A novel method and system for hyperspectral analysis of leaf pigments and structural parameters

Abstract

The application relates to the technical field of deep learning, in particular to a novel leaf pigment and structure parameter hyperspectral analysis method and system, which comprises the following steps: obtaining hyperspectral reflectance data of a target leaf; inputting the hyperspectral reflectance data into a pre-trained multi-task deep learning model to output estimated values of multiple leaf physicochemical parameters corresponding to the target leaf; the multi-task deep learning model is a one-dimensional convolutional neural network, and the network structure comprises the following which are connected in sequence: an input layer used for receiving the hyperspectral reflectance data; a feature extraction module used for extracting multi-scale spectral features, including shallow features and deep features; a feature fusion module used for fusing the shallow features and the deep features; and multiple parallel task output layers used for performing parameter-specific estimation based on the fused features and outputting the same. The application effectively improves the accuracy of the analysis results of the leaf pigment and structure parameters by capturing the correlation between the parameters.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a novel hyperspectral analysis method and system for leaf pigments and structural parameters. Background Technology

[0002] Physicochemical parameters of plant leaves (such as chlorophyll a content, chlorophyll b content, carotenoid content, protein content, dry matter content, and leaf thickness) are core indicators for assessing plant growth status. These parameters not only help identify environmental stress factors during plant phenological stages or growth cycles but also assist in predicting crop yields, thus providing important data for agricultural management and ecological monitoring. For example, during crop growth, chlorophyll content directly reflects photosynthetic efficiency, while leaf nitrogen content indicates the plant's nutrient stress status; this information is crucial for precision agriculture. Furthermore, in the context of climate change, these parameters can reveal plant response mechanisms to adversities such as drought, high temperature, or salinity, helping scientists develop more resistant crop varieties. Currently, the mainstream technical approach for non-destructive, large-scale monitoring of these parameters is to combine leaf reflectance spectra with data analysis models for inversion. With technological advancements, this field has reached several stages of technological development, but significant bottlenecks still exist.

[0003] Firstly, traditional analytical methods, represented by spectral vegetation indices and statistical regression models, construct vegetation indices (such as NDRE) based on specific wavelength combinations or employ algorithms like partial least squares regression to establish empirical statistical relationships between spectral features and single target parameters. However, this approach has inherent limitations: relying on prior knowledge to select bands or design indices makes it difficult to fully extract effective information from hundreds of consecutive bands in hyperspectral data, and it cannot automatically handle spectral noise and multicollinearity issues; when simultaneously inverting multiple parameters, the spectral absorption characteristics of different parameters (such as water and protein) severely overlap, leading to feature utilization conflicts and contradictory estimation results when using the same spectral input for estimation of multiple independent statistical models, thus limiting overall accuracy; empirical models heavily depend on specific conditions of the training data (such as species and growth stage), and their performance significantly deteriorates and their universality is insufficient once the application scenario changes.

[0004] Secondly, deep learning models, represented by convolutional neural networks, can automatically learn deep abstract features from raw spectra, avoiding the limitations of manually designed features and demonstrating superiority in single-parameter inversion tasks such as chlorophyll content. However, they still face key obstacles when dealing with complex multi-parameter inversion tasks: existing research typically uses multiple independent single-task deep learning models to invert different parameters separately. This architecture ignores the close physiological and statistical correlations between various physicochemical parameters of leaves (such as chlorophyll and nitrogen, protein and dry matter), resulting in the models being unable to share learned general spectral features and failing to utilize the constraints between parameters to improve overall inversion consistency. Independent models need to be trained separately for each parameter task. In the context of scarce high-quality labeled data, this fragmented use of data fails to maximize the value of limited samples and is not conducive to learning more robust general feature representations from limited data. Specifically, there is the coexistence of strong overlap in multi-parameter spectral features and high physiological correlation between parameters. The independent model architecture cannot collaboratively utilize correlations to decouple overlapping features, leading to a bottleneck in accuracy improvement when inverting weak spectral signal parameters such as protein and dry matter.

[0005] In summary, current technologies, whether traditional statistical regression models or mainstream single-task independent deep learning models, have failed to effectively design a model architecture that simultaneously addresses the three coupled problems of hyperspectral feature utilization, multi-parameter correlation modeling, and generalization capability improvement under limited data. Specifically, how to design an end-to-end deep learning framework that can automatically learn general spectral features through shared network layers, capture parameter characteristics through specific task branches, and utilize correlation constraints between multiple tasks as prior knowledge to achieve high-precision and high-consistency collaborative inversion of multiple physicochemical parameters under limited data conditions remains an unresolved technical challenge. Summary of the Invention

[0006] To address the technical problem of low accuracy in predicting leaf physicochemical parameters, the present invention aims to provide a novel hyperspectral analysis method and system for leaf pigments and structural parameters. The specific technical solution adopted is as follows: One embodiment of the present invention provides a novel hyperspectral analysis method for leaf pigments and structural parameters, the method comprising the following steps: Acquire the hyperspectral reflectance data of the target leaf; input the hyperspectral reflectance data into a pre-trained multi-task deep learning model, and output the estimated values ​​of multiple physicochemical parameters of the target leaf; The multi-task deep learning model adopts a shared underlying network structure, which is a one-dimensional convolutional neural network. Its network structure includes the following sequentially connected components: The input layer is used to receive the hyperspectral reflectance data; The feature extraction module contains at least one dilated convolutional layer for extracting multi-scale spectral features from hyperspectral data, including shallow and deep features. The feature fusion module is used to fuse the shallow features and deep features output by the feature extraction module; Multiple parallel task output layers, each corresponding to a blade physicochemical parameter, are used to perform parameter-specific estimation and output based on the fused features.

