Method for detecting nitrogen content grade of corn plant based on leaf surface potential time-frequency spectrum

Abstract

The invention relates to the field of corn leaf nitrogen content detection, and particularly discloses a leaf surface potential time-frequency atlas-based corn plant nitrogen content grade detection method, which comprises the following steps: S1, culturing corn plants under different nitrogen content conditions and collecting related data; s2, preprocessing the acquired surface potential data of the blade; s3, performing time-frequency domain feature extraction and data set division on the preprocessed signal; s4, establishing and optimizing a corn nitrogen content grade detection model; and S5, carrying out nitrogen content grade detection by using the optimized detection model. According to the technical scheme, the problems that an existing detection method is high in subjectivity, lossy, prone to interference, high in cost, complex in process, insufficient in single time domain information and the like are solved, the corn nitrogen content detection losslessness and early-stage detection capacity and precision are effectively improved, and a path is provided for plant electric signal physiological and ecological detection.

Description

Technical Field

[0001] This invention relates to the field of nitrogen content detection technology in maize leaves, specifically a method for detecting the nitrogen content level of maize plants based on leaf surface potential time-frequency spectrum. Background Technology

[0002] As an important food crop in my country, nitrogen is a key element essential for the growth of corn. Nitrogen deficiency can lead to slow growth, stunted plants, and even yellowing and death of leaves. Therefore, it is important to quickly obtain the nitrogen content of corn, especially to detect nitrogen deficiency in the early stages.

[0003] There are several different detection methods among existing technologies for detecting nitrogen deficiency in maize. Visual diagnostic methods rely on observation of the appearance of maize leaves, but they are highly subjective, depend on the experience of the testers, and are difficult to meet the requirements of early diagnosis and online detection. Chemical analysis methods such as the Kjeldahl method and colorimetric methods require chemical treatment of maize samples, which are destructive tests with long testing cycles, making rapid detection impossible. In addition, the sampling process is easily affected by human factors, resulting in large detection errors. Spectroscopic analysis methods use spectral characteristics to detect nitrogen content, but are easily affected by factors such as plant leaf moisture, light conditions, and leaf structure, and the detection equipment is expensive. Methods that indirectly detect nitrogen through other physiological indicators such as chlorophyll content and enzyme activity are easily affected by environmental factors, have complex testing processes, and require cumbersome preparation of reagents and equipment.

[0004] Plant electrical signals are widespread and important physiological signals within plants, participating in physiological regulation and transmitting growth-related information. They are the initial responses of plants to external stimuli, and leaf surface potential is a comprehensive reflection of plant electrical signals at the organ level. Under nitrogen deficiency conditions, plants produce physiological responses that affect their own electrical signals. This characteristic provides a new approach to detecting nitrogen content in maize and lays the foundation for developing novel early detection methods for nitrogen deficiency.

[0005] However, existing detection methods still have significant limitations: they struggle to achieve both non-destructive and rapid detection while maintaining accuracy and stability in early detection, thus failing to effectively meet the demand for timely and precise monitoring of nitrogen deficiency in maize production. Therefore, developing a non-destructive, rapid, and accurate method for detecting early nitrogen deficiency in maize based on plant electrical signal characteristics has become a key and challenging issue that urgently needs to be addressed in the field of agricultural production and detection technology. Summary of the Invention

[0006] This invention addresses the aforementioned problems of existing technologies by providing a method for detecting the nitrogen content level of maize plants based on leaf surface potential time-frequency spectrum. This method effectively solves the problems of strong subjectivity, destructiveness, and susceptibility to interference in existing detection methods, and effectively improves the non-destructive nature, early detection capability, and accuracy of maize nitrogen content detection.

