a high-resolution spectral radiometer with a resolution of 1 nm. The plurality of plant trait parameters are chlorophyll content (CHL), leaf water thickness (EWT) and leaf mass per unit area (LMA).

[0021] Figure 3 The distribution characteristics of three plant traits in different data sets (D1-D5) and their relationships are shown, Figure 3The distribution of chlorophyll content (CHL), equivalent water thickness (EWT), and leaf mass per area (LMA) in each of (a) (b) (c) in the figure, (d) the Pearson correlation coefficient between traits. Overall, the distribution of the three traits shows the characteristics of skew normal distribution. Specifically, the distribution curve of CHL, EWT and LMA in each data set shows different peak values, and the peak position and distribution range vary with the data set. In comparing the traits, the ridge distribution of EWT and LMA is relatively similar, both fluctuating on the right side of the peak. In contrast, CHL shows a distribution with a peak moving to the right. In addition, the correlation coefficients between traits also differ: EWT and LMA are strongly positively correlated (0.66), CHL and LMA are moderately negatively correlated (-0.31), and CHL and EWT are weakly negatively correlated (-0.15).

[0022] In step S102, the present application sets up a PROSPECT radiation transfer model based on physical optical principles. This model quantitatively describes the multiple scattering and absorption effects inside the leaf through the radiation transfer equation in layered media, effectively capturing the nonlinear effects of various plant traits on spectral response. By coupling anatomical and biochemical parameters, including chlorophyll content (CHL), equivalent water thickness (EWT), leaf mass per area (LMA), and solar incident angle and leaf surface roughness, the model establishes simulated reflectance and transmittance spectra with a spectral resolution of 1 nm for multiple plant leaves in the 400-2400 nm range. Throughout this process, data generation is limited to empirical observations to maintain biological plausibility, and multiple sets of label pairs are extracted from the simulated reflectance and transmittance spectra. A ResNet-based encoder network encodes these label pairs to obtain spectral features. This framework encodes the biophysical mechanisms of light-trait interactions, and the ResNet encoder network with gradient reversal layers adjusts the feature distribution of heterogeneous ecosystems based on adversarial domain adaptation. This dual approach enables the model to learn domain-invariant representations, effectively mitigating spectral heterogeneity and sensor-induced domain shifts.

[0023] In the implementation of step S102, one embodiment is that the multiple sets of label pairs are extracted from the simulated reflectance and transmittance spectra based on the multiple sets of spectral data and multiple plant trait parameters, including: S1021, based on the reflectance values of the simulated reflectance and transmittance spectra at different wavelengths, obtaining reflectance and transmittance spectral curves; S1022, integrating the reflectance and transmittance spectral curves with the corresponding multiple plant trait parameters to form the label pairs.

[0024] In steps S1021-S1022, the present application draws the reflection spectrum curve and the transmission spectrum curve based on the reflectance values of plant leaves at different wavelengths and the proportion of light passing through plant leaves. The present application uses Latin hypercube sampling strategy to ensure the spatial consistency of parameter sampling. The reflection spectrum curve and the transmission spectrum curve are integrated with the corresponding multiple plant trait parameters to form the tag pair. The tag pair is the reflection spectrum curve and the transmission spectrum curve corresponding to the chlorophyll content, leaf thickness, water content and other parameters of each data of the plant. The number of tag pairs can reach 10,000 pairs. The multiple sets of spectral data can also be aligned (to ensure that the spectral data and the multiple plant trait parameters are correctly matched), noise processed (the spectral data may be affected by environmental factors), and multi-parameter collinearity (some trait parameters may be highly correlated and need to be selected or reduced in dimension).

[0025] In step S103, the present application encodes the feature vectors obtained based on the extracted multiple sets of tag pairs, processes the multiple spectral features based on the ResNet encoder network to establish a spectral feature map, and selects a suitable map form according to the data characteristics and application requirements, such as a two-dimensional heat map, an isogram, or a three-dimensional stereogram. The resolution, color coding rules and other parameters of the map are determined. For example, for a heat map, different colors can be assigned according to the size of the spectral feature value, and the color depth represents the height of the feature value. The spectral feature map is input into a pre-trained trait prediction model. The trait prediction model is established and trained based on deep learning. The core idea of the trait prediction model is to fully utilize large-scale simulated spectral data generated by a radiation transfer model for pre-training. The radiation transfer model combines the physical prior of the simulated spectrum and the adversarial domain self-adaptation, and then uses a limited number of target domain samples for cross-domain adaptive fine-tuning. The pre-training encoder in the first stage is supervised and fine-tuned in the second stage. Multilayer perceptron. MSE loss: mean square error loss, BCE loss: binary cross-entropy loss.