[0007] Furthermore, the physicochemical parameters of the leaves include at least two of the following: chlorophyll a content, chlorophyll b content, carotenoid content, leaf nitrogen content, leaf dry matter content, and leaf thickness.

[0008] Furthermore, the hyperspectral reflectance data are bands selected from the complete spectrum, and the bands include at least one of the following sub-intervals: 430 nm to 900 nm, 960 nm to 1300 nm, 1500 nm to 1700 nm, and 2000 nm to 2400 nm.

[0009] Furthermore, the training steps of the multi-task deep learning model include: Acquire a dataset containing hyperspectral reflectance data of leaves from various plants and corresponding measured physicochemical parameter values ​​of the leaves; The hyperspectral reflectance data containing various plant leaves were subjected to band filtering and normalization preprocessing. Using the filtered and preprocessed hyperspectral reflectance data as input and the corresponding measured physicochemical parameters of the blades as supervision labels, the one-dimensional convolutional neural network is jointly trained on multiple tasks until the model converges.

[0010] Furthermore, the multi-task deep learning model is implemented based on the Python TensorFlow framework, and the training and validation process is performed 50 times in total; during model training, the number of training cycles is set to 250, the batch size is set to 32, and the learning rate is set to the real-time callback learning rate.

[0011] Furthermore, the expression for updating the learning rate is: In the formula, This represents the current learning rate. Indicates the real-time learning rate. This indicates the number of training epochs in which the performance metric has not improved consecutively when monitoring validation set performance.

[0012] Furthermore, the accuracy evaluation metrics of the multi-task deep learning model include at least the coefficient of determination, root mean square error, and normalized root mean square error.

[0013] Furthermore, the multi-task deep learning model is applied to drones or satellite platforms.

[0014] Another embodiment of the present invention provides a novel hyperspectral analysis system for leaf pigments and structural parameters, comprising: The data acquisition module is used to acquire the hyperspectral reflectance data of the target leaf. The processor is configured to run a pre-trained multi-task deep learning model, which adopts a shared underlying network structure and is configured to receive the hyperspectral reflectance data and output estimated values ​​of multiple leaf physicochemical parameters. The result output module is used to output the estimated values ​​of the multiple blade physicochemical parameters; The multi-task deep learning model is a one-dimensional convolutional neural network, and its network structure includes the following sequentially connected components: The input layer is used to receive the hyperspectral reflectance data; The feature extraction module contains at least one dilated convolutional layer for extracting multi-scale spectral features from hyperspectral data, including shallow and deep features. The feature fusion module is used to fuse the shallow features and deep features output by the feature extraction module; Multiple parallel task output layers, each corresponding to a blade physicochemical parameter, are used to perform parameter-specific estimation and output based on the fused features.

[0015] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a novel hyperspectral analysis method for leaf pigments and structural parameters.

[0016] The present invention has the following beneficial effects: This invention provides a novel hyperspectral analysis method and system for leaf pigment and structural parameters. Addressing the shortcomings of traditional methods that rely on manual features, the input layer of the model in this invention directly receives the raw, continuous hyperspectral reflectance sequence. This preserves complete spectral information across the entire band without manual selection, providing a data foundation for subsequent modules to automatically mine deep and nonlinear features, avoiding information loss caused by manually designing vegetation indices or selecting bands. For complex and multi-scale spectral information, this invention utilizes one-dimensional convolution to operate along the spectral dimension. Dilated convolution expands the receptive field by controlling the dilation rate, thereby simultaneously extracting local details (shallow features) and global context (deep features). It can automatically and efficiently learn multi-scale spectral representations from the raw spectrum, ranging from subtle absorption valleys to broad band trends, providing a rich and robust feature foundation for subsequent multi-task estimation. For different parameters that depend on different scale features, the shallow (detail) features output by the feature extraction module are fused with deep (semantic) features (e.g., splicing or weighting). This facilitates the generation of a unified feature representation that integrates multi-scale information, enabling the model to simultaneously utilize features sensitive to certain parameters. The integration of subtle local variations and global patterns important to the overall trend enhances the comprehensive ability of feature representation and helps decouple overlapping spectral signals. The input layer, feature extraction module, and feature fusion module are designed to be shared by all prediction tasks, forming the core architecture of multi-task learning. This forces the network to learn a universal spectral feature useful for the inversion of all physicochemical parameters. Essentially, this utilizes a large amount of task data to discover and encode the intrinsic physiological and statistical correlations between parameters. Furthermore, the shared layer learns from all task samples, which amplifies the utility of limited training data and plays a powerful regularization role, enabling the model to learn a more fundamental and robust spectral-physical mapping relationship, significantly improving generalization performance on small datasets. Above the shared universal high-level features, an independent lightweight output network is set up for each physicochemical parameter. This allows each task layer to focus on learning the unique mapping from the universal feature to its own target parameter, thereby accurately capturing the specific pattern of that parameter. The model outputs the estimated values ​​of all target parameters simultaneously through a forward propagation path, ensuring that the estimation process is based on the same set of feature understandings, thus ensuring physical consistency and computational efficiency between results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a flowchart illustrating the steps of a novel hyperspectral analysis method for leaf pigments and structural parameters, as described in one embodiment of the present invention. Figure 2 This is a schematic diagram of the architecture of a multi-task deep learning model in an embodiment of the present invention; Figure 3 The results of the correlation analysis between spectral reflectance and leaf biochemical parameters in the embodiments of the present invention are shown. Figure 4 The correlation coefficients between various leaf biochemical parameters and vegetation indices in the embodiments of the present invention; Figure 5 The blade parameter estimation results based on RF in this embodiment of the invention; Figure 6 The training results of blade parameters based on the MLTNet model in this embodiment of the invention; Figure 7 The blade parameter estimation results based on MLTNet in the embodiments of the present invention; Figure 8 Chl-a and Chl-b are estimated by different models in the embodiments of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solution proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] One embodiment of the present invention provides a novel hyperspectral analysis method for leaf pigments and structural parameters. This method aims to fully explore and utilize the inherent physiological and physical autocorrelation between multiple physicochemical parameters of leaves through a deep learning model, so as to achieve high-precision, integrated estimation of these parameters.

