Slope soil fertility prediction method

Abstract

The invention relates to the technical field of data prediction. The slope soil fertility prediction method comprises the following steps: S1, constructing a slope soil fertility prediction training set; s2, performing data sample generation on the marked data by using a dynamically selected SMOTE algorithm to realize data expansion of a slope soil fertility prediction training set; s3, inputting the expanded slope soil fertility prediction training set into a feature dimension reduction model to train the feature dimension reduction model; s4, inputting the slope soil fertility prediction training set after dimension reduction into a classifier model to train the classifier model; and S5, constructing a slope soil fertility prediction data set, and performing data processing on the slope soil fertility prediction data set by using the trained model to obtain a slope soil fertility prediction result. The method has the beneficial effect that the accuracy of fertility prediction of the slope soil is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of data prediction, and in particular relates to a method for predicting slope soil fertility. Background Art

[0002] Slope soil fertility is a crucial factor in agriculture, ecological restoration, and soil and water conservation. Its assessment is of great significance for rational planning of resource utilization, optimization of planting strategies, and protection of the ecological environment. However, existing slope soil fertility assessment methods have many deficiencies in data collection, processing, and prediction analysis, making it difficult to meet the needs of practical applications.

[0003] The Chinese invention patent with publication number CN118675306B proposes an intelligent early warning method and system for slope geological disasters based on artificial intelligence; the Chinese invention patent with publication number CN117852717B proposes a typhoon disaster prediction method based on genetic algorithm optimized back propagation neural network. By establishing a typhoon disaster prediction model based on classification decision tree, the impact of typhoon disasters on the environment can be predicted more accurately, thereby scientifically guiding disaster prevention and mitigation work; the above invention methods are innovative, but at the same time there are the following problems that still need to be further solved: 1. The traditional SMOTE algorithm generates samples 1. It is single and cannot dynamically adapt to the local density and distribution of data, resulting in poor generalization ability of the model, insufficient description of the feature distribution of minority class samples, and inconsistency between the generated samples and the true distribution; 2. Traditional feature dimensionality reduction methods perform poorly in dealing with outliers and boundary effects, which easily leads to loss of feature information after dimensionality reduction. The fixed setting or simple attenuation of the learning rate cannot adapt to complex data distribution, resulting in slow convergence; 3. Traditional classifiers are prone to fall into local optimality during the optimization process and lack the ability to explore new parameter areas, resulting in limited classification performance. The classification model training strategy lacks flexibility and is difficult to efficiently adapt to diverse features. Summary of the invention

[0004] In order to address the deficiencies in the prior art, the present invention provides a method for predicting slope soil fertility, which uses a slope soil fertility model trained by a SMOTE algorithm based on dynamic selection and a boundary smoothing autoencoder and a probabilistic neural network based on the quantum tunneling effect to predict new slope soil fertility sample data, so as to improve the accuracy of slope soil fertility prediction.

[0005] To achieve the above object, the present invention is implemented through the following technical scheme: a slope soil fertility prediction method, comprising the following steps: S1. Constructing a slope soil fertility prediction training set, including collecting data on the spectral reflectance value of the slope soil and the thermal radiation value of the soil surface, and annotating the collected data; S2. Generate data samples using the dynamically selected SMOTE algorithm for the labeled data to achieve data expansion of the slope soil fertility prediction training set; S3, inputting the expanded slope soil fertility prediction training set into a feature dimensionality reduction model to train the feature dimensionality reduction model, wherein the feature dimensionality reduction model is formed by a boundary smoothing autoencoder; S4, inputting the slope soil fertility prediction training set after dimensionality reduction into a classifier model to train the classifier model, wherein the classifier model is constructed using a probabilistic neural network of quantum tunneling effect; S5. Construct a slope soil fertility prediction data set, use the trained model to process the data of the slope soil fertility prediction data set, and obtain the slope soil fertility prediction result.

