Drought measurement method and system based on broadband chlorophyll fluorescence and wavelet network

Abstract

The application discloses a drought measurement method and system based on broadband chlorophyll fluorescence and wavelet neural network, which comprises the following steps: a plant leaf is excited by a pseudo-random binary sequence modulated light, and an original chlorophyll fluorescence signal is collected without dark adaptation; the original signal is preprocessed, time sequence features are extracted, and frequency domain features based on discrete wavelet transform and continuous wavelet transform are extracted; the time sequence features and the frequency domain features are input into a pre-trained multi-branch adaptive wavelet attention network, features are processed and fused through multiple parallel branches of the network, and deep features for representing drought stress are extracted; and a drought stress grade of the plant leaf is classified based on the deep features. The application overcomes the limitations of traditional measurement methods, such as the need for dark adaptation and insensitivity to early stress, realizes rapid, non-destructive and high-precision drought stress identification, and is suitable for field high-throughput monitoring.

Description

Technical Field

[0001] This invention relates to the field of smart information agriculture technology, and in particular to a drought measurement method and system based on broadband chlorophyll fluorescence and wavelet networks. Background Technology

[0002] Chlorophyll fluorescence (ChlF) is a weak light signal released by chlorophyll molecules during plant photosynthesis. Changes in its intensity can reflect the energy utilization efficiency of the photosynthetic system and its response to external environmental stresses. Traditionally, pulse amplitude modulation (PAM) technology has been widely used for chlorophyll fluorescence detection, measuring the maximum photochemical efficiency of leaves after dark adaptation (a commonly used parameter is...). Fv / Fm Indicators such as the ratio of variable fluorescence to maximum fluorescence are used to evaluate Photosystem II (PII). PSII The functional state of photosynthetic capacity is assessed. Furthermore, OJIP transient analysis extracts multiple derived parameters through rapid fluorescence kinetic curves, while rapid light response curves characterize photosynthetic capacity through fluorescence responses under different light intensities. However, these methods face significant limitations in practical field applications: a 15–30 minute dark adaptation period is required before measurement, which cannot meet the needs of continuous, high-throughput monitoring; key parameters such as… Fv / Fm It is not sensitive to early or mild drought stress, and a significant decrease often occurs only when the photosynthetic apparatus has already suffered significant damage; while the actual photochemical efficiency ( Φ PSII ) and photochemical quenching ( qP Parameters such as these are easily affected by ambient light intensity and temperature fluctuations, making standardized measurement under natural conditions difficult; furthermore, non-photochemical quenching ( NPQ Its changes are regulated by a variety of biological and abiotic factors, and are difficult to attribute specifically to drought stress.

[0003] To overcome these limitations, researchers have developed frequency-domain excitation methods, such as sinusoidal modulated photoexcitation, to analyze photosynthetic kinetics by analyzing the frequency response of fluorescence signals. However, sinusoidal excitation is inherently a narrowband signal, with energy concentrated at a single frequency, making it difficult to simultaneously capture multi-timescale dynamics ranging from millisecond-level photochemical reactions to minute-level regulatory processes in a single measurement. Therefore, pseudo-random binary sequences (PRBS) have been introduced as a broadband excitation signal for chlorophyll fluorescence detection. PRBS possesses spectral characteristics similar to white noise, enabling simultaneous excitation of the photosynthetic system across a wide frequency range. It allows for the acquisition of physiological response information across multiple timescales in a single measurement without dark adaptation, and is more sensitive to early stresses. However, the fluorescence signals induced by PRBS exhibit high nonlinearity, non-stationarity, and temporal complexity, making it difficult to fully extract the rich physiological information they contain using traditional time-domain or frequency-domain analysis methods. Although deep learning models such as convolutional neural networks, long short-term memory networks, and Transformers have made progress in time series analysis, they often treat high-frequency fluctuations in PRBS signals as noise and fail to effectively capture subtle time-frequency patterns related to drought stress. Therefore, their performance is limited when decoding such signals with multi-scale dynamic characteristics.

[0004] Therefore, it is necessary to develop a dedicated neural network architecture that can adaptively analyze multi-resolution time-frequency features and integrate multi-modal signal representations to achieve high-precision and robust drought stress identification based on broadband excitation of chlorophyll fluorescence. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problems that traditional chlorophyll fluorescence measurement methods in the prior art require dark adaptation and are not sensitive to early stress response, while existing deep learning models are difficult to effectively extract and analyze the complex multi-scale time-frequency characteristics of fluorescence signals generated by broadband excitation.

[0006] To address the aforementioned technical problems, this invention provides a drought measurement method based on broadband chlorophyll fluorescence and wavelet networks, comprising the following steps: S1: Plant leaves are excited by light modulated by a pseudo-random binary sequence, and the raw chlorophyll fluorescence signal induced by the excitation light is collected under the condition that no dark adaptation is required. S2: Preprocess the original chlorophyll fluorescence signal to extract the width temporal features including the original sequence, upper envelope, lower envelope and excitation light intensity abrupt change, and extract the frequency domain features based on discrete wavelet transform and continuous wavelet transform. S3: Input the time-series features and the frequency-domain features into a pre-trained multi-branch adaptive wavelet attention network. Process different types of input features through multiple parallel branches of the multi-branch adaptive wavelet attention network. Perform feature fusion and optimization on the different types of input features to extract deep features for characterizing drought stress. S4: Based on the deep features, classify the drought stress level of plant leaves to obtain the classification results.