[0022] Specifically, in this embodiment, a multi-task deep learning model called MLTNet is constructed. The model uses hyperspectral reflectance data of plant leaves as input, and its core architecture employs a one-dimensional dilated convolutional network to automatically extract multi-scale deep features from the raw spectrum. The model learns common spectral representations of multiple parameters through a shared convolutional layer structure, thereby implicitly capturing and utilizing the dependencies between parameters. Simultaneously, a feature fusion module is designed to combine the network's shallow detail features with deep abstract features. Finally, through parallel output layers dedicated to multiple tasks, the synchronous and collaborative estimation of the aforementioned various leaf physicochemical parameters is achieved. Figure 1 As shown, it includes the following steps: S101, acquire hyperspectral reflectance data of the target leaf.

[0023] Here, the target leaf is the leaf whose pigmentation and structure are to be analyzed, and the target leaf should be a healthy, typical functional leaf without visible lesions or dust. It is worth noting that when measuring and analyzing, care should be taken to avoid the midrib of the leaf and select a uniform area of ​​the leaf surface.

[0024] In this embodiment, a handheld or mounted hyperspectral radiometer is used to perform vertical or near-vertical measurements on the target leaf at a specific distance (e.g., 10 cm to 1 m) under natural light or standard artificial light source conditions to obtain initial spectral data.

[0025] Alternatively, a hyperspectral sensor equipped with a leaf clip or integrating sphere can be used to place the leaf sample at a specific location for contact measurement to obtain a more stable reflectance spectrum that is less affected by ambient light.

[0026] Alternatively, a hyperspectral imager can be mounted on a ground-based mobile platform, an elevated track, or a UAV (unmanned aerial vehicle) or other near-ground remote sensing platform to image crop canopies or individual leaves, and the average reflectance spectrum or pixel-level spectrum of the target leaf can be extracted from the image.

[0027] Before inputting data into the model, necessary preprocessing and quality control steps are typically included to ensure data consistency and reliability, including: First, the original DN values ​​or radiance values ​​are calibrated for radiometrics and converted for reflectance. For non-contact measurements, a standard white board is used for reference calibration. Second, methods such as Savitzky-Golay convolution smoothing, moving average, or wavelet transform are applied to reduce random noise in the spectral signal. Based on spectral shape, reflectance range, or statistical methods (such as standard deviation threshold), abnormal spectral curves caused by measurement errors, leaf anomalies, or mixed pixels are identified and eliminated. Finally, depending on sensor characteristics and model input requirements, specific spectral ranges may be truncated (e.g., 430 nm to 900 nm, 960 nm to 1300 nm, 1500 nm to 1700 nm, 2000 nm to 2400 nm), and uniformly resampled to a fixed number of bands and wavelength intervals to ensure consistency with the training data format. That is, the final input data to the model should be a one-dimensional vector reflectance sequence after the above processing, and its length (number of bands) should strictly match the dimension of the model input layer.

[0028] It should be noted that the input data for the model is selected hyperspectral reflectance, and the band settings are as follows: First, the input bands should include the sensitive spectra of Chl-a, Chl-b, Car, LNC, LCC, and LT parameters to improve the model's parameter estimation performance. Second, the input bands should avoid band redundancy caused by instrument spectral resampling to reduce model parameters and computational load. Third, the input bands should avoid water vapor absorption bands as much as possible, which are classic hyperspectral satellite band locations (e.g., the Gaofen-5 AHSI sensor). Furthermore, the normalized output bands of the MLTNet model should be the Chl-a, Chl-b, Car, LNC, LCC, and LT parameters. Normalization smooths out the range of different output results, preventing the model from losing weight differences due to MSE error for different parameters.

[0029] After preprocessing the initial spectral reflectance data, it becomes hyperspectral reflectance data, which can be directly used as input for subsequent multi-task deep learning models. It is worth noting that the data acquisition and preprocessing standards in step S101 must be consistent with the data specifications used in the model training phase. This is a crucial prerequisite for ensuring the model maintains estimation accuracy in practical applications.

[0030] S102 inputs the hyperspectral reflectance data into a pre-trained multi-task deep learning model and outputs estimated values ​​of multiple physicochemical parameters of the target leaf.

[0031] Here, the physicochemical parameters of the leaves include at least two of the following: chlorophyll a content, chlorophyll b content, carotenoid content, leaf nitrogen content, leaf dry matter content, and leaf thickness.

[0032] The MLTNet model (multi-task deep learning model) constructed in this embodiment can be used for potential applications on UAV or satellite platforms. A schematic diagram of the architecture of the multi-task deep learning model is shown below. Figure 2 As shown, its core architecture is based on a one-dimensional convolutional neural network (1D CNN) and dilated convolution technology, aiming to fully mine the absorption features at different scales in leaf hyperspectral data through multi-level feature extraction and fusion. The model's network structure includes the following sequentially connected components: The input layer is used to receive hyperspectral reflectance data.

[0033] In this embodiment, the input layer receives model input data and connects to the subsequent feature extraction module. Before being input into the model, the acquired hyperspectral reflectance data, i.e., one-dimensional reflectance spectrum data, is reorganized or padded according to the dimensions defined by the model input layer (such as the number of bands and their order), typically converting it to (1, ...). The tensor form of ), where The number of input bands preset for the model.

[0034] The feature extraction module contains at least one dilated convolutional layer for extracting multi-scale spectral features from hyperspectral data, including shallow and deep features.