[0006] Preferably, the sources of data collection are soil spectral data and topographic data; the annotation method adopts manual annotation, and the annotation categories are specifically divided into: low fertility, medium fertility, and high fertility according to the soil fertility level; wherein, the content of data collection includes the spectral reflectance value of the soil and the thermal radiation value of the soil surface, the storage format of the collected data is a structured data file, and the specific data format is a multidimensional array, wherein the pixel resolution of the spectral data is 1 meter / pixel, and the resolution of the infrared image is 0.5 meter / pixel.

[0007] Preferably, the traditional SMOTE algorithm balances the data set by generating synthetic versions of minority class samples, but this method cannot fully capture the complex structure of the data. The present invention improves the traditional SMOTE algorithm to generate a dynamically selected SMOTE algorithm. The dynamic selection refers to dynamically adjusting the generation strategy of the SMOTE algorithm synthetic samples according to the local density and diversity of the data, thereby more accurately simulating the characteristic distribution of minority class samples.

[0008] Preferably, the method of generating data samples using the dynamically selected SMOTE algorithm is: S201, initialize the target number of data expansion and the initial synthetic sample generation ratio; specifically, set is the target number of synthetic samples set according to the degree of data imbalance, The initial synthetic sample generation ratio is used to control the initial proportion of data generated by different categories; S202, evaluating the local density of the current slope soil fertility prediction training set, calculating the sample density in the corresponding neighborhood for the minority class data samples in the training set, and using the sample density information for subsequent dynamic sample generation decisions; S203. Dynamically adjust the synthetic sample generation ratio of each data point according to the evaluated local density. The method of dynamically adjusting the synthetic sample generation ratio of each data point is as follows: ; In the formula, represents the synthetic sample generation ratio of the adjusted i-th sample, represents the initial synthetic sample generation ratio, represents the local density of the i-th sample, Represents the adjustment strength coefficient, and the mathematical model is: ; In the formula, is an adjustable scaling parameter that controls the sensitivity of adjusting the intensity coefficient. is the local inhomogeneity index of the current data set, and the mathematical model is: ; In the formula, is the set of all minority class samples, is the i-th sample The set of nearest neighbor samples, and are the feature vectors of the i-th sample and the j-th sample respectively, is the L2 norm; S204, adopting the interpolation generation method of dual-trajectory learning, selecting strategic data point pairs, using the optimization process to generate a learning trajectory between two data points, interpolating on the two learning trajectories to generate new sample points; the method of using dual-trajectory learning to interpolate and generate synthetic samples is expressed as: ; In the formula, is the i-th synthetic sample, is the i-th synthetic sample point, is the first interpolation coefficient, is the second interpolation coefficient, , , and are two pairs of random synthetic sample points; the interpolation coefficient is dynamically adjusted based on the correlation and spatial structure between the data points, and the calculation method of the first interpolation coefficient is expressed as: ; In the formula, It is a parameter that controls the distance sensitivity of the item and is used to adjust the distribution of the interpolation coefficients; is the local geometric sensitivity index of the i-th synthetic sample; Furthermore, a dynamic adjustment mechanism based on local geometric and topological data structure is adopted to dynamically adjust the synthetic sample generation strategy by calculating the topological relationship and geometric features between data points to enhance the model's adaptability to the potential structure of the data. Specifically, the local geometric sensitivity index of the synthetic sample is used to evaluate the structural complexity and geometric distribution of the data point in its neighborhood. The mathematical model is: ; In the formula, is the i-th sample A set of nearest neighbor samples; The second interpolation coefficient calculation method is expressed as: ; S205, merging the generated synthetic samples with the original slope soil fertility prediction training set to form a new and more balanced slope soil fertility prediction training set, expressed as: ; In the formula, A new training dataset, including the original dataset and all generated synthetic samples; is the original data set, is the i-th synthetic sample, Set to 50% of the number of samples in the original dataset.

[0009] Preferably, the synthetic sample generation ratio of each data point is dynamically adjusted. For those sample points located in low-density areas, the number of synthetic samples generated can be increased to strengthen the model's learning of low-density areas. Relatively speaking, the generation of synthetic samples in high-density areas is reduced to avoid overfitting. At the same time, the interpolation generation method of dual-trajectory learning can ensure that the generation of trajectories can both reflect the main trends of the data and reveal potential nonlinear relationships.