[0007] In one embodiment of the present invention, the method for feature fusion and optimization of the different types of input features in step S3 is as follows: By using a cross-branch attention mechanism, the features of each branch in the multi-branch adaptive wavelet attention network are used as the query vector, while the features of all other branches are used as the key vector and value vector to perform feature interaction and information weighted fusion between branches. By using a cross-modal attention mechanism, the fused branch features are used as query vectors, and the time-encoded features corresponding to the acquisition time are used as key and value vectors to fuse time context information and obtain the deep features.

[0008] In one embodiment of the present invention, in step S3, the method of inputting the temporal features and the frequency domain features into the multi-branch adaptive wavelet attention network is as follows: the original sequence, upper envelope, lower envelope and width feature based on the excitation light intensity change in the temporal features are respectively input into the four sequence branches of the multi-branch adaptive wavelet attention network, and the frequency domain features are input into one feature branch of the multi-branch adaptive wavelet attention network.

[0009] In one embodiment of the present invention, the sequence branches of the multi-branch adaptive wavelet attention network include multiple cascaded processing units. Each processing unit sequentially transforms the input signal to the time-frequency domain through a learnable wavelet convolutional layer to extract multi-resolution features. The multi-resolution features are batch normalized, and the normalized features are nonlinearly transformed by an activation function layer to enhance their expressive power. The nonlinearly transformed features are then pooled and downsampled to reduce dimensionality and retain key information, thus obtaining the output features of the current processing unit.

[0010] In one embodiment of the present invention, the training method of the multi-branch adaptive wavelet attention network is as follows: constructing and combining a cross-entropy loss function that combines class weights and label smoothing; performing data augmentation on the training samples input to the multi-branch adaptive wavelet attention network based on the cross-entropy loss function; and adjusting the learning rate of the multi-branch adaptive wavelet attention network in each cycle of the training process that completes the data augmentation.

[0011] In one embodiment of the present invention, in step S2, the method for extracting the original sequence, upper envelope, lower envelope, and width-time features based on excitation light intensity abrupt changes is as follows: Let the acquired chlorophyll fluorescence original signal be... The pseudo-random binary sequence excitation signal of the first The symbol is The duration of each symbol is The original sequence is The upper envelope By extracting all Fluorescence signals were obtained within the corresponding time period; the lower envelope By extracting all Fluorescence signals were obtained within the corresponding time period; width characteristics were derived from changes in excitation light intensity. Defined as the moment when a pseudo-random binary sequence signal transitions from a low level to a high level. Place.

[0012] In one embodiment of the present invention, in step S2, the method for extracting frequency domain features based on discrete wavelet transform and continuous wavelet transform is as follows: the original chlorophyll fluorescence signal is decomposed by discrete wavelet transform to obtain decomposition coefficients at multiple scales; based on the decomposition coefficients at multiple scales, multiple statistical quantities of each component are extracted to form discrete wavelet transform features. The original chlorophyll fluorescence signal is subjected to continuous wavelet transform to obtain the time-frequency representation of the signal at multiple scales; based on the time-frequency representation, the energy distribution at different scales is calculated and its statistical features are extracted to form continuous wavelet transform features. The discrete wavelet transform features are fused with the continuous wavelet transform features to form the frequency domain features.

[0013] In one embodiment of the present invention, step S2 involves preprocessing the original chlorophyll fluorescence signal, including: standardizing the original signal to eliminate dimensional differences between different samples; and scaling the temporal and frequency domain features.

[0014] This invention also provides a drought measurement system based on broadband chlorophyll fluorescence and wavelet networks, comprising the following modules: The excitation and acquisition module is used to excite plant leaves with pseudo-random binary sequence modulated light and acquire the raw chlorophyll fluorescence signal induced by the excitation light under conditions that do not require dark adaptation. The preprocessing and feature extraction module is used to preprocess the original chlorophyll fluorescence signal, extract the width temporal features including the original sequence, upper envelope, lower envelope and excitation light intensity abrupt change, and extract the frequency domain features based on discrete wavelet transform and continuous wavelet transform. The network model training module is used to input the temporal features and the frequency domain features into a pre-trained multi-branch adaptive wavelet attention network. The multi-branch adaptive wavelet attention network processes different types of input features through multiple parallel branches, performs feature fusion and optimization on the different types of input features, and extracts deep features to characterize drought stress. The classification module is used to classify the drought stress level of plant leaves based on the deep features, and obtain the classification results.

[0015] The present invention also provides a computer storage medium storing a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the drought measurement method based on broadband chlorophyll fluorescence and wavelet networks as described in any one of claims 1 to 8.