[0007] To achieve the above objectives, this invention proposes a method for detecting the nitrogen content level of maize plants based on leaf surface potential time-frequency spectrum, comprising: S1. Cultivating maize plants under different nitrogen content conditions and collecting relevant data: Maize plants cultivated to a specific leaf stage were transplanted into perlite substrate cultivation containers. Hogrange nutrient solution with different nitrogen content ratios was applied regularly to cultivate maize plants under different nitrogen-deficient conditions. After a specific cultivation period, the relative chlorophyll content of maize leaves was measured. Subsequently, in a Faraday electromagnetic grid, a multi-channel data acquisition system with weak electrical signals was used in combination with specific electrodes to collect surface potential information of maize leaves, and the collected data was classified and stored. S2. Preprocessing the collected blade surface potential data: The electrical signal is decomposed using an improved adaptive noise complete set empirical mode decomposition; based on frequency domain analysis, noise-dominant IMF components are screened, and the dominant frequency of each IMF is calculated by fast Fourier transform. High-frequency noise and low-frequency baseline drift components are removed by combining the sample entropy with a preset threshold, and useful IMF components are retained; the useful components are denoised using an improved wavelet threshold function to obtain the preprocessed signal. S3. Extract time-frequency domain features and divide the dataset for the preprocessed signal: Use short-time Fourier transform to convert the preprocessed one-dimensional surface potential signal into a two-dimensional time-frequency domain image, and then divide the time-frequency domain images corresponding to different nitrogen content levels into training set, validation set and test set. S4. Establish and optimize the corn nitrogen content level detection model: Based on the segmented time-frequency domain images, a detection model is established using a hybrid deep learning architecture that combines residual networks and long short-term memory networks. Accuracy, recall, precision and F1 score are used as model performance evaluation indicators. The model parameters are optimized and adjusted to finally obtain the corn leaf nitrogen content level detection model. S5. Use the optimized detection model to detect nitrogen content level: Input the time-frequency domain image obtained by preprocessing and short-time Fourier transforming the surface potential signal of the corn leaf to be detected into the detection model, and output the detection result of nitrogen content level of corn plant.

[0008] Preferably, in S1, the electrode includes an AgCl patch electrode and a silver needle, with the AgCl patch electrode serving as the sampling electrode and the BIAS electrode, and the silver needle serving as the reference electrode; the sampling electrode is uniformly attached to the corn leaf and fixed to prevent displacement.

[0009] Preferably, in S1, the process of collecting surface potential information of corn leaves includes: first, placing the corn in a dark environment and letting it stand for a specific time before starting to collect data, and then collecting potential data under dark conditions, under light-on conditions, and under dark conditions again in sequence. Multiple collections are made for each leaf, and multiple collections are made for each group of corn plants to obtain multiple sets of data.

[0010] Preferably, in S2, when improving the wavelet threshold function for noise reduction, the Daubechies5 wavelet is selected, the number of decomposition layers is set to 5, and the threshold selection rule is heursure; in the process of improved adaptive noise complete set empirical mode decomposition, the noise standard deviation ratio, the number of integrations and the maximum number of iterations need to be set, white noise is added one by one and the residual signal is calculated, each order of IMF is extracted, and the IMF matrix is ​​output.

[0011] Preferably, in S2, the sample entropy is used to quantify the signal complexity. The larger the sample entropy value, the higher the signal complexity. The components that need to be processed are screened by calculating the sample entropy of the effective IMF. After screening, useful IMF components within a specific number range are retained, and noise or redundant components are removed.

[0012] Preferably, in S3, the short-time Fourier transform simultaneously preserves the time-domain and frequency-domain features of the electrical signal on the plant surface. The time-domain features include peak-to-peak value, maximum value, minimum value, and mean value, while the frequency-domain features include the power spectrum.

[0013] Preferably, in S3, the ratio of the training set, validation set, and test set is determined based on the amount of data and the model training requirements, ensuring that the time-frequency domain images corresponding to various nitrogen content levels are reasonably distributed in each dataset.

[0014] Preferably, in S4, the residual network is the front-end module that extracts spatial features and extracts multi-level spatial features through jump connections in the residual blocks. The long short-term memory network is the back-end module that receives the spatial features extracted by the residual network through input gates, forget gates, and output gates and learns the time dynamics and long-term dependencies.