[0010] Preferably, the expanded data is input into a feature dimensionality reduction model for training. The present invention adopts an autoencoder based on boundary smoothing as the feature dimensionality reduction model. The autoencoder includes an encoder and a decoder. The encoder compresses the input data to a low-dimensional representation, and the decoder reconstructs the low-dimensional representation into a feature vector that is as similar as possible to the input data.

[0011] Preferably, the training method for the feature dimensionality reduction model is: S301, initializing the parameters of the autoencoder; S302. During the training process of the autoencoder, the expanded slope soil fertility prediction training set data is processed by forward propagation, and then the autoencoder parameters are updated by the gradient descent method through error back propagation; wherein, the weight coefficient of the autoencoder is used to adjust the contribution of each dimensional feature in the error, and these weights are dynamically allocated according to the importance of the feature, and the calculation method is: ; In the formula, is the characteristic variance of the input data of the i-th dimension autoencoder, is the feature of the input data of the i-th dimension autoencoder, is a constant with a value of 0.001. is the weight coefficient of the i-th dimension autoencoder; S303, adopting an automatic boundary smoothing strategy to enhance the robustness of the model by dynamically adjusting the boundary value, specifically through the smoothing function The output of the encoder is smoothed, and the mathematical model is: ; In the formula, is the output of the encoder; in, ; In the formula, is the output reconstructed by the decoder, It is an automatically adjusted threshold parameter used to control the degree of smoothing; In order to smooth the intensity parameters, an adaptive method based on the local density of data is used for design, and the mathematical model is: ; In the formula, is the kernel density estimation function, mean is the averaging function, For the neighborhood, calculate Kernel density estimation in the corresponding neighborhood; S304, advancing the iteration process through the loss function, and updating the autoencoder parameters; S305, using the convex hull convergence strategy to optimize the training process of the autoencoder, by adjusting the learning step size and weight update strategy, the network converges quickly, and the learning rate is dynamically adjusted according to the data distribution characteristics in the current iteration as follows: ; In the formula, is the initial autoencoder learning rate, is the learning rate of the autoencoder at the tth iteration, is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the t-th iteration, is the diameter of the initial convex hull; S306: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0012] Preferably, the reduced-dimensional data is input into a classifier to train the classifier model. The present invention adopts a probabilistic neural network based on the quantum tunneling effect as the classifier model. The quantum tunneling effect simulates the behavior of quantum particles crossing an energy barrier, so that when the probability neural network algorithm is updating parameters, it can tunnel to a new position outside the barrier, thereby improving the stability and efficiency of the probabilistic neural network training process.

