A method and system for predicting tree root diameter and depth based on deep learning

Abstract

The present invention discloses a method and a prediction system for predicting the root diameter and depth of trees based on deep learning, belonging to the technical field of non-destructive testing. The method uses ground penetrating radar A-Scan data to predict the root radius and depth. Compared with B-Scan images, A-Scan data contains more information about the radius and depth and is more easily learned by the deep model. Therefore, using A-Scan data as the dataset is easier to train a root radius and depth prediction model with high accuracy and strong robustness; a deep learning model composed of a convolutional neural network and an attention mechanism is used to predict the root radius and depth. The attention mechanism is used to highlight the influence of key features on the model prediction, and the convolutional neural network is used to synthesize local features into global features to complete the prediction of the root radius and depth, which can effectively improve the accuracy of root radius and depth prediction.

Description

Technical Field

[0001] The present invention relates to a method and a system for predicting the root diameter and depth of trees based on deep learning, belonging to the technical field of non-destructive testing. Background Art

[0002] Root systems play a crucial role in plant individuals. They can provide water and nutrients required for plant growth, and at the same time, they also support and fix the plant body. However, root systems are more difficult to observe and sample compared to the branches or canopies of fruit trees. During the process of root system research, due to the difficulty of observing and sampling plant root systems under natural conditions, the research on root systems lags behind the above-ground parts of plants. The current root detection methods can be basically divided into two categories: destructive detection and non-destructive testing (NDT). Traditional destructive detection can accurately obtain root diameter information, but it is time-consuming, laborious, and may cause irreversible damage to the root system, which is not advisable in the root detection of fruit trees and ancient and famous trees.

[0003] Ground Penetrating Radar (GPR), as an emerging non-destructive testing technology, has the characteristics of accurate positioning, fast speed, flexible use, and high detection accuracy. Compared with other non-destructive testing methods (such as X-ray tomography, nuclear magnetic resonance method, acoustic method, and electrical resistivity tomography, etc.), it has the advantages of simple operation, convenient carrying, and repeatable measurement. Therefore, it is widely used in detecting unknown objects in shallow underground. GPR also plays an important role in the detection of tree root systems and tree trunks. The physical basis for GPR to detect root systems is the difference in water content between the root system and the surrounding soil, that is, the water content of the root system is significantly higher than that of the surrounding soil, resulting in a large difference in the relative dielectric constant between the two.

[0004] Currently, the research on plant root system detection using GPR mainly focuses on aspects such as root system morphology mapping, root biomass estimation, and root localization. However, using GPR to quantitatively predict the root radius of trees is still a challenging problem.

[0005] Most deep learning methods based on GPR use two-dimensional GPR data (B-Scan) as the basis for research. B-Scan data is a two-dimensional array composed of real numbers and is similar to digital images. Therefore, such methods can utilize the performance advantages of deep learning in image target recognition and classification. However, for the root radius and depth, the prediction accuracy is relatively low. For the prediction of the root radius, most use mathematical methods, which require labor costs and have low robustness. Summary of the Invention

[0006] To solve the problems of low accuracy and low robustness in current methods for predicting tree root radius and depth, the present invention provides a method and a prediction system for predicting tree root diameter and depth using deep learning based on ground penetrating radar A-Scan data. The ground penetrating radar A-Scan data is a one-dimensional array composed of real numbers, and the amplitude in the A-Scan data is related to soil water content, root depth, radius, etc.

[0007] The first object of the present invention is to provide a method for predicting tree root radius and depth based on deep learning, characterized in that the method uses a deep learning prediction model combining a convolutional neural network and an attention mechanism to predict the radius and depth of tree roots. The steps include:

[0008] Step 1: Use the open-source software gprMax to simulate and generate one-dimensional data (A-Scan data) of the ground penetrating radar collecting tree roots.

[0009] Step 2: Process the A-Scan data using the attention mechanism to calculate the feature similarity between the input data and the label data.

[0010] Step 3: Calculate the feature weight coefficient according to the feature similarity.

[0011] Step 4: Multiply the feature weight coefficient element by element with the numerical A-Scan data to obtain the A-Scan data with reallocated weights.