[0015] Preferably, in S4, the residual network contains three residual stages, each stage contains two residual blocks; each residual block contains two convolutional layers, each convolutional layer is followed by a BatchNorm and a ReLU activation function; if the number of channels in the input and output of the residual block is inconsistent or downsampling is required, downsampling is used to adjust the dimension of the residual path.

[0016] Preferably, in S4, the output of the residual network is compressed in height through adaptive average pooling, and then input into the long short-term memory network while keeping the width unchanged. The long short-term memory network is configured with hidden layer dimensions and number of layers, and uses the dropout mechanism to prevent model overfitting. The fully connected layer of the detection model maps the output of the long short-term memory network to the number of corn nitrogen content level categories.

[0017] Therefore, this invention proposes a method for detecting the nitrogen content level of maize plants based on the time-frequency spectrum of leaf surface potential, which has the following beneficial effects: (1) A novel non-destructive detection scheme for nitrogen stress in maize based on plant electrical signals is provided. It does not require damage to maize plants and can achieve early detection of nitrogen deficiency, making up for the shortcomings of existing visual diagnosis and chemical analysis methods in terms of early detection and non-destructiveness.

[0018] (2) The one-dimensional leaf surface potential signal is converted into a two-dimensional time-frequency image by short-time Fourier transform, which fully preserves the time and frequency domain features of the signal. Combined with the ResNet-LSTM hybrid deep learning architecture, the accuracy of nitrogen content level detection of maize plants is significantly improved.

[0019] (3) Taking leaf surface potential as the core detection index avoids the interference of leaf moisture, light and other factors on the spectral analysis method. At the same time, it provides a feasible practical path for the development of physiological and ecological detection technology based on plant electrical signals, which has practical application significance.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a flowchart of the detection method for the nitrogen content level detection method of maize plants based on the time-frequency spectrum of leaf surface potential of the present invention; Figure 2 This is a schematic diagram of the ResNet-LSTM model of the method for detecting nitrogen content levels in maize plants based on leaf surface potential time-frequency spectrum in this invention. Figure 3 This is a flowchart of the ResNet-LSTM model for the method of detecting nitrogen content levels in maize plants based on leaf surface potential time-frequency spectrum, as described in this invention. Figure 4 This is a schematic diagram of perlite matrix filling for the method of detecting nitrogen content in maize plants based on leaf surface potential time-frequency spectrum in this invention; Figure 5 This is an experimental schematic diagram of the data acquisition device for the method of detecting nitrogen content levels in maize plants based on leaf surface potential time-frequency spectrum, as per the present invention. Figure 6 This is a waveform diagram of specific data on the surface potential of maize leaves collected by the maize plant nitrogen content level detection method based on the leaf surface potential time-frequency spectrum of the present invention; Figure 7 This invention presents the time-domain waveforms of the first four IMF components obtained by decomposing maize leaf surface potential signals under different nitrogen content conditions using the leaf surface potential time-frequency spectrum method for detecting nitrogen content levels in maize plants based on leaf surface potential time-frequency spectrum. Figure 8 This is a waveform diagram showing the time-frequency domain characteristics of the remaining low-order IMF components after ICEEMDAN decomposition, based on the leaf surface potential time-frequency spectrum method for detecting nitrogen content levels in maize plants according to the present invention. Figure 9 This is the time-frequency graph after STFT processing of the method for detecting nitrogen content levels in maize plants based on leaf surface potential time-frequency graphs of the present invention; Figure 10 This is the ResNet-LSTM training curve of the method for detecting nitrogen content levels in maize plants based on leaf surface potential time-frequency spectrum, which is the basis of this invention. Detailed Implementation