[0013] Preferably, the specific method of using the probabilistic neural network of the quantum tunneling effect to form a classifier model training is: S401, randomly initialize the model parameters of the probabilistic neural network, including: the initial weights of the probabilistic neural network and biases for probabilistic neural networks ; S402, inputting the slope soil fertility prediction training set data features after dimension reduction, forward propagating through the probabilistic neural network layer, and converting the node output of each layer through the activation function to provide input for the next layer until the final classification probability output is generated; S403, using a cross entropy loss function to calculate the difference between the actual output and the expected output of the probabilistic neural network; S404, using the loss function to calculate the gradient of each parameter of the probabilistic neural network, and then using the gradient update strategy simulated by the quantum tunneling effect to evaluate the potential barriers in the current gradient direction, and decide whether to perform tunneling, that is, skip the local minimum point to reach a new possible area to avoid the local optimum. The parameter gradient of the probabilistic neural network is , the way to adjust the parameter gradient of the probabilistic neural network using the quantum tunneling effect is: ; In the formula, are the parameters of the updated probabilistic neural network, is the quantum barrier parameter, To control the parameters of the intensity of quantum behavior, the gradient jumps to a new position outside the potential barrier, simulating the tunneling phenomenon; S405. According to the parameter gradient adjusted by the quantum tunneling effect algorithm, the parameters of the probabilistic neural network are updated. The weight and bias of the probabilistic neural network are updated in the following manner: ; ; In the formula, is the weight of the updated probabilistic neural network, is the bias of the updated probabilistic neural network, is the weight of the probabilistic neural network before updating, is the bias of the probabilistic neural network before updating; is the learning rate of the probabilistic neural network for the tth iteration, controls the step size of the gradient descent, and dynamically adjusts the learning rate based on the gradient change in the current iteration. The mathematical model is: ; In the formula, is the learning rate of the t-th iteration probabilistic neural network, To adjust the sensitivity parameter, control the speed of learning rate change; and They are the gradient of the loss function at the t-th iteration and the gradient of the loss function at the t-1-th iteration respectively; S406, repeat the above steps until the preset stop iteration condition is met to complete the classifier model training.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: the dynamic smoothing and convex hull convergence strategies are adopted to enhance the processing capability of the dimensionality reduction model for boundary anomalies and local changes, and the quantum tunneling effect enables the classifier to escape from the local optimum more easily, thereby improving the training efficiency and classification performance of slope soil fertility prediction; the dynamic SMOTE method expands the data while retaining the data distribution characteristics, thereby reducing the workload of slope soil fertility data collection and annotation; the feature dimensionality reduction combined with boundary smoothing reduces the demand for computing resources for high-dimensional data, accelerates model training, and significantly enhances the model's prediction performance for complex and nonlinear slope soil fertility data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of the slope soil fertility prediction method.

[0016] Figure 2 This is a training flow chart of the autoencoder algorithm based on boundary smoothing of the present invention. DETAILED DESCRIPTION

[0017] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0018] like Figure 1 As shown, the present invention provides a slope soil fertility prediction method, including S1 to S5.

[0019] S1. Constructing a slope soil fertility prediction training set, including collecting data on the spectral reflectance value of the slope soil and the thermal radiation value of the soil surface, and labeling the collected data.

[0020] Specifically, the data collection method for the slope soil fertility prediction training set adopts low-altitude flight of drones, and uses multispectral sensors and infrared sensors to obtain the surface spectral characteristics and thermal infrared characteristics of the soil, including the spectral reflectance value of the slope soil and the thermal radiation value of the soil surface. The collected data are annotated manually, and the annotated categories include: low fertility, medium fertility, and high fertility. The data is stored in a structured data file, and the specific data format is a multidimensional array, where the pixel resolution of the spectral data is 1 meter / pixel, and the resolution of the infrared image is 0.5 meter / pixel.

[0021] S2. The data samples are generated by using the dynamically selected SMOTE algorithm for the labeled data to achieve data expansion of the slope soil fertility prediction training set.