[0035] In this embodiment, the main feature extraction part of the model consists of four sequentially connected basic convolutional blocks. Each basic convolutional block has the same sub-layer structure, including: Continuous 1D CNN convolutional layers: used for local feature extraction; max pooling layers: used for downsampling, increasing the degree of feature abstraction and receptive field.

[0036] It should be noted that, in order to capture contextual information over a wider spectral range without significantly increasing the number of parameters, the dilation rate used in these four basic convolutional blocks is set to different values ​​that increase sequentially. Through this design, the receptive field of the convolutional operation expands exponentially as the network deepens, enabling the model to simultaneously learn multiple scales of spectral patterns, from local fine features to global broad-band trends, thereby better adapting to the differences in absorption characteristics dependent on different physicochemical parameters.

[0037] The feature fusion module is used to fuse the shallow and deep features output by the feature extraction module.

[0038] In this embodiment, to comprehensively utilize the detailed information of shallow networks and the abstract semantic information of deep networks, a feature fusion module is set up after the four basic convolutional blocks. This feature fusion module can concatenate or weightedly fuse feature maps from different depths (e.g., from shallower and deeper blocks) to form a unified feature representation containing multi-scale information, providing richer input for subsequent parameter estimation.

[0039] Multiple parallel task output layers, each corresponding to a blade physicochemical parameter, are used to perform parameter-specific estimation and output based on the fused features.

[0040] In this embodiment, the fused high-level features are first regularized through a Dropout layer to reduce the risk of overfitting. Subsequently, the features are fed into three consecutive fully connected layers for nonlinear transformation and dimension mapping. Finally, the network generates estimated values ​​for six physicochemical parameters of the target leaf (chlorophyll a, chlorophyll b, carotenoids, protein, dry matter content, and leaf thickness) through an output layer with six neurons.

[0041] In summary, the MLTNet in this embodiment adopts an end-to-end architecture consisting of a multi-scale feature extractor composed of stacked convolutional blocks with different porosities, followed by a feature fusion module and a multi-layer fully connected regressor, achieving integrated high-precision estimation from raw hyperspectral data to multiple leaf physicochemical parameters.

[0042] As an exemplary implementation, the training steps of a multi-task deep learning model include: First, a dataset containing hyperspectral reflectance data of leaves from various plants and corresponding measured physicochemical parameter values ​​of the leaves was obtained.

[0043] In this embodiment, the hyperspectral reflectance data primarily comes from two datasets: AQGRAD and LOPEX. AQGRAD, released in 2024, is a dataset of leaf optical properties used for calibrating leaf radiative transfer models. It focuses on expanding the correlation between leaf spectral measurements and biochemical / biophysical properties to improve leaf optical models. This dataset covers leaves from 53 tree and crop species. The AQGRAD spectral dataset uses a Cary 5000 spectrometer to measure leaf spectra, ranging from 400 nm to 2500 nm, and includes biochemical (e.g., chlorophyll, protein) and biophysical (e.g., leaf thickness) measurements.

[0044] LOPEX is an experimental dataset developed in 1993 by the European Union Joint Research Centre. It aims to correlate the visible or infrared spectra of plant leaves with physical measurements and biochemical analyses. The dataset includes reflectance and transmittance measurements of leaf samples from 45 different plant species, with a spectral range from 400 nm to 2500 nm. The LOPEX dataset also includes metadata such as leaf thickness, chlorophyll, and protein biochemical parameters. In total, the dataset provides approximately 606 sets of leaf reflectance and transmittance spectra, covering both coniferous and broadleaf plants, and is used for the calibration and validation of leaf optical models, such as the PROSPECT model.

[0045] It should be noted that some samples in the AQGRAD and LOPEX datasets had missing leaf parameter data and were therefore not included in the analysis. Similarly, samples containing missing measurements were not included in the model analysis in this embodiment. A total of 476 complete sample datasets were used for model construction and validation in this embodiment; of these, 238 datasets were used to train the model, and the remaining 238 datasets were used to validate the model.

[0046] In this embodiment, the supervised label data required for model training consists of measured values ​​of plant leaf physicochemical parameters acquired synchronously with the hyperspectral data. Specifically, these include chlorophyll a content, chlorophyll b content, carotenoid content, protein content, dry matter content, and leaf thickness. All label data were obtained using standard laboratory measurement methods; for example, chlorophyll content was measured using a spectrophotometer, and dry matter content was measured using a drying and weighing method. These measured values ​​constitute the direct target of model learning.

[0047] Secondly, band filtering and normalization preprocessing were performed on the hyperspectral reflectance data containing various plant leaves.

[0048] In this embodiment, the band selection and normalization preprocessing of the model training data are consistent with the process of obtaining hyperspectral reflectance data in step S101 above, and will not be repeated here.

[0049] In another embodiment, the model input can consist of two parts to fully utilize prior knowledge and enhance the model's feature representation: the original hyperspectral reflectance sequence, which serves as the basic input and retains complete spectral information; and a preferred spectral feature set, which is used to further assist the model. In this embodiment, 28 features sensitive to the target parameters are selected from widely used vegetation indices and feature bands to form a supplementary input feature set, which is detailed in Table 1. Table 1 The vegetation indices and characteristic bands in Table 1 include indices sensitive to chlorophyll a, chlorophyll b, and leaf nitrogen content, such as the normalized red edge index and chlorophyll index; photochemical reflectance indices sensitive to carotenoids; features sensitive to lignin and cellulose within the leaves; typical absorption peak bands of chlorophyll a and b; and the normalized water index sensitive to leaf moisture. It should be noted that these selected features are concatenated with the original spectral data before being input into the model, serving as the model's input vector to provide multi-faceted feature representations with clear physiological and physical significance.