[0026] In the implementation process of step S103, there is an embodiment that the specific steps of establishing and training the trait prediction model based on deep learning include: S1031, based on the range characteristics and spectral patterns of the spectrum, an initial trait prediction model based on deep learning is constructed. S1032, the initial trait prediction model is trained and tested multiple times through the training set and test set generated by the plant data set to obtain the trait prediction model.

[0027] In steps S1031-S1032, the application constructs an initial trait prediction model based on the ResNet PROSPECT overall architecture of deep learning, which is constructed for the range characteristics and spectral patterns of the spectrum, and can completely identify the plant traits. The range characteristics are the wide range characteristics of 400-2400 nm. The initial trait prediction model is trained and tested multiple times by the training set and test set generated by the plant data set, and the initial trait prediction model after multiple training and testing is the required trait prediction model, which can realize cross-ecosystem and non-destructive monitoring.

[0028] During network training, the random gradient descent method is used to update the parameters of the encoder and the two task heads in small batches. The few-shot learning strategy combines a small amount of labeled target domain data with source domain data into the network. The feature alignment mechanism reduces domain transfer and promotes cross-domain knowledge transfer between datasets. The model is trained using a stage optimization strategy. In the first pre-training stage, we use mini-batch training with 64 batches per epoch to configure the training to run for 200 epochs. The Adam optimizer is used with an initial learning rate of 0.001 to ensure fast early convergence speed, and the learning rate is gradually reduced in the later stage with a decay factor of 1x10-5 to reduce overfitting. In the second stage of transfer learning, the epoch number is adjusted to 100 and the batch size is 32 in the case of integrating source and target domain data, and the initial learning rate is updated to 1x10^-4 to facilitate more fine-tuning and domain adaptation.

[0029] The application also evaluates the trait prediction model. To measure the variability of the dependent variable and the prediction error, the normalized root mean square error (nRMSE) and the determination coefficient (R²) are as follows: ; Where n is the number of samples; And represent the measured and predicted yield of sample i, respectively. The higher the R2 value and the lower the nRMSE value, the better the performance of the model.

[0030] In the specific implementation process of step S1031, there is an embodiment that: the range characteristics and spectral patterns based on the spectrum construct an initial trait prediction model based on deep learning, which includes: S10311, an initial trait prediction model based on a pre-trained encoder based on a residual network of deep learning is constructed based on the range characteristics and spectral patterns of the spectrum; S10312, a multilayer perceptron, a gradient reversal layer, and a domain discrimination head are connected to the output end of the pre-trained encoder to construct an initial trait prediction model.

[0031] In steps S10311-S10312, the pre-training encoder is constructed based on the deep learning residual network, and the architecture of the pre-training encoder processes the spectral input through a hierarchical feature extraction pipeline. Initially, the 2000-band spectral data is reshaped into a 224x224x3 tensor to adapt to the convolution operation. The network starts with a 7x7 convolutional layer (64 filters, stride=2), followed by a 3x3 max-pool, to establish a preliminary spatial-spectral representation. The subsequent residual blocks employ a bottleneck structure with concatenated 1x1 and 3x3 convolutions, gradually expanding the channel size from 64 to 1024 through four main stages. Each stage contains multiple identified shortcut blocks, where the 1x1 convolution performs channel dimension matching, and the 3x3 convolution extracts spatially invariant features. Notably, the architecture implements a channel scaling factor of x4 between stages (64→256→512→1024), maintaining computational efficiency through bottleneck compression. The deep stack of 21 convolutional layers maintains gradient flow with residual connections, systematically reducing the spatial resolution of feature maps with stride convolutions in transition blocks. This design effectively learns multi-scale spectral spatial correlations while alleviating the vanishing gradient problem inherent to deep networks. Figure 4 An architecture diagram of ResNet for yield prediction, with batch normalization and ReLU after each convolutional layer. A multilayer perceptron, gradient reversal layer, and domain discriminator head are connected at the output of the pre-training encoder to construct an initial trait prediction model, which is the ResNet PROSPECT overall architecture, informed by pre-training with simulated data initialized weights generated by the PROSPECT-D radiative transfer model. The network is first pre-trained on 10000 synthetic spectral features pairs covering the entire parameter space for 100 epochs, then fine-tuned on the experimental dataset, testing the isolation effect of the physical priors. A fully connected regression head is appended to the encoder output. The final encoder outputs a high-dimensional latent representation suitable for downstream regression tasks through the appended task-specific head. We optimize the network by minimizing the mean squared error (MSE) loss function to predict LMA, EWT, and CHL values, and learn the latent representation strongly correlated with plant traits from large-scale simulated data through gradient backpropagation in the network. The ResNet-GRL architecture enables physically unconstrained domain adversarial learning by adding GRL between the feature extractor and the domain classifier. The dual-objective network simultaneously minimizes the feature prediction error and maximizes the domain confusion through adversarial training (λ=0.3), thus evaluating the independent contribution of domain-invariant feature learning.