[0022] Specifically, the method of generating data samples using the dynamically selected SMOTE algorithm is as follows: S201, initialize the target number of data expansion and the initial synthetic sample generation ratio; specifically, set is the target number of synthetic samples set according to the degree of data imbalance, The initial synthetic sample generation ratio is used to control the initial proportion of data generated by different categories; S202, evaluate the local density of the current slope soil fertility prediction training set, calculate the sample density in the corresponding neighborhood for the minority class data samples in the training set, and use the sample density information for subsequent dynamic sample generation decisions; the calculation method in the implementation process is: ; In the formula, is the local density of the ith sample, is the number of nearest neighbors, is the i-th sample The set of nearest neighbor samples, and is the feature vector of the i-th sample and the j-th sample; S203. Dynamically adjust the synthetic sample generation ratio of each data point according to the evaluated local density. The calculation method during the implementation is: ; In the formula, represents the synthetic sample generation ratio of the adjusted i-th sample, represents the initial synthetic sample generation ratio, represents the local density of the i-th sample, Represents the adjustment strength coefficient, and the mathematical model is: ; In the formula, is an adjustable scaling parameter that controls the sensitivity of adjusting the intensity coefficient. is the local inhomogeneity index of the current data set, and the mathematical model is: ; In the formula, is the set of all minority class samples, is the i-th sample The set of nearest neighbor samples, and are the feature vectors of the i-th sample and the j-th sample respectively, is the L2 norm; S204, adopting the interpolation generation method of dual-trajectory learning, selecting strategic data point pairs, using the optimization process to generate a learning trajectory between two data points, interpolating on the two learning trajectories to generate new sample points; the method of using dual-trajectory learning to interpolate and generate synthetic samples is expressed as: ; In the formula, is the i-th synthetic sample, is the i-th synthetic sample point, is the first interpolation coefficient, is the second interpolation coefficient, , , and are two pairs of random synthetic sample points; the interpolation coefficient is dynamically adjusted based on the correlation and spatial structure between the data points, and the calculation method of the first interpolation coefficient is expressed as: ; In the formula, It is a parameter that controls the distance sensitivity of the item and is used to adjust the distribution of the interpolation coefficients; is the local geometric sensitivity index of the i-th synthetic sample; Furthermore, a dynamic adjustment mechanism based on local geometric and topological data structure is adopted to dynamically adjust the synthetic sample generation strategy by calculating the topological relationship and geometric features between data points to enhance the model's adaptability to the potential structure of the data. Specifically, the local geometric sensitivity index of the synthetic sample is used to evaluate the structural complexity and geometric distribution of the data point in its neighborhood. The mathematical model is: ; In the formula, is the i-th sample A set of nearest neighbor samples; The second interpolation coefficient calculation method is expressed as: ; S205, merging the generated synthetic samples with the original slope soil fertility prediction training set to form a new and more balanced slope soil fertility prediction training set, expressed as: ; In the formula, A new training dataset, including the original dataset and all generated synthetic samples; is the original data set, is the i-th synthetic sample, Set to 50% of the number of samples in the original dataset.

[0023] Specifically, the data augmentation model is used to increase the number of samples. During the implementation process, the original collected samples are set to 800, and the data augmentation model is used to expand and generate 400 samples. The expanded data set contains 1,200 samples.

[0024] S3. Inputting the expanded slope soil fertility prediction training set into a feature dimensionality reduction model to train the feature dimensionality reduction model, wherein the feature dimensionality reduction model is constructed using a boundary smoothing autoencoder.

[0025] Specifically, the training process of the autoencoder algorithm based on boundary smoothing is as follows: Figure 2 As shown, during the implementation process, the training method for the feature dimensionality reduction model is: S301, initialize the parameters of the autoencoder, including the encoder and the decoder, wherein the encoder parameter initialization method is expressed as: ; In the formula, is the weight of the encoder, is a random function, is the dimension of the encoder input layer, is the dimension of the middle layer of the encoder; ; In the formula, is the encoder bias, is a function that generates a vector of all zeros; The decoder parameter initialization method is expressed as: ; In the formula, is the weight of the decoder, is a random function, is the dimension of the encoder input layer, is the dimension of the middle layer of the encoder, is the dimension of the encoder output layer; ; In the formula, is the bias of the decoder; S302. During the training process of the autoencoder, the expanded slope soil fertility prediction training set data is processed by forward propagation, and then the autoencoder parameters are updated by the gradient descent method through error back propagation; wherein, the weight coefficient of the autoencoder is used to adjust the contribution of each dimensional feature in the error, and these weights are dynamically allocated according to the importance of the feature. The calculation method during the implementation process is: ; In the formula, is the characteristic variance of the input data of the i-th dimension autoencoder, is the feature of the input data of the i-th dimension autoencoder, is a constant with a value of 0.001. is the weight coefficient of the i-th dimension autoencoder; S303, adopting an automatic boundary smoothing strategy to enhance the robustness of the model by dynamically adjusting the boundary value, specifically through the smoothing function The output of the encoder is smoothed, and the mathematical model is: ; In the formula, is the output of the encoder; in, ; In the formula, is the output reconstructed by the decoder, It is an automatically adjusted threshold parameter used to control the degree of smoothing; In order to smooth the intensity parameters, an adaptive method based on the local density of data is used for design, and the mathematical model is: ; In the formula, is the kernel density estimation function, mean is the averaging function, For the neighborhood, calculate Kernel density estimation in the corresponding neighborhood; S304, advancing the iteration process through the loss function, and updating the autoencoder parameters; S305, using the convex hull convergence strategy to optimize the training process of the autoencoder, by adjusting the learning step size and weight update strategy, the network converges quickly, and the learning rate is dynamically adjusted according to the data distribution characteristics in the current iteration as follows: ; In the formula, is the initial autoencoder learning rate, is the learning rate of the autoencoder at the tth iteration, is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the t-th iteration, is the diameter of the initial convex hull; S306: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0026] Specifically, the preset stopping condition during the implementation process is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0027] S4. Inputting the slope soil fertility prediction training set after dimensionality reduction into the classifier model to train the classifier model, wherein the classifier model is constructed using a probabilistic neural network of quantum tunneling effect.