[0022] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0024] like Figures 1-10 As shown, the method for detecting the nitrogen content level of maize plants based on leaf surface potential time-frequency spectrum provided by the present invention includes: S1. Cultivating maize plants under different nitrogen content conditions and collecting relevant data: Maize plants cultivated to a specific leaf stage were transplanted into perlite substrate cultivation containers. Hogrange nutrient solution with different nitrogen content ratios was applied regularly to cultivate maize plants under different nitrogen-deficient conditions. After a specific cultivation period, the relative chlorophyll content of maize leaves was measured. Subsequently, in a Faraday electromagnetic grid, a multi-channel data acquisition system with weak electrical signals was used in combination with specific electrodes to collect surface potential information of maize leaves, and the collected data was classified and stored. The electrodes include an AgCl patch electrode and a silver needle. The AgCl patch electrode serves as both the acquisition electrode and the BIAS electrode, while the silver needle serves as the reference electrode. The acquisition electrode is uniformly attached to the corn leaf and fixed in place to prevent displacement.

[0025] The process of collecting surface potential information of maize leaves includes: first, placing the maize in a dark environment and letting it stand for a specific time before starting the collection; then, collecting potential data under dark conditions, under light-on conditions, and under dark conditions again in sequence; collecting data multiple times for each leaf; and collecting data multiple times for each group of maize plants to obtain multiple sets of data.

[0026] S2. Preprocessing the collected blade surface potential data: The electrical signal is decomposed using an improved adaptive noise complete set empirical mode decomposition; based on frequency domain analysis, noise-dominant IMF components are screened, and the dominant frequency of each IMF is calculated by fast Fourier transform. High-frequency noise and low-frequency baseline drift components are removed by combining the sample entropy with a preset threshold, and useful IMF components are retained; the useful components are denoised using an improved wavelet threshold function to obtain the preprocessed signal. When improving the wavelet threshold function for noise reduction, the Daubechies5 wavelet is selected, the number of decomposition layers is set to 5, and the threshold selection rule is heursure. In the process of improving the adaptive noise complete set empirical mode decomposition, the noise standard deviation ratio, the number of integrations and the maximum number of iterations need to be set. White noise is added one by one and the residual signal is calculated. IMFs of each order are extracted and the IMF matrix is ​​output.

[0027] Sample entropy is used to quantify signal complexity. The larger the sample entropy value, the higher the signal complexity. By calculating the sample entropy of the effective IMF, the components that need to be processed are screened. After screening, useful IMF components within a specific number range are retained, and noise or redundant components are removed.

[0028] S3. Extract time-frequency domain features and divide the dataset for the preprocessed signal: Use short-time Fourier transform to convert the preprocessed one-dimensional surface potential signal into a two-dimensional time-frequency domain image, and then divide the time-frequency domain images corresponding to different nitrogen content levels into training set, validation set and test set. The short-time Fourier transform preserves both the time-domain and frequency-domain features of the electrical signals on the plant surface. The time-domain features include peak-to-peak value, maximum value, minimum value, and mean value, while the frequency-domain features include the power spectrum.

[0029] The proportions of the training set, validation set, and test set are determined based on the amount of data and the model training requirements, ensuring that the time-frequency domain images corresponding to various nitrogen content levels are reasonably distributed in each dataset.

[0030] S4. Establish and optimize the corn nitrogen content level detection model: Based on the segmented time-frequency domain images, a detection model is established using a hybrid deep learning architecture that combines residual networks and long short-term memory networks. Accuracy, recall, precision and F1 score are used as model performance evaluation indicators. The model parameters are optimized and adjusted to finally obtain the corn leaf nitrogen content level detection model. The residual network is the front-end module that extracts spatial features. It extracts multi-level spatial features through jump connections in the residual blocks. The long short-term memory network is the back-end module that receives the spatial features extracted by the residual network through input gates, forget gates, and output gates, and learns the temporal dynamics and long-term dependencies.

[0031] The residual network consists of three residual stages, each containing two residual blocks; each residual block contains two convolutional layers, followed by BatchNorm and ReLU activation functions; if the number of input and output channels of the residual block is inconsistent or downsampling is required, downsampling is used to adjust the dimension of the residual path.