[0032] The present application utilizes the spectral feature representation ability obtained from large-scale simulation data, and at the same time fine-tunes and domain aligns through limited target domain samples, so as to improve the prediction accuracy in the actual target environment. In order to solve the problems of plant trait prediction and domain adaptation at the same time, two parallel task heads are introduced at the output end of the pre-training encoder: (1) a multi-layer perceptron (MLP) is used to regress three plant traits (LMA, EWT and CHL) as a downstream task. (2) A domain discrimination head is used to distinguish between source domain and target domain inputs, which is implemented as a compact convolutional or fully connected classifier. In order to extract region-robust features.

[0033] In the implementation process of step S10312, there is an embodiment that: the multi-layer perceptron, the gradient reversal layer and the domain discrimination head are connected at the output end of the pre-training encoder to construct an initial trait prediction model, including: S103121, a corresponding loss function is set in the multi-layer perceptron and the domain discrimination head to optimize the initial trait prediction model for the first time; S103122, based on the positional relationship of the multi-layer perceptron and the domain discrimination head, the gradient reversal layer is set to optimize the initial trait prediction model again.

[0034] In this step, the mean square error (MSE) loss function is set for the multi-layer perceptron to predict LMA, EWT and CHL values, and the binary cross entropy (BCE) loss is used for training the domain discrimination head, and the MSE loss function is defined as follows: ; Where m represents the trait index, and m represents the number of samples. The loss of all features is combined: ; Based on the positional relationship of the multi-layer perceptron and the domain discrimination head, the gradient reversal layer is set to optimize the initial trait prediction model again, and the gradient reversal layer (GRL) is integrated before the domain discrimination head. In the forward propagation process, GRL acts as an identity transformation, while in the backward propagation process, it reverses the gradient from the domain discrimination loss, thereby encouraging the encoder to learn domain-invariant features through adversarial confusion. By simultaneously minimizing feature prediction error and maximizing domain discriminator confusion, the network achieves improved cross-domain generalization while maintaining downstream task performance. The comprehensive objective function of this stage includes two parts:

[0035] Where represents the MSE loss of plant trait regression, represents the adversarial loss of domain discrimination, which can be realized by binary cross entropy or other equivalent forms. and are hyperparameters that control the weights of the two loss terms, and their optimal balance can be determined by hyperparameter tuning on a validation set.

[0036] In the implementation process of step S1032, there is an embodiment that: the training set and the test set generated by the plant data set are used to train and test the initial trait prediction model multiple times to obtain the trait prediction model, including: S10321, multiple generation modes for the training set and the test set are set to perform corresponding verification and generate corresponding verification results; S10322, the verification results corresponding to the multiple generation modes are compared to evaluate the trait prediction model according to the comparison results.