[0016] The technical solution of the present invention has the following advantages over the prior art: The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks described in this invention effectively overcomes the limitations of traditional pulse-modulated fluorescence measurements, which require long-term dark adaptation and are insensitive to early drought stress, by using pseudo-random binary sequence-modulated light for broadband excitation. This achieves rapid and non-destructive fluorescence signal acquisition under natural light conditions. By extracting multi-dimensional temporal information such as the original sequence, upper and lower envelopes, and width features based on excitation light intensity abrupt changes, and combining discrete and continuous wavelet transforms to obtain frequency domain representations of the signal at different scales, a multimodal feature set comprehensively reflecting photosynthetic dynamics is constructed. Furthermore, a multi-branch adaptive wavelet attention network is used to process and deeply fuse these features in parallel. The method demonstrates high accuracy and robustness in classifying plant drought stress levels, significantly improving the sensitivity of early drought identification and exhibiting good environmental adaptability and model calibrability. It provides a reliable technical means for high-throughput, real-time crop stress monitoring and precision irrigation management in the field. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the drought measurement method based on broadband chlorophyll fluorescence and wavelet networks provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the architecture of the multi-branch adaptive wavelet attention network in an embodiment of the present invention; Figure 3 This is a schematic diagram of the response of chlorophyll fluorescence to progressive drought stress in the experiment; Figure 4 This is a schematic diagram of the original fluorescence signal waveform under progressive drought stress in the experiment; Figure 5 This is a schematic diagram comparing the five most significant PAM chlorophyll fluorescence parameters and PRBS wavelet-derived features under drought stress levels in the experiment. Figure 6 This is a schematic diagram illustrating the performance of the multi-branch adaptive wavelet attention network model in classifying drought levels by PRBS-induced chlorophyll fluorescence under different drought treatment durations in the experiment. Figure 7 This is a schematic diagram of the spatiotemporal distribution of classification errors in the multi-branch adaptive wavelet attention network model during drought stress classification in the experiment; Figure 8 This is a schematic diagram of the ROC curves for each category of drought stress classification in the experiment; Figure 9 This is a diagram comparing and evaluating the classification performance of different methods on the PRBS dataset in the experiment; Figure 10 This is a schematic diagram showing the improvement in recall rates for each category under different ablation configurations under four drought treatment durations in the experiment. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0020] Example 1: like Figure 1 As shown, this invention provides a drought measurement method based on broadband chlorophyll fluorescence and wavelet networks, comprising the following steps: S1: Plant leaves are excited by light modulated by a pseudo-random binary sequence, and the raw chlorophyll fluorescence signal induced by the excitation light is collected under the condition that no dark adaptation is required. S2: Preprocess the original chlorophyll fluorescence signal to extract the width temporal features including the original sequence, upper envelope, lower envelope and excitation light intensity abrupt change, and extract the frequency domain features based on discrete wavelet transform and continuous wavelet transform. S3: Input the time-series features and the frequency-domain features into a pre-trained multi-branch adaptive wavelet attention network. Process different types of input features through multiple parallel branches of the multi-branch adaptive wavelet attention network. Perform feature fusion and optimization on the different types of input features to extract deep features for characterizing drought stress. S4: Based on the deep features, classify the drought stress level of plant leaves to obtain the classification results.

[0021] This invention provides an intelligent measurement method for plant drought stress based on broadband chlorophyll fluorescence and wavelet attention networks. The method uses PRBS-modulated light to excite plant leaves, acquiring raw chlorophyll fluorescence signals without dark adaptation. The signal is then preprocessed to extract temporal features including the original sequence, upper and lower envelopes, and broadband features based on abrupt changes in excitation light intensity, as well as frequency domain features based on discrete and continuous wavelet transforms. These features are then input into a pre-trained multi-branch adaptive wavelet attention network, where feature fusion and optimization are performed through multi-branch parallel processing and cross-branch, cross-modal attention mechanisms to extract deep stress characterization features. Finally, automatic classification of drought stress levels is achieved based on these features. This method offers advantages such as no dark adaptation required, short measurement time, high sensitivity to early stress responses, and suitability for high-throughput field monitoring, providing reliable technical support for non-destructive, rapid, and accurate diagnosis of plant drought stress.

[0022] In step S1, pseudo-random binary sequence modulated light is used to excite plant leaves, and the original chlorophyll fluorescence signal induced by the excitation is collected under the condition that no dark adaptation is required.

[0023] Specifically, in this embodiment of the invention, the drought stress measurement method is implemented using maize as an example. The field experiment was conducted at an experimental farm in Wuxi, Jiangsu Province. Maize seeds were sown in mid-July, and the plants were irrigated twice daily to ensure optimal growth. By early September, when the maize plants reached the trumpet stage, drought stress treatment was initiated. The soil was fully irrigated to field capacity the night before treatment, and then irrigation was completely stopped for four days, thus forming a progressive drought stress gradient represented by the number of days without irrigation, labeled D1, D2, D3, and D4. To comprehensively monitor the canopy response, a total of 42 chlorophyll fluorescence sensors based on a wireless sensor network were deployed, mounted on tripods, and evenly distributed in three vertical positions in the upper, middle, and lower layers of the maize canopy. All sensors were wirelessly synchronized and controlled by a computer. Measurements were conducted daily from 8:30 AM to 5:30 PM, with data acquisition completed hourly for a total of 10 measurement periods. During each period, both dark-adapted PAM chlorophyll fluorescence signals and non-dark-adapted PRBS-induced chlorophyll fluorescence signals were recorded simultaneously. Under natural light conditions without dark adaptation, broadband excitation of maize leaves was performed using PRBS-modulated light. Specific excitation parameters were as follows: high-level (corresponding to binary "1") light intensity was set to... The low-level (corresponding to binary "0") light intensity is set to .