[0032] The output of the residual network is compressed in height through adaptive average pooling, and then fed into the long short-term memory network while keeping the width unchanged. The long short-term memory network is configured with hidden layer dimensions and number of layers, and uses the dropout mechanism to prevent model overfitting. The fully connected layer of the detection model maps the output of the long short-term memory network to the number of corn nitrogen content level categories.

[0033] S5. Use the optimized detection model to detect nitrogen content level: Input the time-frequency domain image obtained by preprocessing and short-time Fourier transforming the surface potential signal of the corn leaf to be detected into the detection model, and output the detection result of nitrogen content level of corn plant.

[0034] Example 1 This embodiment uses Suke waxy corn as the test object. By controlling the nitrogen content gradient and standardizing the data collection and processing process, the nitrogen content level of corn plants is detected. The specific steps are as follows: I. Maize Plant Cultivation and Experimental Environment Control: Select Suke waxy corn seeds and cultivate them first in seedling trays with seedling soil, watering them daily in measured amounts. When the corn reaches the 2-3 leaf stage, transplant them into perlite substrate cultivation containers. The bottom and top layers of the cultivation containers are filled with large-particle hard perlite (8-12mm), and the middle layer is covered with 3-6mm small-particle perlite. This utilizes capillary action to improve water retention and stabilize seedling root development. Figure 4 As shown. The entire experiment was conducted indoors, with the ambient temperature controlled at 24℃-26℃ and the humidity at 50%RH-60%RH. Plant fluorescent lights and supplemental lights were used to provide the light source.

[0035] Four different nitrogen concentration ratios of Hogland nutrient solution were applied regularly. The 100% standard formula of Hogland nutrient solution was: 2.185 g / L containing NH4+. + 1.0 mmol, NO3 -Nutrient solutions containing 14 mmol of P, 1.0 mmol of P, 6.0 mmol of K, 4.0 mmol of Ca, 2.0 mmol of Mg, and 2.0 mmol of S, and trace elements (at standard doses) of Cu 0.02 mg, Fe 2.8 mg, Zn 0.05 mg, Mn 0.5 mg, B 0.5 mg, and Mo 0.01 mg, with nitrogen concentrations of 0 mmol / L, 0.075 mmol / L, 0.75 mmol / L, and 7.5 mmol / L, respectively, corresponding to categories 0-3, were cultured continuously for 40 days. The four nitrogen concentration gradients are shown in Table 1.

[0036] Table 1 Nitrogen concentrations in different groups

[0037] II. Data Collection Preparation and Execution: After 40 days of cultivation, it was found that the corn leaves with extremely low nitrogen levels had developed surface traits such as yellowing, wilting, and stunting. Further cultivation revealed a stepped shape pattern, at which point a plant electrical sampling experiment was conducted.

[0038] First, the relative chlorophyll content (SPAD) of maize leaves was measured using a plant nutrient analyzer. The content was measured multiple times and averaged. After analyzing the SPAD values, it was determined that category 0 was severe nitrogen deficiency, category 1 was moderate nitrogen deficiency, category 2 was mild nitrogen deficiency, and category 3 was normal. The average SPAD results of maize cultured in each nutrient solution are shown in Table 2.

[0039] Table 2. SPAD values ​​of maize leaves under nutrient solutions with different nitrogen contents.

[0040] Subsequently, a layer of 100-mesh brass mesh and a layer of 100-mesh copper mesh were tightly glued to a custom-made iron cage with hot melt adhesive. Inside the custom-made 90cm×90cm×200cm grounded Faraday cage, a weak multi-channel electrical signal acquisition system was used to collect the surface potential of the blade.

[0041] Using AgCl medical patch electrode as the acquisition electrode and BIAS electrode and silver needle as the reference electrode, the acquisition electrode was evenly attached to the leaf and pressed and fixed. The experimental schematic diagram of the acquisition device is shown below. Figure 5 As shown.