[0012] Step 5: Use a convolutional neural network to extract features from the A-Scan data with reallocated weights, and predict the tree root radius and depth based on the extracted features.

[0013] Optionally, the processing of the A-Scan data by the attention mechanism includes: calculating the feature similarity between the A-Scan data and the label data, setting the activation function as Sigmoid, importing the A-Scan data and the label into a multi-layer neural network MLP for calculation. The calculation formula for the feature similarity is:

[0014] sim i (inputs i ,labels)=MLP(inputs i ,labels) (1)

[0015] In the formula, sim i is the feature similarity, inputs i is the input A-Scan data, labels is the label value, and MLP(·) represents the MLP network.

[0016] Optionally, the calculation method of the feature weight coefficient includes:

[0017] The Softmax function is introduced for calculation. On the one hand, it completes the normalization operation, organizing the original feature similarity into a probability distribution where the sum of the weights of all elements is 1. On the other hand, it highlights the weights of important elements through the internal mechanism of Softmax. The calculation formula for the feature weight coefficient is:

[0018]

[0019] In the formula, α i is the feature weight coefficient, and k represents the total number of features of the A-Scan data.

[0020] Optionally, the process of calculating the A-Scan data with the reallocated weights includes:

[0021] Multiplying the feature weight coefficient by the input data to obtain the input data with reallocated weights:

[0022]

[0023] where x i is the original input data; is the input data with allocated weights;

[0024]

[0025] In the formula, Attention Value is the Attention value.

[0026] Optionally, using a convolutional neural network to extract features from the A-Scan data with the reallocated weights and predicting the root radius and depth based on the extracted features includes:

[0027] Using a convolutional layer to extract feature information. In the present invention, a 3×1 one-dimensional convolutional kernel is used, and convolution is performed with a stride of 1 to complete feature extraction; the pooling layer reduces the parameters of the network structure and keeps the features unchanged; the fully connected layer synthesizes the feature information extracted by the previous convolutional layer into the global features of the A-Scan data, and finally predicts the radius and depth of the root based on the global features.

[0028] Optionally, the activation function of the convolutional layer is the ReLU function, and the activation functions of the fully connected layers are all PReLU functions.

[0029] Optionally, the training process of the deep learning prediction model includes:

[0030] Step 1: Model the soil and the root, construct a three-dimensional space model, and set the space model parameters;

[0031] Step 2: Randomly generate a data set using the three-dimensional space model;

[0032] Step 3: Use the said data set to complete the parameter tuning and training of the said deep learning prediction model.

[0033] Optionally, the said Step 1 further includes:

[0034] Step 11: Use the open-source software gprMax that solves Maxwell's equations based on the finite-difference time-domain (FDTD) method to generate simulation data, set the size of the model and the spatial discretization step length, where the tree root is idealized as a cylinder, and the medium between the ground-penetrating radar antenna and the tree root is set as dry sand;

[0035] Step 12: Set the dielectric constant of the dry sand, the relative dielectric constant of the root system, the radius and depth of the tree root;

[0036] Step 13: Set the antenna used in the simulation data, and a single A-Scan generated by the simulation is used as one piece of data in the said data set.

[0037] Optionally, the said Step 3 further includes:

[0038] Step 31: Use the said data set for parameter tuning. The loss function of the model uses the mean squared error, and the Adam function is used to minimize the mean squared error between the actual value and the predicted value. Set the exponential decay factor of the biased first moment estimate, the exponential decay factor of the biased second moment estimate, and the learning rate of the prediction model;

[0039] Step 32: Use the model with tuned parameters for training: Set the number of model iterations and the size of each batch of samples. During the training process, record the loss value Loss and the coefficient of determination R for each iteration 2 。

[0040] The second object of the present invention is to provide a tree root radius and depth prediction system based on deep learning, characterized in that the said system includes: an input module, a tree root radius prediction module and a tree root depth prediction module respectively connected to the said input module;

[0041] Both the said tree root radius prediction module and the tree root depth prediction module include: an attention module, a convolutional layer, a pooling layer, a fully connected layer and an output module connected in sequence;

[0042] The said prediction system takes the numerical A-Scan data as the input, highlights important feature information through the said attention module, then uses the said convolutional layer to extract feature information, the pooling layer reduces the parameters of the network structure and keeps the features unchanged, improves the generalization ability of the model, and finally synthesizes the features extracted by the said convolutional layer into the global features of the A-Scan data through the said fully connected layer, completes the prediction of the root system radius or depth, and outputs the prediction result through the said output module.