[0024] The PRBS signal is generated by a 7-bit linear feedback shift register, and its characteristic polynomial is: , in Indicates the shift operator. Indicates the first Output of the level register.

[0025] The feedback equation for the feedback loop is: , in, Indicates the clock cycle. This represents the XOR operation. Indicates the first At the clock cycle, the The state of the level register.

[0026] The PRBS signal is generated by a 7-bit linear feedback shift register, with a sequence length of 127 bits, a single pulse width of 0.5s, and a duration of 63.5s for each PRBS cycle. Each measurement is performed continuously for three cycles.

[0027] Through the above excitation, the system synchronously acquires the original chlorophyll fluorescence signal induced by PRBS-modulated light, thus realizing the rapid, multi-location and non-destructive acquisition of fluorescence signals without dark adaptation in natural field environment, providing a data foundation for subsequent drought stress characteristic analysis and classification.

[0028] Further, in step S2, the original chlorophyll fluorescence signal is preprocessed to extract the original sequence, upper envelope, lower envelope, and width temporal features based on excitation light intensity abrupt changes, and frequency domain features based on discrete wavelet transform and continuous wavelet transform.

[0029] Specifically, dual-domain feature extraction was performed on the collected PRBS-induced chlorophyll fluorescence signals to comprehensively capture the complex temporal dynamics and drought stress response information contained therein.

[0030] In the time domain, let the original chlorophyll fluorescence signal be... , No. Each PRBS symbol is Each symbol contains Each sampling point is used. The upper envelope is extracted through excitation mode decomposition. With lower envelope The fluorescence response during the high-intensity excitation phase and the relaxation and recovery dynamics during the low-intensity excitation phase were characterized, respectively: , .

[0031] This decomposition has clear physiological significance, enabling the separation of the photochemical quenching-dominated response from the non-photochemical quenching recovery process. To further capture the dynamic characteristics of the signal during the light intensity transition transient, a width feature based on the abrupt change in excitation light intensity is defined. This is used to quantify the signal amplitude change at each low-to-high transition moment: , This feature is directly related to the redox state of plastoquinone libraries and PSII efficiency, and can effectively reflect the transient response of photosynthetic electron transfer.

[0032] In the frequency domain, a combination of Discrete Wavelet Transform (DWT) and Continuous Wavelet Transform (CWT) is used to extract the multi-resolution spectral characteristics of the signal. Bior wavelets are employed to analyze the signal. Perform a 6-level discrete wavelet decomposition to obtain the approximation coefficients of each level. With detail coefficient : , , in, and These are low-pass and high-pass decomposition filters, respectively. The number of layers in the decomposition. This is the index variable for the output coefficients.

[0033] For each set of coefficients Extracting the comprehensive statistical feature vector: , in The absolute value of the average. Standard deviation For the difference of squares, For skewness, For kurtosis, For energy, The maximum value of the absolute coefficient. This is the 75th percentile of the absolute coefficients. The approximate symmetry and linear phase properties of the bioorthogonal wavelet family enable it to preserve the signal shape during decomposition.

[0034] To further improve frequency resolution, Morlet wavelet transform is used for continuous wavelet transform, and the CWT coefficients are calculated as follows: , in For scale parameters, For translation parameters, The original time-domain signal, The Morlet mother wavelet is defined as: , exist The CWT coefficients are calculated at each scale, thus yielding the scale-dependent energy distribution: .

[0035] Construct a continuous wavelet eigenvector composed of the statistical moments of the normalized energy distribution and the energy at each scale. .

[0036] By fusing time-domain envelope features with frequency-domain wavelet features, a complete time-frequency hybrid representation is formed: , in, The width characteristics of the signal based on the abrupt change in excitation light intensity. The feature vector extracted from DWT. The feature vector (36-dimensional) extracted from CWT.

[0037] The complete time-frequency hybrid representation includes both the photosynthetic response reflecting the fast transient state and the oscillation mode characterizing the slow regulatory feedback, providing a comprehensive and discriminative feature base for subsequent neural network classification.

[0038] Prior to model training, data quality control was implemented to exclude anomalous records caused by leaf clip detachment, optical path obstruction, or sensor malfunction. The validated dataset comprised 1587 pairs of PAM and PRBS chlorophyll fluorescence signals, collected at ten time intervals (8:30–17:30) and three vertical canopy locations (L1–L3). Outputs from 42 chlorophyll fluorescence sensors were captured simultaneously at each sampling time to ensure comprehensive spatial coverage of the maize canopy. Drought stress severity was categorized into four progressive levels (D1–D4), corresponding to cumulative stress durations of one to four days, as shown in Table 1.