[0037] In the step, two different verification methods are used to comprehensively evaluate the model performance: (1) aggregated cross-validation (such as Figure 5 (a)): all samples from five data sets (D1-D5) are collected into a composite repository, and then a stratified five-fold cross-validation scheme is used, in which each fold maintains the proportion of the original data set distribution. That is, in this way, the generation mode of the training set and the test set is data set merging, and in each iteration, 80% of the samples are allocated for training (15% are reserved for internal validation to facilitate early stopping), and 20% are used for testing. This method evaluates the overall prediction accuracy under the assumption that the data distribution is uniform. (2) Cross-data set validation (such as Figure 5 (b)): In order to strictly evaluate the transferability across domains, a data set exclusion experiment is performed. In each experiment, four data sets are used as the training set, and the remaining data set is used for independent testing. That is, in this way, each data set is a training set and a test set, and spatial spectral standardization is applied to each data set using the respective training statistics to prevent information leakage. This protocol specifically quantifies the generalization ability of the model under heterogeneous data acquisition conditions.

[0038] Figure 6t-SNE visualizations showing the distribution of features from multiple datasets comparing ResNet-extracted features to those processed by PPADA Net using adversarial domain adaptation. The left panel shows ResNet-derived features where color-coded data points from different datasets form distinct, well-separated clusters with minimal overlap. This sharp separation stems from domain-specific biases in spectral reflectance patterns leading to distribution mismatch, which hinders the accuracy of cross-dataset predictions in transfer learning. In contrast, the right panel shows PPADA Net-processed features where adversarial learning leads to two key transformations: (1) the boundaries between datasets become less distinct, with most data points forming a mixed region neighborhood; (2) previously compact clusters spread out into overlapping distributions, particularly in high-dimensional manifolds related to invariant spectral features. These structural adjustments indicate that PPADA Net enhances knowledge transfer across heterogeneous datasets by learning transfer-invariant representations, effectively mitigating domain shift. This alignment mechanism supports PPADA Net's superior cross-domain generalization performance, where PPADA Net is an acronym for the architecture of the trait prediction model of the present application.

[0039] The performance of PLSR models and the proposed PPADA Net in predicting three plant traits was evaluated through various train-test set assessments, where Figure 8 , Figure 9 , Figure 10 are used to represent PLSR and PPADA Net, respectively, with D1-D2 indicating dataset 1 as the training set and dataset 2 as the test set. PPADA-Net consistently outperformed PLSR models for all three traits across most data set combinations. On average, PPADA Net exhibited higher R² values and lower nRMSE values than PLSR. For example, in predicting CHL, PPADA Net achieved an average R² of 0.53 and an average nRMSE of 0.12 across all data set combinations, while PLSR had an average R² of 0.18 and an average nRMSE of 0.25. Similarly, PPADA Net achieved an average R² of 0.72 and an average nRMSE of 0.10 in predicting LMA, while PLSR had an R² of 0.46 and an nRMSE of 0.15. This demonstrates that PPADA Net, which combines data pre-training with adversarial domain alignment, significantly improves prediction accuracy and reliability.

[0040] The performance of trait prediction models varies greatly across different dataset combinations. For example, in predicting LMA, PPADA Net has an R² of 0.68 and nRMSE of 0.11 when using D1 as the training set and D2 as the test set, while PLSR has an R² of 0.34 and nRMSE of 0.15. Similarly, for the D5-D4 combination, PPADA Net has an R² of 0.81 and nRMSE of 0.07, while PLSR has an R² of 0.22 and nRMSE of 0.22. These examples further emphasize the significant influence of the training set on the prediction accuracy of data-driven models. Under these conditions, PPADA Net exhibits greater stability and adaptability across various dataset combinations. For example, in predicting EWT across different dataset combinations, PLSR has an R² ranging from 0.18 to 0.58 and nRMSE ranging from 0.13 to 0.24. In contrast, PPADA Net has an R² ranging from 0.53 to 0.82 and nRMSE ranging from 0.07 to 0.12, indicating that the integration of simulation data pre-training and adversarial domain alignment enables the model to better handle domain transfer and maintain consistent performance across different data sources, which is particularly valuable in the presence of high data heterogeneity.