[0043] The beneficial effects of the present invention are as follows:

[0044] The present invention predicts the root radius and depth through the ground penetrating radar A-Scan data. Compared with the B-Scan image, the A-Scan data contains more information about the radius and depth and is more easily learned by the depth model. Therefore, using the A-Scan data as the data set is more likely to train a root radius and depth prediction model with high accuracy and strong robustness;

[0045] The present invention uses a deep learning model composed of a convolutional neural network and an attention mechanism to predict the root radius and depth. The attention mechanism is used to highlight the influence of key features on the model prediction, and the convolutional neural network synthesizes local features into global features to complete the prediction of the root radius and depth. The maximum errors of the radius and depth predictions are 0.0029m and 0.0112m, respectively, which are far better than other prediction schemes using common regression models. Therefore, the prediction method and system of the present invention can effectively improve the accuracy of root radius and depth prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0047] Figure 1 is the flowchart of the present invention.

[0048] Figure 2 is the longitudinal sectional view of the three-dimensional model constructed by the present invention to generate simulation data.

[0049] Figure 3 is the radius prediction model of the present invention; where Attention represents the attention module; Conv1D,n represents a convolutional kernel with size n; MaxPooling1D,3 represents a max pooling layer with a stride of 3; Dense,m represents a fully connected layer with m nodes; Output is the predicted value.

[0050] Figure 4 is the depth prediction model of the present invention.

[0051] Figure 5 is the attention module in the two prediction models of the present invention, where α i is the feature weight coefficient, x i is the original input data, x newi is the input data with weights assigned, and "⊙" is element-wise multiplication.

[0052] Figure 6 It is the prediction result diagram of the root radius of the present invention.

[0053] Figure 7 It is the prediction result diagram of the root depth of the present invention. Specific embodiments

[0054] To make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0055] Example 1:

[0056] This embodiment provides a method for predicting the root radius and depth based on deep learning. The method uses a deep learning prediction model combining a convolutional neural network and an attention mechanism to predict the root radius and depth. The steps include:

[0057] Step 1: Use the open-source software gprMax to simulate and generate one-dimensional data (A-Scan data) of the ground-penetrating radar collecting tree roots.

[0058] Step 2: Process the A-Scan data using the attention mechanism to calculate the feature similarity between the input data and the label data.

[0059] Step 3: Calculate the feature weight coefficient according to the feature similarity.

[0060] Step 4: Multiply the feature weight coefficient element by element with the numerical A-Scan data to obtain the A-Scan data with reallocated weights.

[0061] Step 5: Use a convolutional neural network to extract features from the A-Scan data with reallocated weights, and predict the root radius and depth according to the extracted features.

[0062] Example 2:

[0063] This embodiment provides a method for predicting the root radius and depth based on deep learning. The method includes:

[0064] 1. Training the prediction model

[0065] Step 1: Model the soil and tree roots to construct a three-dimensional space model and set the space model parameters. Specifically, it includes:

[0066] Step 11: Use the open-source software gprMax that solves Maxwell's equations based on the finite-difference time-domain (FDTD) method to generate simulation data. The size of the model is 0.5m × 0.3m × 0.6m, and the spatial discretization step size is 0.001m. Among them, the tree root is idealized as a cylinder, and the medium between the ground-penetrating radar antenna and the tree root is set as dry sand.

[0067] Step 12: Set a random number for the dielectric constant of dry sand ranging from 3 to 6, a random number for the relative dielectric constant of the root system ranging from 15 to 21, a random number for the root radius ranging from 0.005 to 0.03 m, and a random number for the depth ranging from 0 to 0.3 m;

[0068] Step 13: The antenna used in the simulation data is a 400 MHz commercial GPR antenna, and a single A-Scan generated by the simulation is used as one piece of data in the dataset.