[0039] Table 1:

[0040] In the data preprocessing stage, the raw PRBS fluorescence signal was Z-score normalized: , in and These represent the mean and standard deviation of each signal sequence, respectively. The numerical stability constant is used to eliminate dimensional differences and preserve temporal dynamics. The extracted wavelet coefficient features are standardized using StandardScaler to ensure comparability across different feature dimensions.

[0041] The dataset was divided using a stratified random sampling strategy to maintain consistency in class distribution across subsets. Specifically, 20% of the total samples were reserved as a separate test set, while the remaining 80% constituted the training-validation set.

[0042] To obtain robust performance estimates and mitigate potential overfitting, five-fold hierarchical cross-validation is implemented on the training-validation partition, providing a statistically reliable assessment of the model's generalization ability. Model performance is evaluated using multiple complementary metrics. Classification accuracy quantifies the proportion of correctly classified samples across all drought stress categories. Precision, recall, and F1 score are calculated for each category to evaluate category-specific performance, and a macro-mean is calculated to ensure equal weighting for each category. A one-to-many strategy is used to construct Receiver Operating Characteristic (ROC) curves, and the Area Under the Curve (AUC) quantifies the discriminative ability at different decision thresholds. To address class imbalance, inverse frequency weighting is introduced into the cross-entropy loss function. Label smoothing with a coefficient of 0.1 is used to enhance model calibration, and an early stopping strategy with a patience value of 30 epochs is employed to prevent overfitting.

[0043] Furthermore, such as Figure 2 As shown, in step S3, the temporal features and the frequency domain features are input into a pre-trained multi-branch adaptive wavelet attention network (MB-AWAN). The multi-branch adaptive wavelet attention network processes different types of input features through multiple parallel branches, performs feature fusion and optimization on the different types of input features, and extracts deep features to characterize drought stress.

[0044] Specifically, the multi-branch adaptive wavelet attention network addresses the complexity of drought stress characteristics across multiple time scales through a multi-branch, multi-scale feature extraction paradigm. Inspired by advances in multimodal learning and multi-scale network research, the architecture employs five parallel processing branches to extract complementary signal features, thereby achieving deep analysis of the rich physiological information embedded in PRBS-induced chlorophyll fluorescence signals.

[0045] The input preprocessing module first converts the measured fluorescence signal into four complementary temporal features: the original sequence, the upper envelope, the lower envelope, and a width feature based on the excitation light intensity abrupt change, which is based on the light intensity variation. The original sequence completely preserves the original temporal fluctuations of the signal response to high and low light excitation; the upper and lower envelopes extract the fluorescence response under high and low light intensity conditions, respectively, with the former approximating the maximum fluorescence yield (…). Fm The latter serves as the minimum fluorescence ( Fo Functional proxy; based on the width characteristics of excitation light intensity abrupt changes, by quantifying the amplitude difference between the upper and lower envelopes at each light transition point, it provides a function similar to variable fluorescence. FvThe four features together constitute the inputs to the four sequence branches of the network. Simultaneously, the 92-dimensional frequency domain features extracted from the original signal, based on the fusion of Discrete Wavelet Transform (DWT) and Continuous Wavelet Transform (CWT), are input to an independent high-dimensional feature branch, thus forming five parallel input channels.

[0046] At the core of each sequence branch is a processing pipeline consisting of multiple cascaded WaveletBlock modules. WaveletBlock aims to implement learnable time-frequency transforms, with its first layer being a learnable wavelet convolutional layer initialized using a Morlet wavelet kernel. The Morlet wavelet kernel is defined as follows: , in The center frequency.

[0047] The filters in this convolutional layer are initialized at multiple scales and optimized through backpropagation during network training, enabling adaptive learning of the optimal time-frequency representation and providing excellent time-frequency localization capabilities for non-stationary physiological signals. Each WaveletBlock performs the following operations sequentially: first, the input signal is transformed to the time-frequency domain through the learnable wavelet convolution to extract multi-resolution features; then, batch normalization is performed on the extracted features to stabilize the training process; subsequently, a nonlinear transformation is introduced through a GELU activation function layer to enhance the model's expressive power; finally, downsampling is performed through a max-pooling layer to reduce feature dimensionality, suppress noise, and retain key information. The four cascaded WaveletBlock modules progressively extract hierarchical abstract features. Subsequently, a projection module normalizes the features extracted from each sequence branch to 128 dimensions through linear transformation, layer normalization, and Dropout operations to facilitate subsequent cross-modal interaction.

[0048] To incorporate the crucial contextual information of measurement time into feature learning, the network introduces conditional feature modulation based on the Feature Linear Modulation (FiLM) mechanism. Given the input features of a certain branch... and its corresponding time code The FiLM mechanism generates scaling parameters through two independent linear layers. Translation parameters : , , in, For learnable weight matrix, It is a learnable bias vector.

[0049] Subsequently, the input features are modulated through element-wise multiplication and addition operations: , in This represents element-wise multiplication. This mechanism enables the network to dynamically scale and translate features based on specific measurement time points, thereby adaptively capturing the continuous diurnal variation of photosynthetic activity over time.

[0050] After feature extraction and conditional modulation, the network achieves deep feature fusion and optimization through a two-level attention mechanism.