[0041] In step S104, the present application inputs the spectral feature map into the trait prediction model of the deep learning-based ResNetPROSPECT overall architecture, then controls the trait prediction model to process the spectral feature map to obtain plant trait data, so as to complete non-destructive monitoring of plant traits in different ecosystems based on the plant trait data, and achieve the monitoring effect across ecosystems, that is, across domains. The present application also systematically compares four regression methods for CHL, EWT, and LMA prediction: (1) traditional multivariate PLSR, (2) data-driven ResNet, and two ablation variants, (3) physically enhanced ResNet foreground, and (4) domain-adaptive ResNet GRL. These methods are rigorously benchmarked against our proposed PPADA network to evaluate the incremental benefits of comprehensive physical adversarial learning. As shown in Figure 7As shown, models exhibited significant differences in R2and nRMSE. Overall, the accuracy of the traditional machine learning model PLSR was relatively low, with average R2values of 0.59, 0.63, and 0.72 for CHL, EWT, and LMA, respectively. In contrast, the deep learning-based ResNet improved the R2for CHL to 0.63 (an increase of 6.8%) through nonlinear feature extraction, although its EWT prediction slightly decreased (R2= 0.61 vs 0.63), highlighting the limitations of purely data-driven end-to-end training for cross-feature generalization. Notably, combining physical priors and domain adaptation strategies significantly improved model performance. For example, the ResNet PROSPECT pre-trained on simulated data generated from the forward radiative transfer model, achieved R2values of 0.67, 0.74, and 0.80 for CHL, EWT, and LMA, respectively, an average improvement of 8.3% over the basic ResNet, and showed the largest error reduction in EWT prediction (nRMSE = 0.09 vs 0.15). Furthermore, the ResNet GRL employed adversarial learning to align features between the source and target domains, resulting in a marginal improvement in CHL prediction (R2= 0.68 vs 0.67), but slightly worse LMA performance (R2= 0.78 vs 0.80) compared to the ResNet PROSPECT, indicating that physical simulation data pre-training is more beneficial for certain traits, such as LMA. The dual-strategy PPADA network, which combined physical constraints and domain adaptation, achieved the best overall performance by increasing the R2values for CHL, EWT, and LMA to 0.72, 0.77, and 0.86, respectively, an average improvement of 5.1% compared to the single-strategy models, and showed balanced improvements across different traits. From an individual trait perspective, LMA had the highest prediction accuracy (PPADA net R2= 0.86), which could be due to the more apparent physical link between its spectral features and dry matter content.

[0042] The PPADA network of the present application introduces a two-stage training strategy that integrates radiative transfer simulation and adversarial learning. By pre-training a ResNet encoder of synthetic spectra generated based on foreground D, the framework encodes the biophysical mechanisms of light-trait interactions, while adversarial domain adaptation with gradient reversal layers adjusts the feature distribution of heterogeneous ecosystems. This dual approach enables the model to learn domain-invariant representations, effectively mitigating spectral heterogeneity and sensor-induced domain shifts. Rigorous validation on five independent datasets demonstrates the robustness of PPADA Net. The framework achieves state-of-the-art accuracy in predicting chlorophyll content (CHL: R²=0.72), equivalent water thickness (EWT: R²=0.77), and leaf mass per area (LMA: R²=0.86), outperforming traditional PLSR (R² improvement of 35-52%) and data-driven ResNet models (gain of 5.1-22.4%). Cross-dataset validation highlights its generalization capability, with stable performance in challenging scenarios. The adversarial component reduces feature separation in specific domains, as shown in t-SNE visualizations of aligned latent spaces. In summary, PPADA network reconciles physics-driven prior knowledge with data-driven adaptability, providing a transformative solution for plant trait estimation in heterogeneous environments.

[0043] The transferability of spectral feature prediction models is crucial for practical applications, as ecosystems exhibit great variability in spectral features and plant trait distributions due to differences in species composition, environmental conditions, and measurement protocols. Traditional machine learning and deep learning models often struggle to generalize across heterogeneous datasets, limiting their effectiveness in large-scale monitoring. For example, PLSR, while computationally efficient, is less adaptable to spectral heterogeneity. In cross-dataset validation, PLSR's average R² for CHL was only 0.18 ( Figure 8 ), and 0.46 ( Figure 10 ) for LMA, with an nRMSE value of 0.25. Similarly, deep learning models like ResNet, despite their multi-task learning capabilities, are still susceptible to specific domain biases.