[0069] Step 2: Use the three-dimensional space model to randomly generate a dataset, with a total of 3100 A-Scan data, and each piece of data has 3000 features;

[0070] Step 3: Use the dataset to complete the parameter tuning and training of the deep learning prediction model, specifically including:

[0071] Step 31: Use the dataset for parameter tuning. The loss function of the model uses the mean squared error, and the Adam function is used to minimize the mean squared error between the actual value and the predicted value. Among them, the exponential decay factor of the biased first-order moment estimate is 0.9, the exponential decay factor of the biased second-order moment estimate is 0.999, and the learning rate of the prediction model is 0.0003;

[0072] Step 32: Use the model with tuned parameters for training: The model iterates 2000 times, and the number of samples in each batch is 64. During the training process, record the loss value Loss and the coefficient of determination R for each iteration 2 .

[0073] II. Predict the root radius and depth through the trained prediction model

[0074] Step 1: Use the open-source software gprMax to simulate and generate one-dimensional data (A-Scan data) of the ground-penetrating radar collecting tree roots;

[0075] Step 2: Use the attention mechanism to process the A-Scan data and calculate the feature similarity between the input data and the label data;

[0076] The attention mechanism processes the A-Scan data as follows: Calculate the feature similarity between the A-Scan data and the label data, set the activation function to Sigmoid, import the A-Scan data and the label into the multi-layer neural network MLP for calculation, and the calculation formula for the feature similarity is:

[0077] sim i (inputs i ,labels) = MLP(inputs i ,labels) (1)

[0078] where sim i is the feature similarity, inputs i is the input A-Scan data, labels is the label value, and MLP(·) represents the MLP network.

[0079] Step 3: Calculate the feature weight coefficient according to the feature similarity;

[0080] Introduce the Softmax function for calculation. On the one hand, it completes the normalization operation, sorting the original feature similarity into a probability distribution where the sum of the weights of all elements is 1. On the other hand, it highlights the weights of important elements through the internal mechanism of Softmax. The calculation formula for the feature weight coefficient is:

[0081]

[0082] where α i is the feature weight coefficient, k represents the total number of features of the A-Scan data, and in this embodiment, k is 3000.

[0083] Step 4: Multiply the feature weight coefficient by each element of the numerical A-Scan data to obtain the A-Scan data with redistributed weights;

[0084] Multiply the feature weight coefficient by the input data to obtain the input data with redistributed weights:

[0085]

[0086] where, x i is the original input data; is the input data with weights assigned;

[0087] Perform weighted summation on the feature weight coefficient:

[0088]

[0089] where Attention Value is the Attention value.

[0090] Step 5: Use a convolutional neural network to extract features from the A-Scan data with redistributed weights, and predict the root radius and depth based on the extracted features;

[0091] Use the convolutional layer to extract feature information, the pooling layer reduces the parameters of the network structure and keeps the features unchanged, and the fully connected layer synthesizes the feature information extracted by the previous convolutional layer into the global features of the A-Scan data. Finally, predict the radius and depth of the root based on the global features.

[0092] Example 4:

[0093] This embodiment provides a root radius and depth prediction system based on deep learning, including: an input module, a root radius prediction module and a root depth prediction module respectively connected to the input module;

[0094] Both the root radius prediction module and the root depth prediction module include: an attention module, a convolutional layer, a pooling layer, a fully connected layer and an output module connected in sequence;

[0095] The root radius prediction module is used to predict the root system radius, and the root depth prediction module is used to predict the root system depth. They are deep learning models based on the attention mechanism and convolutional neural network, taking numerical A-Scan data as input. First, the attention module highlights important feature information. Then, the convolutional layer is used to extract feature information. The pooling layer can reduce the parameters of the network structure and keep the features unchanged, improving the generalization ability of the model. Finally, the fully connected layer synthesizes the local features learned by the previous convolutional layer into the global features of the A-Scan data. Thus, the constructed model can automatically learn and obtain the main features of the root system radius or depth to complete the prediction of the root system radius or depth.