[0050] To lay the physical foundation of this invention and to clarify the necessity of designing a multi-task deep learning model, a systematic correlation analysis was conducted on the training dataset: First, analyze the complex and differentiated correlation patterns between hyperspectral data and various target parameters: First, the parameters of photosynthetic pigments (such as chlorophyll a, chlorophyll b, and carotenoids) show a significant negative correlation trough in the visible light band (400-700 nm), the position of which precisely matches the absorption peaks of known pigments (e.g., chlorophyll a at ~660 nm, chlorophyll b at ~640 nm). Figure 3 As shown, this confirms the direct physical basis for using spectral inversion to determine pigment content.

[0051] Secondly, structural and non-pigmentary biochemical parameters (such as leaf nitrogen content LNC, carbon content LCC, and leaf thickness LT): their correlation with the spectrum is more complex, such as... Figure 4 As shown, in the near-infrared (NIR) region, LT exhibits a strong correlation due to the influence of internal scattering within the leaf; while in the short-wave infrared (SWIR) region, the strong absorption of water dominates the spectral signal, severely masking the weak characteristics of parameters such as LNC and LCC, resulting in generally low linear correlation between them and the broadband spectrum.

[0052] The above analysis shows that the spectral response signals of the target parameters have strong overlap (such as water and LNC / LCC) and uneven intensity (strong pigment signal, weak protein / carbon signal). This directly explains from the data level why traditional spectral indices and statistical models that rely on linear or simple nonlinear relationships are difficult to achieve high-precision synchronous inversion of multiple parameters.

[0053] Second, the autocorrelation relationship between the target parameters: Significant statistical autocorrelation exists among the various leaf physicochemical parameters, as shown in Table 2. For example, chlorophyll a and b show a high positive correlation with carotenoids (correlation coefficient reaching as high as 0.873), reflecting their close synergy in physiological function. Meanwhile, some parameters also exhibit negative or weak correlations.

[0054] Table 2 The above analysis shows that multiple prediction tasks are not independent of each other, but rather have an inherent physiological network of connections. This provides direct data support for adopting a multi-task learning architecture—by learning the common features of these parameters through a shared underlying network, it is expected to achieve more consistent and efficient performance estimation than a single-task model that models them independently.

[0055] Third, verifying the characterization ability of the selected vegetation index: To further incorporate prior knowledge, this embodiment selects 28 vegetation indices and characteristic bands that are sensitive to the target parameters. The following... Figure 5 Correlation analysis validated the effectiveness of the feature set: for example, chlorophyll index (LCI) was most correlated with pigment parameters, normalized water index (NDII1) was highly correlated with leaf thickness (LT), and cellulose uptake index (CAI) was correlated with LCC.

[0056] The above analysis shows that these preferred features constitute a powerful complement to the original hyperspectral data, each enhancing the sensitivity to specific parameters. In the implementation of this invention, this feature set, along with the original spectral data, is used as input to the model to provide more discriminative information.

[0057] Fourth, analytical conclusions and implications for this invention: Hyperspectral data does contain rich information for retrieving various target parameters, but the information overlaps and the signal-to-noise ratio is uneven. The inherent correlation between parameters provides an opportunity to adopt a collaborative learning paradigm. The complexity of spectral features requires the model to have strong nonlinear feature mining and decoupling capabilities.

[0058] The aforementioned correlation analysis directly confirms and supports the core design of this invention: an end-to-end model must be employed that can automatically learn hierarchical, multi-scale features, implicitly utilize task correlations through parameter sharing, and is compatible with both the original spectrum and prior features. Therefore, the MLTNet multi-task deep learning model described below is a necessary and optimized technical solution designed specifically for the aforementioned data characteristics.

[0059] It should be noted that in the subsequent training process, the model uses the above-mentioned fused feature vector as input and the measured physicochemical parameter values ​​in the laboratory as supervision labels. The network weights are optimized by minimizing the error between the predicted value and the true label. The above-mentioned vegetation index and feature band are clues or hints to help the model learn, rather than the learning target itself.

[0060] Finally, using the filtered and preprocessed hyperspectral reflectance data as input and the corresponding measured physicochemical parameters of the blades as supervision labels, a one-dimensional convolutional neural network is jointly trained on multiple tasks until the model converges.

[0061] In this embodiment, the multi-task deep learning model is implemented using the Python TensorFlow framework. The training and validation process is repeated 50 times, and the final estimated model is selected. During model training, the number of training epochs is set to 250, and the batch size is set to 32. The learning rate is set to a real-time callback learning rate; if the loss does not improve after 10 epochs, the new learning rate can be 0.3 times the current learning rate. The expression for the dynamic learning rate callback strategy is: ; In the formula, This represents the current learning rate. Indicates the real-time learning rate. This represents the number of training epochs in which the performance metric has not improved consecutively while monitoring the validation set performance. In this embodiment, the initial learning rate is 0.001, and the minimum learning rate is 0.00001.

[0062] It's important to note that the loss refers to the loss value calculated on the validation set, independent of the training set. Using the validation set loss as a monitoring metric is to accurately reflect the model's generalization ability to unseen data, thus preventing incorrect judgments due to the model overfitting only to the training set.

[0063] The aforementioned improvement can be defined as follows: the validation set loss value calculated in the current epoch is strictly less than the best validation set loss value in the historical records; conversely, no improvement means that the current validation set loss value is greater than or equal to the historical best value. Therefore, the above loss not being improved means that in N consecutive training epochs (e.g., 10), the loss function value calculated by the model on the independent validation set has not fallen below at least a preset small threshold of the previously recorded best loss value.