[0044] Adversarial domain adaptation has emerged as a powerful strategy to adjust functional distribution across domains. For example, Ma et al. (2022) proposed a domain adaptation scheme called adversarial entropy optimization to learn domain-invariant features, achieving state-of-the-art performance across different domain adaptation tasks. In this application, the PPADA network combines this technique with physical priors to address spectral heterogeneity. The average R² of CHL, LMA, and EWT is 0.53, 0.72, and 0.65, respectively, outperforming PLSR by 35–52% in terms of R² across all dataset combinations. Notably, the PPADA Net maintains stable performance even in challenging scenarios such as D5→LMA (R²=0.81 vs. 0.22 for PLSR), demonstrating its robustness to domain-specific noise. The enhanced transferability comes from two synergistic mechanisms. First, PROSPECT-D-simulated spectra provide feature-specific spectral patterns, enabling the model to prioritize domain-invariant biochemical signals over dataset-specific artifacts. Second, by minimizing the difference between source and target domains in the high-dimensional feature space, the PPADA network alleviates overfitting to local spectral variations. Future research can extend this framework to dynamically adapt to emerging ecosystems or sensor types, further bridging the gap between controlled simulations and field applications. By combining physical mechanisms with adaptive learning, the PPADA network sets a precedent for developing robust, scalable tools in precision ecology.

[0045] In the implementation process of step S104, there is an embodiment that after the control of the trait prediction model processes the spectral feature map to obtain plant trait data, the following steps are included: S1041, respectively evaluate the plant trait data based on multiple evaluation dimensions to obtain corresponding evaluation results; S1042, fuse the evaluation results corresponding to the multiple evaluation dimensions to obtain a fusion evaluation result, and evaluate the trait prediction model based on the fusion evaluation result.

[0046] In steps S1041-S1042, Figure 11(a) Statistical analysis of spectral reflectance at different wavelengths was presented. The average reflectance values from the five datasets showed consistent trends across the spectral range of 400-2400 nm, all showing different spectral peaks and troughs at specific wavelengths. In the visible spectrum (500-700 nm), relatively low and stable average reflectance values were observed, followed by a rapid increase to higher values in the near-infrared (NIR, 700-1300 nm) region. Subsequently, a fluctuating decrease occurred in the short-wave infrared (SWIR, 1300-2400 nm) range, with a significant decrease in reflectance in the 1900-2100 nm sub-band. Significant differences in average reflectance values between datasets were observed, particularly in specific spectral bands. Standard deviation curves showed the dispersion patterns of reflectance in different spectral regions. The larger standard deviation in the visible spectrum indicated greater variability between datasets, while the deviation decreased in the NIR and SWIR regions, especially at 1000-1300 nm and 1900-2100 nm. The coefficient of variation (CV) curves further elucidated the patterns of relative reflectance variation. The substantial CV fluctuations across the entire spectral range indicated significant differences in reflectance variability between bands.

[0047] Figure 11 (b) The correlations between spectral reflectance at different wavelengths and plant traits were presented. Overall, the relationships between reflectance and plant traits showed significant differences across different spectral bands. Chlorophyll had a higher correlation coefficient in the visible band (400-700 nm), while the correlation coefficient gradually decreased in the near-infrared and short-wave infrared bands. All coefficients in these regions remained below 0.4, except for the 1900-2400 nm band. The correlation pattern of EWT showed distinct characteristics. In the 400-700 nm range, the correlation coefficient was relatively low. Higher correlation coefficients (>0.5) mainly occurred in the 700-1900 nm range, indicating the effectiveness of these spectral regions for EWT prediction. LMA was highly significantly positively correlated with leaf dry matter content (700-1400 nm), indicating that reflectance in this range was significantly positively correlated with leaf dry matter content. Within specific spectral bands, correlation coefficients varied between different datasets. For example, dataset D2 had significantly higher correlations with LMA than other datasets, while dataset D5 had lower correlations with EWT, highlighting the complexity of trait spectrum relationships and dataset heterogeneity.