[0096] The activation function can make the neural network approximate any non-linear function and enhance the learning ability of the model. Therefore, the selection of the activation function in the deep neural network has an important impact on the model performance. In the prediction model of this embodiment, the activation function of the convolutional layer is ReLU, and the activation function of the fully connected layer is PReLU. PReLU can solve the problems of model gradient disappearance and gradient explosion. The activation function of the last layer is linear, which is used for continuous value prediction.

[0097] The attention module (Attention) is as Figure 5 shown. The numerical A-Scan data is used to calculate the feature weight coefficients through two fully connected layers composed of 3000 nodes. The weight coefficients are multiplied by the numerical A-Scan data element by element to obtain the A-Scan data with reallocated weights, so as to achieve the purpose of highlighting key features.

[0098] To further prove the beneficial effects that the present invention can achieve, the following experiments are carried out:

[0099] The dataset is generated using the open-source software gprMax based on the finite-difference time-domain method to solve Maxwell's equations. Figure 2 It represents the longitudinal sectional view of the spatial model constructed by generating one A-Scan data, where the size of the model is 0.5m×0.3m×0.6m, and the GSSI 400MHz antenna is placed on the ground, directly above the measured root system.

[0100] The measured single A-Scan serves as one piece of data in the dataset, without the need for a complete B-Scan data, which can reduce the computational requirements for generating simulation data. The tree root is modeled as a cylinder and kept parallel to the ground penetrating radar antenna.

[0101] The dataset consists of 3100 A-Scan data, and each piece of data has 3000 features. The dataset is divided into three parts, with 65% of the data for training, 15% for validation, and 20% for testing. All training and validation are carried out within the range of a radius from 0.005 to 0.03 m and a depth from 0 to 0.3 m. The model is iterated (Epoch) 2000 times, and the number of samples per batch (Batch Size) is 64.

[0102] Figure 6 and Figure 7 Shows the prediction effect of the model on the simulation data. The 600 pieces of simulation data not included in the training set are used for prediction. The abscissa is the true value, that is, the label value, and the ordinate is the predicted value of the model. One point represents one piece of data. When the predicted value is closer to the true value, the data point will be closer to the line y = x, which means the prediction effect and fitting degree of the model are better. It can be seen from the figure that the model has a good fitting degree, and the fitting degree of the root depth prediction model is the best.

[0103] Using the maximum error and the coefficient of determination as evaluation parameters, this embodiment is compared with other common regression models, and the comparison results are shown in Table 1.

[0104] Table 1 Comparison of the performance of different models

[0105]

[0106]

[0107] It can be seen from Table 1 that under the condition of using the same training set and test set, the maximum error of the model radius prediction of the present invention is 0.0029 m, and the coefficient of determination R 2 value is 0.990, which can accurately predict the root radius of the tree. Compared with other models, it has obvious advantages both in terms of the maximum error and the coefficient of determination.

[0108] The maximum error of the depth prediction is 0.0112 m, and R 2The value is 0.996. Although the coefficient of determination is only slightly higher than that of other models, the maximum error is significantly better than that of other models, indicating that the model has better prediction accuracy when the fitting effects are not much different. It can be seen from this that the prediction scheme in the present invention is feasible and effective for predicting the root radius and depth. At the same time, through the trained model, the rapid prediction of the root radius and depth can be realized, saving a large amount of labor costs.

[0109] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0110] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the radius and depth of tree roots based on deep learning, characterized in that, The method uses a deep learning prediction model that combines a convolutional neural network and an attention mechanism to predict the radius and depth of tree roots. The steps include: Step 1: Use the open-source software gprMax to simulate and generate one-dimensional data of ground-penetrating radar for collecting tree roots: A-Scan data; Step 2: Process the A-Scan data using the attention mechanism to calculate the feature similarity between the input data and the label data; Step 3: Calculate the feature weight coefficient based on the feature similarity; Step 4: Multiply each element of the feature weight coefficient and the numerical A-Scan data to obtain the A-Scan data with reallocated weights; Step 5: Use a convolutional neural network to extract features from the A-Scan data with reallocated weights, and predict the radius and depth of the tree roots based on the extracted features; The training process of the deep learning prediction model includes: Step 1: Model the soil and tree roots, construct a three-dimensional space model, and set the space model parameters; Step 2: Randomly generate a dataset using the three-dimensional space model; Step 3: Use the dataset to complete the parameter tuning and training of the deep learning prediction model.