[0028] Specifically, the specific method of using the probabilistic neural network with quantum tunneling effect to form a classifier model training process is: S401, randomly initialize the model parameters of the probabilistic neural network, including: the initial weights of the probabilistic neural network and biases for probabilistic neural networks ; S402, input the slope soil fertility prediction training set data features after dimension reduction, forward propagate through the probabilistic neural network layer, and the node output of each layer is converted through the activation function to provide input for the next layer until the final classification probability output is generated; set is the input feature vector of the probabilistic neural network, then the calculation method of the output of the first layer is expressed as: ; In the formula, is the output of the first layer of the probabilistic neural network, is the Sigmoid activation function, and are the weights and biases of the first layer of the probabilistic neural network; S403, using the cross entropy loss function to calculate the difference between the actual output and the expected output of the probabilistic neural network, the calculation method is expressed as: ; In the formula, is the loss function of the probabilistic neural network, To sort out the samples input to the probabilistic neural network for the current batch, is the true label of the i-th sample input to the probabilistic neural network, For the i-th sample input to the probabilistic neural network, the model predicts the probability that the sample is a positive class; further, the model prediction probability is calculated by the softmax function, and the calculation method is expressed as: ; In the formula, is the linear output of the model for samples belonging to the i-th category, K is the total number of categories, is the linear output of the model for samples belonging to the kth category; S404, using the loss function to calculate the gradient of each parameter of the probabilistic neural network, and then using the gradient update strategy simulated by the quantum tunneling effect to evaluate the potential barriers in the current gradient direction, and decide whether to perform tunneling, that is, skip the local minimum point to reach a new possible area to avoid the local optimum. The parameter gradient of the probabilistic neural network is , the way to adjust the parameter gradient of the probabilistic neural network using the quantum tunneling effect is: ; In the formula, are the parameters of the updated probabilistic neural network, is the quantum barrier parameter, To control the parameters of the intensity of quantum behavior, the gradient jumps to a new position outside the potential barrier, simulating the tunneling phenomenon; S405. According to the parameter gradient adjusted by the quantum tunneling effect algorithm, the parameters of the probabilistic neural network are updated. The weight and bias of the probabilistic neural network are updated in the following manner: ; ; In the formula, is the weight of the updated probabilistic neural network, is the bias of the updated probabilistic neural network, is the weight of the probabilistic neural network before updating, is the bias of the probabilistic neural network before updating; is the learning rate of the probabilistic neural network for the tth iteration, controls the step size of the gradient descent, and dynamically adjusts the learning rate based on the gradient change in the current iteration. The mathematical model is: ; In the formula, is the learning rate of the t-th iteration probabilistic neural network, To adjust the sensitivity parameter, control the speed of learning rate change; and They are the gradient of the loss function at the t-th iteration and the gradient of the loss function at the t-1-th iteration respectively; S406, repeat the above steps until the preset stop iteration condition is met to complete the classifier model training.

[0029] Specifically, the preset stopping condition during the implementation process is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0030] S5. Construct a slope soil fertility prediction data set, use the trained model to process the data of the slope soil fertility prediction data set, and obtain the slope soil fertility prediction result.