[0048] Example 2 The present application also provides a deep learning-based non-destructive monitoring device for plant traits, such as Figure 12Fig. 1 shows a block diagram of a device for non-destructive monitoring of plant traits based on deep learning, which realizes the functions corresponding to the steps of the method for non-destructive monitoring of plant traits based on deep learning executed on a terminal device. The device can be understood as a component of a server including a processor. The device for non-destructive monitoring of plant traits based on deep learning includes: a measurement module 1201 configured to perform spectral measurements on a plurality of plant data sets with plant species in a target spectral range, and obtain a plurality of sets of spectral data by a spectral radiometer, wherein different plant data sets correspond to different ecosystems; a generation module 1202 configured to extract a plurality of sets of label pairs from reflection simulation spectra and transmission simulation spectra generated based on the plurality of sets of spectral data and a plurality of plant trait parameters, and encode the plurality of sets of label pairs to obtain a plurality of spectral features; an establishment module 1203 configured to establish a spectral feature map based on the plurality of spectral features, and input the spectral feature map into a pre-trained trait prediction model, wherein the trait prediction model is established and trained based on deep learning; a monitoring module 1204 configured to control the trait prediction model to process the spectral feature map to obtain plant trait data, and complete non-destructive monitoring of plant traits in different ecosystems based on the plant trait data.

[0049] In an available implementation, the establishment module includes: a first construction module configured to construct an initial trait prediction model based on deep learning based on spectral range features and spectral patterns; a test module configured to perform multiple training and testing on the initial trait prediction model by a training set and a test set generated by the plant data set, to obtain the trait prediction model.

[0050] In an available implementation, the establishment module further includes: a second construction module configured to construct a pre-training encoder based on deep learning residual network in advance based on spectral range features and spectral patterns; a connection construction module configured to connect a multi-layer perceptron, a gradient reversal layer, and a domain discriminator head at an output end of the pre-training encoder to construct the initial trait prediction model.

[0051] In an available implementation, the establishment module also includes: an optimization module configured to set corresponding loss functions in the multi-layer perceptron and the domain discriminator head to initially optimize the initial trait prediction model; The setting module is configured to set the gradient reversal layer to re-optimize the initial trait prediction model based on a positional relationship of the multi-layer perceptron and the domain discriminator head.

[0052] In an implementation, the establishing module further includes: The verification module is configured to set multiple generation manners for the training set and the test set to perform corresponding verifications and generate corresponding verification results. The comparison module is configured to compare the verification results of the multiple generation manners to generate a comparison result to evaluate the trait prediction model.

[0053] In an implementation, the generation module includes: The reflection construction module is configured to obtain a reflection spectrum curve and a transmission spectrum curve based on the reflectance values of the reflection simulation spectrum and the transmission simulation spectrum at different wavelengths. The integration module is configured to integrate the reflection spectrum curve and the transmission spectrum curve with the corresponding multiple plant trait parameters to form the label pair.

[0054] In an implementation, the monitoring module includes: The evaluation module is configured to evaluate the plant trait data based on multiple evaluation dimensions to obtain corresponding evaluation results. The fusion module is configured to fuse the evaluation results of the multiple evaluation dimensions to obtain a fusion evaluation result, and evaluate the trait prediction model based on the fusion evaluation result. Embodiment 3 The present application also provides an electronic device, such as Figure 13 As shown in the figure, it includes a processor 1301, a memory 1302, and a bus 1303. The memory 1302 stores machine-readable instructions executable by the processor 1301. When the electronic device is running, the processor 1301 communicates with the memory 1302 through the bus 1303. The machine-readable instructions are executed by the processor 1301 to perform the steps of any one of the plant trait nondestructive monitoring methods based on deep learning.

[0055] Embodiment 4 The present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it performs the steps of any one of the plant trait nondestructive monitoring methods based on deep learning.

[0056] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the system and the device described above can refer to the corresponding process in the method embodiment, and will not be repeated in the present application. In the several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented by other means. The above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and the actual implementation can have another division manner, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed mutual elements can be indirect coupling or communication connection through some communication interface, device or module, which can be electrical, mechanical or other forms.

[0057] The modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical units, i.e. can be located in one place or distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0058] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0059] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a non-volatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present application essentially or the parts of the prior art that make contributions or parts of the technical solutions can be embodied in the form of software products, which are stored in a storage medium and include a plurality of instructions for causing a computer device (which can be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk and various program code storage media.