[0042] Before data collection, the corn was placed in a dark environment within a Faraday cage. Four AgCl electrodes were attached to the leaves, and the collection was initiated after 10 minutes of inactivity. Potential data were collected sequentially under conditions of 2 minutes in darkness, 3 minutes with the lights on, and 2 minutes in darkness, with each data session lasting 7 minutes. Five measurements were taken per leaf, and 50 measurements were taken per plant group. Each data collection session contained four channels, resulting in 200 data points for each category, for a total of 800 data points. Each data point contained 105,000 data points, which were then stored separately. The data collection results for different nitrogen content groups are shown in Table 3.

[0043] Table 3 Data Collection Results

[0044] The specific waveforms of the collected corn leaf surface potential data are as follows: Figure 6 As shown in the image, most plant electrical signals are below 10 Hz, with amplitudes generally around 20 μF.

[0045] III. Data Preprocessing Operations: A combined denoising method of "ICEEMDAN + sample entropy + improved wavelet threshold" was used to reduce the noise of the collected maize leaf surface potential signal. First, the signal is input into ICEEMDAN for signal decomposition: the noise standard deviation ratio Nstd is set to 0.1, the number of integrations NE is 50, and the maximum number of iterations MaxIter is 300. White noise is added successively, and the residual signal is calculated. The IMF components of each order are extracted. By adding adaptive white noise and performing multiple integration averaging, the mode mixing problem is reduced, and the IMF matrix is ​​output. The ICEEMDAN decomposition diagram is shown below. Figures 7-8 As shown.

[0046] The noise standard deviation ratio Nstd ranges from 0.1 to 0.3, and the number of integrations NE ranges from 50 to 100. In other application embodiments, the noise standard deviation ratio Nstd, the number of integrations NE, and the maximum number of iterations MaxIter are set as needed.

[0047] Next, IMF component screening: Based on frequency domain analysis, noise-dominant IMF components are screened. The dominant frequency of each IMF is calculated by Fast Fourier Transform (FFT). Combined with the sample entropy SpEn preset threshold of 0.1, high-frequency noise and low-frequency baseline drift components are removed. The sample entropy values ​​and frequency values ​​of each IMF component are shown in Table 4.

[0048] The results showed that the main frequencies after IMF15 were only below 0.0071Hz, which often represented extremely low frequency trends or baseline drift other than residual components. After frequency screening and setting the sample entropy threshold to 0.1, only the useful IMF components numbered 5-15 were retained.

[0049] Table 4. IMF component sample entropy and frequency values

[0050] Finally, wavelet denoising: Wavelet thresholding is performed to suppress residual noise. Daubechies5 (db5) wavelet is selected, with a decomposition level of 5 and a threshold rule of heursure. The retained IMF components are denoised to obtain the preprocessed signal.

[0051] The signal-to-noise ratio (SNR), mean square error (MSE), and normalized cross-correlation (NCC) were selected as evaluation metrics. The comparison showed that the ICEEMDAN filtering method performed better, as shown in Table 5.

[0052] Table 5 Evaluation Indicators for Different Filtering Methods

[0053] IV. Time-Frequency Domain Feature Extraction and Dataset Partitioning: The preprocessed one-dimensional potential signal is converted into a two-dimensional time-frequency domain image using a short-time Fourier transform (STFT), preserving time-domain characteristics, peak-to-peak value, maximum value, minimum value, mean value, and frequency-domain power spectrum. The transformed time-frequency image is shown below. Figure 9 As shown.

[0054] The transformed data was divided into training, validation, and test sets in a ratio of 8:1:1. The dataset distribution is shown in Table 6. 200 images were originally collected for each data category, with each image lasting 7 minutes, totaling 2800 images.