[0031] Specifically, the raw data collected during the implementation process is input into the trained feature dimension reduction model for feature processing. Further, the processed features are input into the classifier model for classification to obtain the classification results. In this implementation, the classification categories include: low fertility, medium fertility, and high fertility.

Claims

1. A method for predicting slope soil fertility, characterized in that: The following steps are involved: S1. Constructing a slope soil fertility prediction training set, including collecting data on the spectral reflectance value of the slope soil and the thermal radiation value of the soil surface, and annotating the collected data; S2. Generate data samples using the dynamically selected SMOTE algorithm for the labeled data to achieve data expansion of the slope soil fertility prediction training set; S3, inputting the expanded slope soil fertility prediction training set into a feature dimensionality reduction model to train the feature dimensionality reduction model, wherein the feature dimensionality reduction model is formed by a boundary smoothing autoencoder; S4, inputting the slope soil fertility prediction training set after dimensionality reduction into a classifier model to train the classifier model, wherein the classifier model is constructed using a probabilistic neural network of quantum tunneling effect; S5. Construct a slope soil fertility prediction data set, use the trained model to process the data of the slope soil fertility prediction data set, and obtain the slope soil fertility prediction result.

2. A slope soil fertility prediction method according to claim 1, characterized in that: The data collected in S1 are from soil spectrum data and topographic data; the annotation method is manual annotation, and the annotation categories are specifically divided into low fertility, medium fertility, and high fertility according to the soil fertility level.

3. A slope soil fertility prediction method according to claim 1, characterized in that: The SMOTE algorithm dynamically selected in S2 is a generation strategy for synthesizing samples by dynamically adjusting the standard SMOTE algorithm according to the local density and diversity of the data. The specific method for generating data samples using the dynamically selected SMOTE algorithm is as follows: S201, initializing the target number of data expansion and the initial synthetic sample generation ratio; S202, evaluating the local density of the current slope soil fertility prediction training set, calculating the sample density in the corresponding neighborhood for the minority class data samples in the training set, and using the sample density information for subsequent dynamic sample generation decisions; S203, dynamically adjusting the synthetic sample generation ratio of each data point according to the evaluated local density, wherein the number of synthetic samples is increased for sample points in low-density areas and the sample generation is reduced for high-density areas; the method of dynamically adjusting the synthetic sample generation ratio of each data point is as follows: ； In the formula, represents the synthetic sample generation ratio of the adjusted i-th sample, represents the initial synthetic sample generation ratio, represents the local density of the i-th sample, Represents the adjustment strength coefficient, and the mathematical model is: ； In the formula, is an adjustable scaling parameter that controls the sensitivity of adjusting the intensity coefficient. is the local inhomogeneity index of the current data set, and the mathematical model is: ； In the formula, is the set of all minority class samples, is the i-th sample The set of nearest neighbor samples, and are the feature vectors of the i-th sample and the j-th sample respectively, is the L2 norm; S204, adopting the interpolation generation method of dual-trajectory learning, selecting strategic data point pairs, using the optimization process to generate a learning trajectory between two data points, interpolating on the two learning trajectories to generate new sample points; the method of using dual-trajectory learning to interpolate and generate synthetic samples is expressed as: ； In the formula, is the i-th synthetic sample, is the i-th synthetic sample point, is the first interpolation coefficient, is the second interpolation coefficient, , , and are two pairs of random synthetic sample points; the interpolation coefficient is dynamically adjusted based on the correlation and spatial structure between the data points, and the calculation method is expressed as: ； In the formula, It is a parameter that controls the distance sensitivity of the item and is used to adjust the distribution of the interpolation coefficients; is the local geometric sensitivity index of the i-th synthetic sample, which is used to evaluate the structural complexity and geometric distribution of the data point in its neighborhood; the mathematical model is: ； S205: merging the generated synthetic samples with the original slope soil fertility prediction training set to form a new and more balanced slope soil fertility prediction training set.