[0060] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A non-destructive monitoring method for plant traits based on deep learning, characterized in that: The method comprises: Spectral measurements are performed on multiple plant datasets containing plant species within the target spectral range. Spectral radiometers are used to obtain various plant trait parameters during the measurement process to generate multiple sets of spectral data. Different plant datasets correspond to different ecosystems. Extracting a plurality of tag pairs from the reflection simulation spectra and the transmission simulation spectra generated based on the plurality of spectral data sets and a plurality of plant trait parameters, respectively, and encoding the plurality of tag pairs to obtain a plurality of spectral features; Establishing a spectral feature map based on the multiple spectral features, and inputting the spectral feature map into a pre-trained trait prediction model; the trait prediction model is established and trained based on deep learning; The trait prediction model is controlled to process the spectral feature map to obtain plant trait data, so as to complete non-destructive monitoring of plant traits in different ecosystems based on the plant trait data.

2. The method according to claim 1, characterized in that The specific steps of establishing and training the trait prediction model based on deep learning include: Based on the wide range of spectral features and spectral patterns, an initial trait prediction model based on deep learning was constructed; The initial trait prediction model is trained and tested multiple times using a training set and a test set generated from a plant data set to obtain a trait prediction model.

3. The method according to claim 2, characterized in that The initial trait prediction model based on deep learning is constructed based on the wide range characteristics and spectral patterns of the spectrum, including: A pre-trained encoder based on a deep learning residual network is constructed based on the wide range characteristics and spectral patterns of the spectrum; A multilayer perceptron, a gradient reversal layer, and a domain discrimination head are connected to the output end of the pre-trained encoder to construct an initial trait prediction model.

4. The method according to claim 3, characterized in that The method further comprises connecting a multilayer perceptron, a gradient reversal layer, and a domain discrimination head to the output end of the pre-trained encoder to construct an initial trait prediction model, including: Setting corresponding loss functions in the multilayer perceptron and the domain discrimination head respectively to initially optimize the initial trait prediction model; Based on the positional relationship between the multilayer perceptron and the domain discriminant head, the gradient reversal layer is set to re-optimize the initial trait prediction model.

5. The method according to claim 2, characterized in that The initial trait prediction model is trained and tested multiple times using the training set and the test set generated by the plant data set to obtain the trait prediction model, including: Setting multiple generation methods for the training set and the test set to perform corresponding verification and generate corresponding verification results; The verification results corresponding to the multiple generation methods are compared, and the trait prediction model is evaluated based on the generated comparison results.

6. The method according to claim 1, characterized in that The extracting of multiple sets of label pairs from the reflection simulation spectra and the transmission simulation spectra generated based on the multiple sets of spectral data and multiple plant trait parameters respectively includes: Obtaining a reflection spectrum curve and a transmission spectrum curve based on the reflectance values ​​of the reflection simulation spectrum and the transmission simulation spectrum at different wavelengths; The reflectance spectrum curve and the transmission spectrum curve are integrated with corresponding multiple plant trait parameters to form the label pair.

7. The method according to claim 1, characterized in that After controlling the trait prediction model to process the spectral feature map to obtain plant trait data, the method includes: Evaluate the plant trait data based on multiple evaluation dimensions to obtain corresponding evaluation results; The evaluation results corresponding to the multiple evaluation dimensions are fused to obtain a fused evaluation result, so as to evaluate the trait prediction model based on the fused evaluation result.

8. A non-destructive monitoring device for plant traits based on deep learning, characterized in that: The device comprises: The measurement module is used to perform spectral measurements on multiple plant datasets containing plant species within the target spectral range, and to generate multiple sets of spectral data by obtaining various plant trait parameters during the measurement process through the spectroradiometer; different plant datasets correspond to different ecosystems; A generating module, configured to extract a plurality of tag pairs from the reflection simulated spectra and the transmission simulated spectra generated based on the plurality of spectral data sets and a plurality of plant trait parameters, and encode the plurality of tag pairs to obtain a plurality of spectral features; An establishment module is used to establish a spectral feature map based on the multiple spectral features, and input the spectral feature map into a pre-trained trait prediction model; the trait prediction model is established and trained based on deep learning; The monitoring module is used to control the trait prediction model to process the spectral feature map to obtain plant trait data, so as to complete non-destructive monitoring of plant traits in different ecosystems based on the plant trait data.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate via the bus. When the machine-readable instructions are executed by the processor, the steps of the method for non-destructive monitoring of plant traits based on deep learning are performed.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of a non-destructive monitoring method of plant traits based on deep learning as described in any one of claims 1 to 7.

Images