[0055] Table 6 Dataset Distribution

[0056] V. Model Establishment, Optimization, and Testing: A detection model based on a ResNet-LSTM hybrid architecture: based on Figure 2 and Figure 3 The ResNet-LSTM architecture is designed with a 224×224 image input in 3-channel RGB in the convolutional layer, using a 3×3 convolutional kernel with a stride of 1 and padding of 1, followed by BatchNorm and ReLU activation functions, and then through a 2×2 max pooling layer, outputting a 16-channel 112×112 feature map.

[0057] Three residual stages are designed after the convolutional layers. Each stage contains two residual blocks, and each residual block contains two convolutional layers. Each convolutional layer is followed by BatchNorm and ReLU activation functions. When the number of input and output channels is inconsistent or downsampling is required, downsampling is used to adjust the dimension of the residual path.

[0058] The first stage inputs 16 channels and outputs 16 channels, with the feature map maintaining a size of 112×112; the second stage inputs 16 channels and outputs 32 channels, with a feature map size of 56×56; the third stage inputs 32 channels and outputs 64 channels, with a feature map size of 28×28.

[0059] The residual output is compressed to a height of 14 while maintaining a width of 28 using adaptive average pooling, resulting in an output size of 14×28. This is then fed into an LSTM with an input dimension of 64×28, two hidden layers of 256, and a dropout of 0.4 to prevent overfitting, resulting in an output dimension of 256. Finally, a fully connected layer maps the output to four classes, producing the detection results. The training curve is shown below. Figure 10 As shown.

[0060] The model parameters were optimized using accuracy, recall, precision, and F1 score as evaluation metrics. The time-frequency domain images of the test set were input into the model, and the detection results were obtained: Category 0 had the highest number of correct predictions (246) with an accuracy of 87.86%; Category 1 had the highest number of correct predictions (196) with an accuracy of 70.50%; Category 2 had the highest number of correct predictions (209) with an accuracy of 81.01%; and Category 3 had the highest number of correct predictions (233) with an accuracy of 76.64%, thus achieving effective detection of nitrogen content levels in corn.

[0061] Therefore, this invention provides a method for detecting the nitrogen content level of maize plants based on the time-frequency spectrum of leaf surface potential. This method involves phased completion of maize plant cultivation and data acquisition, leaf surface potential signal preprocessing, time-frequency domain feature extraction and dataset partitioning, detection model establishment and optimization, and level detection. Using leaf surface potential as the core detection indicator, and combining an improved adaptive noise complete set empirical mode decomposition, short-time Fourier transform, and ResNet-LSTM hybrid deep learning architecture, this method achieves non-destructive and early detection of maize nitrogen content while avoiding the problems of strong subjectivity, susceptibility to interference, high cost, and complex processes inherent in existing detection methods. This significantly improves detection accuracy and provides a practical solution for the application of crop physiological and ecological detection technology based on plant electrical signals, effectively meeting the needs of accurate monitoring of nitrogen deficiency in maize during agricultural production.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting the nitrogen content level of a corn plant based on the time-frequency spectrum of leaf surface potential, characterized in that, The application relates to a method for detecting the nitrogen content level of corn leaves. S1, cultivating corn plants under different nitrogen content conditions and collecting relevant data: corn plants cultivated to a specific leaf stage are transplanted into perlite substrate cultivation containers, and different nitrogen content ratios of Hoagland nutrient solution are regularly applied to cultivate corn plants under different nitrogen deficiency conditions; After a specific time of cultivation, the relative content of chlorophyll in the corn leaves is measured, then the surface potential information of the corn leaves is collected in a Faraday electromagnetic net using a weak electric signal multi-channel data acquisition system combined with specific electrodes, and the collected data is classified and stored; S2, preprocessing the collected leaf surface potential data: the electric signal is decomposed by improved adaptive noise complete ensemble empirical mode decomposition; Based on frequency domain analysis, noise-dominant IMF components are screened, the main frequency of each IMF is calculated by fast Fourier transform, high-frequency noise and low-frequency baseline drift components are removed by combining a preset threshold value of sample entropy, and useful IMF components are retained; the useful components are denoised by using an improved wavelet threshold function to obtain preprocessed signals; S3, extracting time-frequency domain features from the preprocessed signals and dividing the data set: the one-dimensional surface potential signal after preprocessing is converted into a two-dimensional time-frequency domain image by using a short-time Fourier transform, then the time-frequency domain images corresponding to different nitrogen content levels are divided into a training set, a validation set and a test set; S4, establishing and optimizing a corn nitrogen content level detection model: based on the divided time-frequency domain images, a hybrid deep learning architecture combining a residual network and a long short-term memory network is used to establish the detection model, the accuracy, recall rate, precision and F1 value are used as the model performance evaluation indexes, the model parameters are optimized and adjusted, and finally the corn leaf nitrogen content level detection model is obtained; S5, using the optimized detection model to detect the nitrogen content level: the time-frequency domain image of the surface potential signal of the corn leaf to be detected after preprocessing and short-time Fourier transform is input into the detection model, and the corn plant nitrogen content level detection result is output.