2. The method according to claim 1, characterized in that, The attention mechanism processes the A-Scan data by calculating the feature similarity between the A-Scan data and the label data. The activation function is set to Sigmoid. The A-Scan data and the label are imported into a multi-layer neural network MLP for calculation. The calculation formula for the feature similarity is: sim i (inputs i , labels) = MLP(inputs i , labels) (1) where sim i is the feature similarity, inputs i is the input A-Scan data, labels is the label value, and MLP(·) represents the MLP network.

3. The method according to claim 2, characterized in that, The calculation method of the feature weight coefficient includes: Introduce the Softmax function for calculation. On the one hand, complete the normalization operation to organize the original feature similarity into a probability distribution where the sum of the weights of all elements is 1. On the other hand, highlight the weights of important elements through the internal mechanism of Softmax. The calculation formula for the feature weight coefficient is: where α i is the feature weight coefficient, and k represents the total number of features of the A-Scan data.

4. The method according to claim 3, characterized in that, The process of calculating the A-Scan data with reallocated weights includes: Multiply the feature weight coefficient and the input data to obtain the input data with reallocated weights: where x i is the original input data; is the input data for weight assignment; Perform weighted summation on the feature weight coefficient: In the formula, Attention Value is the Attention value.

5. The method according to claim 4, wherein Using a convolutional neural network to extract features from the A-Scan data with reallocated weights and predicting the radius and depth of the tree roots based on the extracted features includes: Use a convolutional layer to extract feature information. The convolutional layer uses a 3×1 one-dimensional convolutional kernel and performs convolution in a step size of 1 to complete feature extraction; the pooling layer reduces the parameters of the network structure and keeps the features unchanged; the fully connected layer synthesizes the feature information extracted by the convolutional layer into the global features of the A-Scan data, and finally predicts the radius and depth of the tree roots based on the global features.

6. The method according to claim 5, wherein The activation function of the convolutional layer is the ReLU function, and the activation functions of the fully connected layer are all PReLU functions.

7. The method according to claim 6, wherein Step 1 also includes: Step 11: Use the open-source software gprMax that solves Maxwell's equations based on the finite-difference time-domain (FDTD) method to generate simulation data. Set the size of the model and the spatial discretization step length, where the tree root is idealized as a cylinder, and the medium between the ground-penetrating radar antenna and the tree root is set as dry sand; Step 12: Set the dielectric constant of the dry sand, the relative dielectric constant of the root system, the radius and depth of the tree root; Step 13: Set the antenna used in the simulation data, and a single A-Scan generated by the simulation is used as one piece of data in the dataset.

8. The method according to claim 7, wherein Step 3 further includes: Step 31: Use the dataset for parameter tuning. The loss function of the model uses the mean squared error, and the Adam function is used to minimize the mean squared error between the actual value and the predicted value. Set the exponential decay factor of the biased first moment estimate, the exponential decay factor of the biased second moment estimate, and the learning rate of the prediction model; Step 32: Train using the model with adjusted parameters: Set the number of model iterations and the size of each batch of samples. During the training process, record the loss value Loss and the coefficient of determination R for each iteration 2 .

9. A root radius and depth prediction system based on deep learning, applied to the method according to any one of claims 1-8, characterized in that, The system includes: an input module, a tree root radius prediction module and a tree root depth prediction module respectively connected to the input module; Both the tree root radius prediction module and the tree root depth prediction module include: an attention module, a convolutional layer, a pooling layer, a fully connected layer and an output module connected in sequence; The prediction system takes the numerical A-Scan data as input, highlights important feature information through the attention module, then uses the convolutional layer to extract feature information, the pooling layer reduces the parameters of the network structure and keeps the features unchanged, improving the generalization ability of the model. Finally, the fully connected layer synthesizes the features extracted by the convolutional layer into the global features of the A-Scan data, completes the prediction of the root system radius or depth, and outputs the prediction result through the output module.

Images