[0051] The cross-branch attention mechanism enables information exchange and weighted fusion among five parallel branches through a multi-head attention module with a gating mechanism.

[0052] Specifically, attention is calculated using the features of each branch as the query vector, and the features of all other branches as the key and value vectors. This allows each branch to dynamically focus on and integrate complementary information from other modalities. Subsequently, the cross-modal attention mechanism further integrates temporal context information with branch features. This mechanism uses the fused joint branch features as the query and the temporal embedding features generated by the TimeEncoder as the key and value vectors. Through asymmetric attention calculation, the model can recalibrate features based on temporal context, thereby enhancing the model's ability to discriminate stress features within specific time periods. The TimeEncoder combines learnable embeddings for 10 discrete measurement time points with continuous sinusoidal positional encoding to simultaneously represent discrete time labels and continuous temporal flow.

[0053] All features fused and optimized through the attention mechanism are passed to the classification head. The classification head employs a progressive dimensionality reduction strategy, gradually reducing the feature dimension through multiple fully connected layers, and using a decreasing dropout rate between layers to prevent overfitting. Finally, it outputs classification probabilities corresponding to four drought stress levels (D1-D4), accurately identifying the drought stress level of plant leaves. The entire MB-AWAN architecture, through the synergistic effect of the aforementioned multi-branch feature extraction, conditional feature modulation, and hierarchical attention fusion, achieves efficient and robust decoding of multi-scale drought stress features in complex PRBS fluorescence signals.

[0054] Furthermore, in this embodiment, the implementation and training process of the multi-branch adaptive wavelet attention network is completed in the PyTorch 2.7.0 deep learning framework, and GPU-accelerated inference is achieved using CUDA 12.8. The experimental environment is a computing workstation equipped with an NVIDIA GeForce RTX 2080 Ti GPU and running the Ubuntu 24.04.2 LTS operating system, with Python version 3.12.3.

[0055] The model training employs mini-batch stochastic gradient descent with a batch size of 64 and a maximum training epoch of 200 epochs. An early stopping mechanism (with a patience value of 30 epochs) is introduced to suppress overfitting. The optimizer used is AdamW with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001. The learning rate scheduling uses a cosine annealing hot restart strategy, with the update formula as follows: , in, The learning rate for the current cycle. and These are the minimum and maximum learning rates, respectively. This is the current cycle count since the last restart. For the first The restart cycle length is specified. The training objective uses a labeled smoothed weighted cross-entropy loss function, with a label smoothing coefficient set to 0.1. The class weights are calculated based on the inverse frequency of the training set samples, specifically... ,in The total number of samples, For the number of categories, For category The sample size was increased to alleviate class imbalance. To enhance the model's generalization ability, various data augmentation techniques were implemented during training, including Gaussian noise injection, random scaling, time shifting, and Mixup, along with regularization techniques such as Dropout, gradient clipping, and Kaiming initialization.

[0056] experiment: Furthermore, after step S4, the performance of MB-AWAN in the maize drought stress classification task was systematically verified, specifically including the following aspects.

[0057] (1) Analysis of the response characteristics of maize chlorophyll fluorescence to progressive drought stress: Based on the PRBS-induced chlorophyll fluorescence signals collected in this embodiment, their time-frequency response characteristics under four drought stress levels (D1–D4) were analyzed. Figure 3As shown, the signal was decomposed and visualized using Discrete Wavelet Transform (DWT) and Continuous Wavelet Transform (CWT). The results show that as drought stress intensifies (D1 to D3), the energy distribution of the signal at multiple scales exhibits systematic changes, especially showing significant dynamic shifts in the low-frequency and mid-frequency bands. In the D4 stage, its characteristics tend to overlap with those of D3, reflecting the steady-state adaptation of the photosynthetic mechanism under long-term stress.

[0058] Furthermore, such as Figure 4 As shown, the original waveforms of the PAM-induced chlorophyll fluorescence signal (eh) and PRBS-induced chlorophyll fluorescence signal are compared under four drought stress levels (D1-D4). The solid line in the figure represents the average signal, and the shaded area represents... Standard deviation envelope. It can be seen that as stress intensifies, the dynamic range and variability of the PRBS signal exhibit systematic changes, providing an intuitive basis for further time-frequency feature extraction and classification.

[0059] Furthermore, such as Figure 5 As shown, five of the most discriminative wavelet features were extracted by one-way ANOVA, which showed highly significant differences among the drought levels (*p* < 0.001), verifying the effectiveness and sensitivity of PRBS-excited chlorophyll fluorescence in characterizing the process of drought stress.

[0060] (2) Evaluation of the classification performance of the MB-AWAN model: The preprocessed time-series and frequency-domain features are input into the trained MB-AWAN model for drought level classification. The confusion matrix and performance metrics obtained on the independent test set are as follows: Figure 6 As shown, the overall accuracy of the model reached 75.47%, with precision, recall, and F1 score of 75.83%, 75.46%, and 75.53%, respectively, demonstrating balanced discriminative ability. Among them, the classification accuracy of category D4 (severe stress) was the highest (81.2%), while there was some inter-class confusion between D2 and D3, reflecting the continuity of drought stress development.