[0064] In practical applications, to avoid frequent callbacks due to numerical fluctuations (such as extremely small decreases), a small threshold is usually introduced, denoted as . The judgment logic can then be adjusted to: the current loss is less than the optimal loss and... The difference between the current epoch and the optimal loss indicates that the current epoch has improved; while when the loss is greater than or equal to the optimal loss and the optimal loss, the current epoch has improved. The difference indicates that the current epoch has not improved. To implement learning rate judgment, a non-improvement counter can be set, i.e. This counter increments each time the loss on the validation set fails to improve, and resets to zero once it does. Only when this counter reaches a preset patience value (e.g., 10) does the system finally determine that the loss has not improved after 10 epochs and trigger a learning rate adjustment.

[0065] For evaluating the accuracy of the model, the accuracy evaluation metrics for the multi-task deep learning model in this embodiment include at least the coefficient of determination. ), Root Mean Square Error (RMSE), and Normalized Root Mean Square Error (nRMSEs).

[0066] ; ; ; In the formula, The coefficient of determination is denoted by m, where m represents the total number of samples evaluated. This represents the predicted value of the i-th sample. This represents the true value of the i-th sample. This represents the average of the true values ​​of all evaluated samples; Indicates the root mean square error; This represents the sum of the normalized root mean square errors of all the physicochemical parameters of the blade to be estimated, where t represents the number of physicochemical parameters of the blade to be estimated. This represents the normalized root mean square error of the physicochemical parameters of the i-th leaf.

[0067] In summary, this embodiment analyzes whether deep learning technology can fully utilize the autocorrelation of leaf pigment and structural parameters. To achieve high-precision estimation, an MLTNet model is proposed. This model takes leaf hyperspectral reflectance as input and simultaneously estimates multiple parameters. Specifically, the model extracts hyperspectral data features through 1D dilated convolutions and captures the dependencies between parameters through shared convolutional layers. It fuses features from deep and shallow networks and uses multi-head output to simultaneously estimate chlorophyll a content, chlorophyll b content, carotenoid content, protein content, dry matter content, and leaf thickness.

[0068] To objectively evaluate the performance of the MLTNet model proposed in this invention, three widely used hyperspectral estimation models for leaf physicochemical parameters were simultaneously established on the same dataset as performance comparison benchmarks: Multiple linear regression (LR): As a classic statistical method, LR establishes the relationship between spectral features (independent variables) and target parameters (dependent variables) through linear fitting, and is used to measure the strength of the most basic linear association between parameters and spectra.

[0069] Partial Least Squares Regression (PLSR): This method combines the features of principal component analysis and multiple regression, building a model by extracting the latent variable with the largest covariance in the independent and dependent variable spaces. PLSR can effectively handle the high degree of multicollinearity between bands in hyperspectral data and is a commonly used robust statistical method in vegetation parameter inversion.

[0070] Random Forest Regression (RF): As an ensemble learning algorithm, RF performs regression by constructing a large number of decision trees and integrating their predictions. This model has good fitting ability and noise resistance for nonlinear relationships and is suitable for modeling complex, high-dimensional spectral-parameter relationships.

[0071] In this embodiment, all the benchmark models mentioned above are implemented based on the scikit-learn library in Python. It should be noted that since LR, PLSR, and RF are essentially single-task regression models, in order to achieve multi-parameter estimation, a corresponding model needs to be trained independently for each target physicochemical parameter (chlorophyll a, chlorophyll b, etc.). This independent modeling approach is in direct contrast to the single-model multi-task collaborative learning architecture adopted by MLTNet in this invention.

[0072] The estimation accuracy of the physicochemical parameters of each blade based on the PLSR, LR and RF models is shown in Table 3: Table 3 Table 3 shows that the three methods achieve similar accuracy in estimating blade parameters. Using nRMSEs as the metric, LR achieves the best Chl-a estimation result; PLSR achieves the best Car estimation result; and RF achieves the best Chl-b, LCC, and LT estimation results. The estimation results for different types of blade parameters based on RF are as follows: Figure 5 As shown in the figure. Furthermore, this embodiment calculated the nRMSEs of all benchmark models and the MLTNet model of this invention. The results show that the nRMSEs of the three comparative models can be ranked as RF > PLSR > LR. In Table 3, * indicates the traditional statistical regression model that achieves the best estimation result, and ** indicates the model that achieves the best estimation result.

[0073] The training results of blade parameters based on the MLTNet model are as follows: Figure 6 As shown, by Figure 6 It can be seen that the MLTNet model's accuracy increases rapidly during the first 50 epochs of training, but remains almost unchanged after approximately 200 epochs. The blade parameter estimation results based on MLTNet are as follows: Figure 7 As shown in Table 3, using RMSE as the metric, MLTNet achieved the best estimation results for all parameters. This embodiment further calculated the nRMSEs of all benchmark models and the MLTNet model of this invention. The results show that the nRMSEs of the four models can be ranked as MLTNet > RF > PLSR > LR.

[0074] Chl-a and Chl-b exhibit similar spectral absorption characteristics, and their ratios tend to be stable within the same growth stage. However, significant differences exist between them at different growth stages or in different vegetation types. The estimation performance of Chl-a and Chl-b can test the model's ability to capture the high autocorrelation between leaf physicochemical parameters. Different models estimate Chl-a and Chl-b as follows: Figure 8 As shown, the results indicate that the MLTNet model exhibits the strongest Chl-a / Chl-b estimation results, with RMSE=0.448. The Chl-a / Chl-b estimation results of the four tested models can be ranked as MLTNet > RF > PLSR > LR, which is the same as the ranking of the plant leaf parameter estimation results.