2. The method according to claim 1, wherein, In S1, the electrodes include AgCl patch electrodes and silver needles, the AgCl patch electrodes are used as the collection electrodes and BIAS electrodes, and the silver needles are used as reference electrodes; the collection electrodes are uniformly attached to the corn leaves and fixed to prevent displacement.

3. The method according to claim 1, wherein, In S1, the corn leaf surface potential information collection process includes: first, the corn is placed in a dark environment and rested for a specific time, then the potential data under dark conditions, light conditions and dark conditions again are collected in sequence, each leaf is collected multiple times, and each group of corn plants is collected multiple times to obtain multiple groups of data.

4. The method according to claim 1, wherein, In S2, when the improved wavelet threshold function is used for denoising, Daubechies5 wavelet is selected, the decomposition layer number is set to 5 layers, and the threshold selection rule is heursure; in the improved adaptive noise complete ensemble empirical mode decomposition process, the noise standard deviation ratio, the integration number and the maximum iteration number need to be set, white noise is added step by step and the residual signal is calculated, each IMF is extracted, and an IMF matrix is output.

5. The method according to claim 1, wherein, In S2, sample entropy is used to quantify the complexity of the signal, and the larger the sample entropy value, the higher the signal complexity. The sample entropy of the effective IMF is calculated to screen the components that need to be finely processed. After screening, the useful IMF components within a specific number range are retained, and the noise or redundant components are removed.

6. The method according to claim 1, wherein, In S3, the short-time Fourier transform retains the time domain characteristic information and frequency domain characteristic information of the plant surface electrical signal. The time domain characteristic information includes peak-to-peak value, maximum value, minimum value, and mean value. The frequency domain characteristic information includes power spectrum.

7. The method according to claim 1, wherein, In S3, the division ratio of the training set, validation set, and test set is determined according to the data volume and model training requirements to ensure that the time-frequency domain images corresponding to each nitrogen content level are reasonably distributed in each data set.

8. The method according to claim 1, wherein, In S4, the residual network extracts spatial features as the front-end module, extracts multi-level spatial features through the skip connection in the residual block, and the long short-term memory network is the back-end module, which receives the spatial features extracted by the residual network through the input gate, forget gate, and output gate, and learns the time dynamic changes and long-term dependencies.

9. The method according to claim 1, wherein, In S4, the residual network contains three residual stages, each stage contains two residual blocks; each residual block contains two convolution layers, each convolution layer is followed by BatchNorm and ReLU activation function; if the channel number of the input and output of the residual block is inconsistent or needs to be down-sampled, downsample is used to adjust the dimension of the residual path.

10. The method according to claim 1, wherein the method is characterized by, In S4, the output of the residual network is compressed in height by adaptive average pooling and input into the long short-term memory network with unchanged width; the long short-term memory network is set with hidden layer dimension and number of layers, and adopts dropout mechanism to prevent model overfitting; the full connection layer of the detection model maps the output of the long short-term memory network to the number of categories of corn nitrogen content level.

Images