[0061] To further analyze the distribution characteristics of classification errors, such as Figure 7 As shown, the spatiotemporal distribution of classification errors in the MB-AWAN model during drought stress classification is illustrated. (a) shows the overall classification performance; (b) is a heatmap of the error rate by leaf location and measurement time; (c) and (e) show the spatial distribution and temporal dynamics of the errors, respectively; and (d) stratifies the error rate by measurement time and drought treatment duration. This analysis helps to understand the robustness of the model under different environmental and physiological conditions.

[0062] The robustness of the model was further evaluated using five-fold cross-validation, with an average accuracy of 73.36% ± 2.31%, indicating good generalization performance. Furthermore, feature visualization based on t-SNE showed that the deep features extracted by MB-AWAN could clearly separate overlapping samples in the original signal space into four ordered clusters, further confirming the model's effectiveness in feature learning and category discrimination.

[0063] (3) Comparative analysis with existing advanced methods: To comprehensively evaluate the performance advantages of MB-AWAN, it was compared with several state-of-the-art time series classification models, including LSTM, Transformer, FEDformer, DLinear, Mamba, and TimeXer. The performance comparison results on the same PRBS dataset are shown in Table 2. MB-AWAN outperformed the other methods in all metrics, achieving the highest accuracy (0.755).

[0064] Table 2:

[0065] like Figure 8 As shown, the ROC curves for each category (D1-D4) are further compared. MB-AWAN exhibits the highest AUC values ​​across all categories, with significant improvements, particularly in categories D2 (0.890) and D3 (0.917), demonstrating its superior discriminative ability even during the more challenging moderate drought phase.

[0066] like Figure 9 As shown, further analysis using receiver operating characteristic (ROC) curves and calibration curves revealed that MB-AWAN achieved an area under the curve (AUC) of 0.926, and its predicted probability closely matched the true distribution, demonstrating superior discriminative stability and reliability.

[0067] (4) Ablation experiments and module contribution analysis: To clarify the contribution of each component of MB-AWAN to the final classification performance, ablation experiments were conducted, and the results are shown in Table 3. Starting from the baseline convolutional neural network (CNN-Base), multi-branch temporal features (MB-CNN), adaptive wavelet attention mechanism (AWAN), temporal-frequency domain feature fusion (MB-AWAN-Base), and time coding mechanism (MB-AWAN) were gradually introduced. The classification accuracy of the model improved to 0.616, 0.629, 0.682, and 0.755, respectively. Among them, the complete MB-AWAN improved the performance by 24.6% compared to the baseline model, and the Kappa coefficient reached 0.673, showing the synergistic effect among the modules.

[0068] Table 3:

[0069] like Figure 10 As shown, further analysis of the recall rates of each category revealed that MB-AWAN showed a significant improvement, particularly in the early drought stage (D2) and the severe stress stage (D4), increasing from 29.3% and 40.0% at baseline to 73.2% and 81.2%, respectively. This indicates that the network structure has a significant advantage in capturing subtle physiological changes under different stress intensities.

[0070] Example 2: Based on the same inventive concept as Embodiment 1, the present invention also provides a drought measurement system based on broadband chlorophyll fluorescence and wavelet networks, used to implement the steps of the drought measurement method based on broadband chlorophyll fluorescence and wavelet networks described in Embodiment 1, including the following modules: The excitation and acquisition module is used to excite plant leaves with pseudo-random binary sequence modulated light and acquire the raw chlorophyll fluorescence signal induced by the excitation light under conditions that do not require dark adaptation. The preprocessing and feature extraction module is used to preprocess the original chlorophyll fluorescence signal, extract the width temporal features including the original sequence, upper envelope, lower envelope and excitation light intensity abrupt change, and extract the frequency domain features based on discrete wavelet transform and continuous wavelet transform. The network model training module is used to input the temporal features and the frequency domain features into a pre-trained multi-branch adaptive wavelet attention network. The multi-branch adaptive wavelet attention network processes different types of input features through multiple parallel branches, performs feature fusion and optimization on the different types of input features, and extracts deep features to characterize drought stress. The classification module is used to classify the drought stress level of plant leaves based on the deep features, and obtain the classification results.

[0071] The excitation and acquisition module, preprocessing and feature extraction module, network model training module, and classification module of the drought measurement system based on broadband chlorophyll fluorescence and wavelet network proposed in this embodiment are used to implement steps S1, S2, S3, and S4 in the drought measurement method based on broadband chlorophyll fluorescence and wavelet network in Embodiment 1, respectively. To avoid redundancy, they will not be described in detail here.

[0072] Example 3: The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the drought measurement method based on broadband chlorophyll fluorescence and wavelet networks described in Embodiment 1.

[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A drought measurement method based on broadband chlorophyll fluorescence and wavelet networks, characterized in that, Includes the following steps: S1: Plant leaves are excited by light modulated by a pseudo-random binary sequence, and the raw chlorophyll fluorescence signal induced by the excitation light is collected under the condition that no dark adaptation is required. S2: Preprocess the original chlorophyll fluorescence signal to extract the width temporal features including the original sequence, upper envelope, lower envelope and excitation light intensity abrupt change, and extract the frequency domain features based on discrete wavelet transform and continuous wavelet transform. S3: Input the time-series features and the frequency-domain features into a pre-trained multi-branch adaptive wavelet attention network. Process different types of input features through multiple parallel branches of the multi-branch adaptive wavelet attention network. Perform feature fusion and optimization on the different types of input features to extract deep features for characterizing drought stress. S4: Based on the deep features, classify the drought stress level of plant leaves to obtain the classification results.