[0075] The spectral absorption characteristics and high autocorrelation of leaf pigments: Leaf pigment parameters (such as Chl-a, Chl-b, and Car) exhibit high autocorrelation in plant physiological processes, stemming from their synergistic effects in photosynthesis. Chl-a and Chl-b, as major photosynthetic pigments, are responsible for capturing light energy and driving the electron transport chain, while Car, as an accessory pigment, protects chlorophyll from photo-oxidative damage and expands the light absorption range. Spectrally, the absorption peaks of leaf pigments in the visible light band (approximately 400-700 nm) highly overlap. For example, Chl-a has strong absorption peaks near 430 nm and 660 nm, Chl-b near 460 nm and 640 nm, while Car mainly absorbs in the 400-500 nm blue light region. This overlap may lead to a negative correlation of spectral reflectance in the corresponding bands, and their band correlation differs from their spectral absorption characteristics. This high autocorrelation and similarity in spectral absorption characteristics pose challenges to traditional vegetation indices and statistical regression methods. Vegetation indices (such as NDRE and LCI) are typically designed based on ratios or differences in specific wavelength bands to capture absorption characteristics such as chlorophyll. However, due to the similar spectral responses of Chl-a, Chl-b, and Car, these indices often fail to distinguish subtle differences between them. For example, NDRE is sensitive to total chlorophyll but responds weakly to the Chl-a / Chl-b ratio because this ratio is affected by environmental stresses (such as nitrogen deficiency or light intensity), and the index design does not adequately account for this dynamic change. The analysis of the above RF-based leaf parameter estimation results confirms this, namely that most VIs are positively correlated with pigment parameters, but the correlation coefficients are similar (e.g., LCI has a correlation of 0.61 with Chl-a, 0.53 with Chl-b, and 0.47 with Car), indicating that VIs are difficult to separate with high precision across multiple parameters.

[0076] Therefore, in practical applications, this high autocorrelation can lead to estimation biases for Chl-a, Chl-b, and Car. For example, in the early stages of crop growth, the Chl-a / Chl-b ratio is low (typically 2-3:1 for C3 plants), reflecting optimized photosynthesis; while in later stages of growth, chlorophyll b decomposes rapidly, and this ratio may increase (typically 5-6:1 for C3 plants). This is particularly evident in multi-species datasets such as AQGRAD and LOPEX, which cover 53 tree species and 45 plant species, where species diversity amplifies the effects of autocorrelation. If vegetation indices are used in conjunction with statistical regression (such as PLSR or RF), the model tends to capture the total pigment signal while ignoring the variability of individual parameters.

[0077] To overcome the aforementioned shortcomings, the advantages of deep learning in vegetation parameter analysis are utilized. Hyperspectral remote sensing data are characterized by high dimensionality (hundreds to thousands of bands), redundancy (high correlation between adjacent bands), and nonlinearity (nonlinear relationship between parameters and reflectance). Traditional methods, such as vegetation indices combined with statistical regression, struggle to address these challenges. Traditional vegetation indices, such as NDVI or OSAVI, are designed with linear or simple ratios and cannot handle the complex interactions between parameters. For example, chlorophyll content is highly correlated with nitrogen (LNC) because nitrogen is a key element in chlorophyll synthesis; similarly, dry matter content (LCC) is correlated with leaf thickness (LT), reflecting the synergy of structural parameters.

[0078] The MLTNet model proposed in this invention is based on 1D CNN, fusing shallow and deep features and using dilated convolutions to expand the receptive field to handle the complexity of hyperspectral data. MLTNet avoids water vapor absorption by selecting sensitive bands (430-900 nm, 960-1300 nm, etc.) and uses max-pooling layers for noise reduction. Dilated convolutions further expand context awareness, helping the model capture subtle changes such as red-edge shifts. The advantages of deep learning in capturing nonlinear relationships are evident. MLTNet uses shared convolutional layers to simultaneously estimate six parameters (Chl-a, Chl-b, Car, LNC, LCC, LT), fully utilizing autocorrelation. Through multi-head output, the model learns the dependencies between parameters, improving overall performance. Results show that MLTNet's nRMSEs are 1.125, outperforming RF (1.423), PLSR (1.462), and LR (1.513), demonstrating its generalization ability in processing multi-species datasets (AQGRAD and LOPEX, covering more than 90 plant species). Overall, MLTNet demonstrates the advantages of deep learning in capturing complexity, improving accuracy, and facilitating joint estimation in vegetation parameter analysis, providing efficient technical support for vegetation leaf parameter analysis.

[0079] Furthermore, traditional vegetation index combined with statistical regression methods requires separate modeling for each parameter, while MLTNet's joint estimation reduces computational redundancy and utilizes transfer learning between parameters. For example, Chl-a features can assist in Car estimation because they share visible light absorption characteristics.

[0080] This invention provides a novel hyperspectral analysis method for leaf pigment and structural parameters. This method achieves the strongest leaf parameter estimation performance. Specifically, the leaf parameter estimation results of MLTNet and the widely used vegetation index combined with classical regression method were compared, showing that the accuracy ranking is MLTNet (nRMSEs=1.125) > RF (nRMSEs=1.423) > PLSR (nRMSEs=1.462) > LR (nRMSEs=1.513). This method is better able to capture the high autocorrelation between leaf physicochemical parameters than the widely used vegetation index combined with classical regression method. Specifically, the estimation performance of different models for Chl-a / Chl-b was tested, showing that the accuracy ranking is MLTNet (RMSE=0.448) > RF (RMSE=0.500) > PLSR (RMSE=0.503) > LR (RMSE=0.512). Therefore, the MLTNet model in this invention has the ability to capture variables with similar spectral absorption characteristics and high autocorrelation. Specifically, the deep learning model has advantages in capturing complex interactions. By sharing convolutional layers, multiple parameters can be estimated simultaneously to make full use of these autocorrelations and improve overall performance.