4. A slope soil fertility prediction method according to claim 1, characterized in that: The training method for the feature dimensionality reduction model in S3 is: S301, initializing the parameters of the autoencoder; S302. During the training process of the autoencoder, the expanded slope soil fertility prediction training set data is processed by forward propagation, and then the autoencoder parameters are updated by the gradient descent method through error back propagation; wherein, the weight coefficient of the autoencoder is used to adjust the contribution of each dimensional feature in the error, and these weights are dynamically allocated according to the importance of the feature, and the calculation method is: ； In the formula, is the characteristic variance of the input data of the i-th dimension autoencoder, is the feature of the input data of the i-th dimension autoencoder, is a constant with a value of 0.

001. is the weight coefficient of the i-th dimension autoencoder; S303, adopting an automatic boundary smoothing strategy to enhance the robustness of the model by dynamically adjusting the boundary value, specifically through the smoothing function The output of the encoder is smoothed, and the mathematical model is: ； In the formula, is the output of the encoder; in, ; In the formula, , It is an automatically adjusted threshold parameter used to control the degree of smoothing; In order to smooth the intensity parameters, an adaptive method based on the local density of data is used for design, and the mathematical model is: ； In the formula, is the kernel density estimation function, mean is the averaging function, For the neighborhood, calculate Kernel density estimation in the corresponding neighborhood; S304, advancing the iteration process through the loss function, and updating the autoencoder parameters; S305, using the convex hull convergence strategy to optimize the training process of the autoencoder, by adjusting the learning step size and weight update strategy, the network converges quickly, and the learning rate is dynamically adjusted according to the data distribution characteristics in the current iteration as follows: ； In the formula, is the initial autoencoder learning rate, is the learning rate of the autoencoder at the tth iteration, is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the t-th iteration, is the diameter of the initial convex hull; S306: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

5. A slope soil fertility prediction method according to claim 1, characterized in that: The specific method of using the probabilistic neural network with quantum tunneling effect to form a classifier model training in S4 is: S401, randomly initialize the model parameters of the probabilistic neural network, including: the initial weights of the probabilistic neural network and biases for probabilistic neural networks ; S402, inputting the slope soil fertility prediction training set data features after dimension reduction, forward propagating through the probabilistic neural network layer, and converting the node output of each layer through the activation function to provide input for the next layer until the final classification probability output is generated; S403, using a cross entropy loss function to calculate the difference between the actual output and the expected output of the probabilistic neural network; S404, using the loss function to calculate the gradient of each parameter of the probabilistic neural network, and then using the gradient update strategy simulated by the quantum tunneling effect to evaluate the potential barriers in the current gradient direction, and decide whether to perform tunneling, that is, skip the local minimum point to reach a new possible area to avoid the local optimum. The parameter gradient of the probabilistic neural network is , the way to adjust the parameter gradient of the probabilistic neural network using the quantum tunneling effect is: ； In the formula, are the parameters of the updated probabilistic neural network, is the quantum barrier parameter, To control the parameters of the intensity of quantum behavior, the gradient jumps to a new position outside the potential barrier, simulating the tunneling phenomenon; S405. According to the parameter gradient adjusted by the quantum tunneling effect algorithm, the parameters of the probabilistic neural network are updated. The weight and bias of the probabilistic neural network are updated in the following manner: ； ； In the formula, is the weight of the updated probabilistic neural network, is the bias of the updated probabilistic neural network, is the weight of the probabilistic neural network before updating, is the bias of the probabilistic neural network before updating; is the learning rate of the probabilistic neural network for the tth iteration, controls the step size of the gradient descent, and dynamically adjusts the learning rate based on the gradient change in the current iteration. The mathematical model is: ； In the formula, is the learning rate of the t-th iteration probabilistic neural network, To adjust the sensitivity parameter, control the speed of learning rate change; and They are the gradient of the loss function at the t-th iteration and the gradient of the loss function at the t-1-th iteration respectively; S406, repeat the above steps until the preset stop iteration condition is met to complete the classifier model training.

Images