2. The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks according to claim 1, characterized in that: In step S3, the method for feature fusion and optimization of the different types of input features is as follows: By using a cross-branch attention mechanism, the features of each branch in the multi-branch adaptive wavelet attention network are used as the query vector, while the features of all other branches are used as the key vector and value vector to perform feature interaction and information weighted fusion between branches. By using a cross-modal attention mechanism, the fused branch features are used as query vectors, and the time-encoded features corresponding to the acquisition time are used as key and value vectors to fuse time context information and obtain the deep features.

3. The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks according to claim 1, characterized in that: In step S3, the method of inputting the temporal features and the frequency domain features into the multi-branch adaptive wavelet attention network is as follows: the original sequence, upper envelope, lower envelope and width feature based on the excitation light intensity change in the temporal features are respectively input into the four sequence branches of the multi-branch adaptive wavelet attention network, and the frequency domain features are input into one feature branch of the multi-branch adaptive wavelet attention network.

4. The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks according to claim 1 or 3, characterized in that: The sequence branches of the multi-branch adaptive wavelet attention network include multiple cascaded processing units. Each processing unit sequentially transforms the input signal to the time-frequency domain through a learnable wavelet convolutional layer to extract multi-resolution features. The multi-resolution features are batch normalized, and the normalized features are nonlinearly transformed by an activation function layer to enhance their expressive power. The nonlinearly transformed features are then pooled and downsampled to reduce dimensionality and retain key information, thus obtaining the output features of the current processing unit.

5. The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks according to claim 1, characterized in that: The training method for the multi-branch adaptive wavelet attention network is as follows: constructing and combining a cross-entropy loss function that smooths the class weights and labels; and performing data augmentation on the training samples input to the multi-branch adaptive wavelet attention network based on the cross-entropy loss function. The learning rate of the multi-branch adaptive wavelet attention network is adjusted in each cycle of the data augmentation training process.

6. The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks according to claim 1, characterized in that: In step S2, the method for extracting the original sequence, upper envelope, lower envelope, and width-time features based on excitation light intensity abrupt changes is as follows: Let the acquired chlorophyll fluorescence original signal be... The pseudo-random binary sequence excitation signal of the first The symbol is The duration of each symbol is The original sequence is The upper envelope By extracting all Fluorescence signals were obtained within the corresponding time period; the lower envelope By extracting all Fluorescence signals were obtained within the corresponding time period; width characteristics were derived from changes in excitation light intensity. Defined as the moment when a pseudo-random binary sequence signal transitions from a low level to a high level. Place.

7. The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks according to claim 1, characterized in that: In step S2, the method for extracting frequency domain features based on discrete wavelet transform and continuous wavelet transform is as follows: the original chlorophyll fluorescence signal is decomposed by discrete wavelet transform to obtain decomposition coefficients at multiple scales; based on the decomposition coefficients at multiple scales, multiple statistics of each component are extracted to form discrete wavelet transform features. The original chlorophyll fluorescence signal is subjected to continuous wavelet transform to obtain the time-frequency representation of the signal at multiple scales; based on the time-frequency representation, the energy distribution at different scales is calculated and its statistical features are extracted to form continuous wavelet transform features. The discrete wavelet transform features are fused with the continuous wavelet transform features to form the frequency domain features.

8. The drought measurement method based on broadband chlorophyll fluorescence and wavelet networks according to claim 1, characterized in that: In step S2, the original chlorophyll fluorescence signal is preprocessed, including: standardizing the original signal to eliminate dimensional differences between different samples; and scaling the time-series features and frequency-domain features.

9. A drought measurement system based on broadband chlorophyll fluorescence and wavelet networks, characterized in that, Includes the following modules: The excitation and acquisition module is used to excite plant leaves with pseudo-random binary sequence modulated light and acquire the raw chlorophyll fluorescence signal induced by the excitation light under conditions that do not require dark adaptation. The preprocessing and feature extraction module is used to preprocess the original chlorophyll fluorescence signal, extract the width temporal features including the original sequence, upper envelope, lower envelope and excitation light intensity abrupt change, and extract the frequency domain features based on discrete wavelet transform and continuous wavelet transform. The network model training module is used to input the temporal features and the frequency domain features into a pre-trained multi-branch adaptive wavelet attention network. The multi-branch adaptive wavelet attention network processes different types of input features through multiple parallel branches, performs feature fusion and optimization on the different types of input features, and extracts deep features to characterize drought stress. The classification module is used to classify the drought stress level of plant leaves based on the deep features, and obtain the classification results.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, the computer software product including several instructions for causing a computer device to execute the drought measurement method based on broadband chlorophyll fluorescence and wavelet networks as described in any one of claims 1 to 8.

Images