[0081] Based on the same inventive concept, another embodiment of the present invention provides a novel hyperspectral analysis system for leaf pigments and structural parameters, comprising: The data acquisition module is used to acquire the hyperspectral reflectance data of the target leaf. The processor is configured to run a pre-trained multi-task deep learning model, which adopts a shared underlying network structure and is configured to receive the hyperspectral reflectance data and output estimated values ​​of multiple leaf physicochemical parameters. The result output module is used to output the estimated values ​​of the multiple blade physicochemical parameters; The multi-task deep learning model is a one-dimensional convolutional neural network, and its network structure includes the following sequentially connected components: The input layer is used to receive the hyperspectral reflectance data; The feature extraction module contains at least one dilated convolutional layer for extracting multi-scale spectral features from hyperspectral data, including shallow and deep features. The feature fusion module is used to fuse the shallow features and deep features output by the feature extraction module; Multiple parallel task output layers, each corresponding to a blade physicochemical parameter, are used to perform parameter-specific estimation and output based on the fused features.

[0082] Since this system embodiment is based on the aforementioned method embodiment, its operations and core model architecture are exactly the same as those of the method. Therefore, this system can naturally inherit and realize all the beneficial effects of the method. For example, through multi-task shared learning and multi-scale feature fusion, it can achieve high-precision, high-consistency collaborative inversion of multiple blade physicochemical parameters and has good generalization ability. Further details will not be elaborated here.

[0083] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a novel hyperspectral analysis method for leaf pigments and structural parameters.

[0084] Since the computer program stored in the storage medium is used to execute all the steps of the aforementioned method embodiments, by running the program, the corresponding computing device or system can implement the method and will necessarily obtain all the beneficial effects consistent with the method, including but not limited to improving the accuracy and consistency of multi-parameter estimation, making full use of spectral information and parameter correlation, and achieving better generalization performance with limited data.

[0085] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A novel hyperspectral analysis method for leaf pigments and structural parameters, characterized in that, Includes the following steps: Acquire hyperspectral reflectance data of the target leaf; The hyperspectral reflectance data is input into a pre-trained multi-task deep learning model, which outputs estimated values ​​of multiple physicochemical parameters of the target leaf. The multi-task deep learning model adopts a shared underlying network structure, which is a one-dimensional convolutional neural network. Its network structure includes the following sequentially connected components: The input layer is used to receive the hyperspectral reflectance data; The feature extraction module contains at least one dilated convolutional layer for extracting multi-scale spectral features from hyperspectral data, including shallow and deep features. The feature fusion module is used to fuse the shallow features and deep features output by the feature extraction module; Multiple parallel task output layers, each corresponding to a blade physicochemical parameter, are used to perform parameter-specific estimation and output based on the fused features.

2. The novel hyperspectral analysis method for leaf pigments and structural parameters according to claim 1, characterized in that, The physicochemical parameters of the leaves include at least two of the following: chlorophyll a content, chlorophyll b content, carotenoid content, leaf nitrogen content, leaf dry matter content, and leaf thickness.

3. The novel hyperspectral analysis method for leaf pigments and structural parameters according to claim 1, characterized in that, The hyperspectral reflectance data are bands selected from the complete spectrum, and the bands include at least one of the following sub-intervals: 430 nm to 900 nm, 960 nm to 1300 nm, 1500 nm to 1700 nm, and 2000 nm to 2400 nm.

4. The novel hyperspectral analysis method for leaf pigments and structural parameters according to claim 1, characterized in that, The training steps of the multi-task deep learning model include: Acquire a dataset containing hyperspectral reflectance data of leaves from various plants and corresponding measured physicochemical parameter values ​​of the leaves; The hyperspectral reflectance data containing various plant leaves were subjected to band filtering and normalization preprocessing. Using the filtered and preprocessed hyperspectral reflectance data as input and the corresponding measured physicochemical parameters of the blades as supervision labels, the one-dimensional convolutional neural network is jointly trained on multiple tasks until the model converges.

5. The novel hyperspectral analysis method for leaf pigments and structural parameters according to claim 4, characterized in that, The multi-task deep learning model is implemented based on the Python TensorFlow framework, and the training and validation process is performed 50 times in total. During model training, the number of training cycles is set to 250, the batch size is set to 32, and the learning rate is set to the real-time callback learning rate.

6. The novel hyperspectral analysis method for leaf pigments and structural parameters according to claim 5, characterized in that, The expression for the dynamic learning rate callback strategy is: In the formula, This represents the current learning rate. Indicates the real-time learning rate. This indicates the number of training epochs in which the performance metric has not improved consecutively when monitoring validation set performance.

7. The novel hyperspectral analysis method for leaf pigments and structural parameters according to claim 1, characterized in that, The accuracy evaluation metrics for the multi-task deep learning model include at least the coefficient of determination, root mean square error, and normalized root mean square error.

8. The novel hyperspectral analysis method for leaf pigments and structural parameters according to claim 1, characterized in that, The multi-task deep learning model is applied to drones or satellite platforms.

9. A novel hyperspectral analysis system for leaf pigments and structural parameters, characterized in that, include: The data acquisition module is used to acquire the hyperspectral reflectance data of the target leaf. The processor is configured to run a pre-trained multi-task deep learning model, which adopts a shared underlying network structure and is configured to receive the hyperspectral reflectance data and output estimated values ​​of multiple leaf physicochemical parameters. The result output module is used to output the estimated values ​​of the multiple blade physicochemical parameters; The multi-task deep learning model is a one-dimensional convolutional neural network, and its network structure includes the following sequentially connected components: The input layer is used to receive the hyperspectral reflectance data; The feature extraction module contains at least one dilated convolutional layer for extracting multi-scale spectral features from hyperspectral data, including shallow and deep features. The feature fusion module is used to fuse the shallow features and deep features output by the feature extraction module; Multiple parallel task output layers, each corresponding to a blade physicochemical parameter, are used to perform parameter-specific estimation and output based on the fused features.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of a novel hyperspectral analysis method for leaf pigments and structural parameters as described in any one of claims 1 to 8.

Images