Soil organic matter content inversion method and system based on improved GOOSE-SVM

Abstract

The application is suitable for the field of soil inversion technology, and provides a soil organic matter content inversion method and system based on improved GOOSE-SVM, which comprises the following steps: collecting hyperspectral data of soil samples; pre-processing the hyperspectral data to obtain pre-processed data; performing dimension reduction processing on the pre-processed data based on the CARS method to obtain dimension-reduced data; constructing an SVM model according to the dimension-reduced data, and optimizing the SVM model parameters based on a preset improved goose swarm optimization algorithm; and using the optimized SVM model to invert the soil organic matter content. The improved goose swarm optimization algorithm is used to optimize the SVM model parameters, which can effectively balance global exploration and local development, and avoid falling into local optimization; the method can realize high-precision inversion of the soil organic matter content, provides a new idea for rapid detection of soil organic matter, and has practical significance for evaluation of the organic matter condition in black soil areas and protection of soil resources.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of soil inversion, and particularly relates to a soil organic matter content inversion method and system based on improved GOOSE-SVM. BACKGROUND

[0002] Soil organic matter is a key indicator of soil fertility. As an important grain-producing base in the northeast, the long-term high-intensity cultivation of North Great Field has caused the black soil layer to thin and the soil organic matter content to rapidly decline. The serious degradation of black soil threatens the ecological environment and food production. Therefore, scientifically and efficiently evaluating the organic matter condition of black soil has become a key to the management and protection of black soil resources.

[0003] Conventional soil organic matter detection relies on laboratory chemical analysis, which is not only cumbersome and time-consuming, but also difficult to obtain continuous information on the spatial distribution of soil fertility and to achieve large-scale rapid monitoring. In contrast, high-spectral technology, with its closely continuous spectral information, can quickly and accurately detect soil physical and chemical properties, and has unique advantages in soil monitoring. The combination of high-spectral technology and machine learning significantly improves the accuracy of soil inversion models, but the spatial heterogeneity of different soil regions is significantly different. The composition of black soil organic matter is complex, and is affected by climate, cultivation and other factors. The spectral data is complex, and the model monitoring faces challenges. Model parameter optimization is a key link to improve model accuracy.

[0004] Model optimization often relies on intelligent optimization algorithms. Among common swarm intelligence algorithms, particle swarm optimization (PSO) is prone to local optimization due to inertia weight parameter settings, and grey wolf optimization (GWO) has the risk of premature convergence in complex problems. In recent years, emerging algorithms such as HO, APO, BKA, and CPO have their own characteristics, but they still tend to lose diversity when dealing with high-dimensional and strongly nonlinear problems, leading to performance degradation. In contrast, the goose optimization algorithm (GOOSE) has shown significant advantages in convergence speed and accuracy in benchmark testing and engineering optimization. This algorithm simulates the cooperative alert mechanism of the goose guard standing on one leg and grasping the stone while resting: the sound of the stone falling can wake up the guard, so as to timely wake up the group to respond to threats. GOOSE models the "stone falling time and sound propagation distance" to adaptively adjust the search step and range, effectively balancing global exploration and local exploitation. However, the robustness and solution accuracy of this algorithm in extreme or high-dimensional complex scenarios still need to be further improved. SUMMARY

[0005] The purpose of the present application is to provide a soil organic matter content inversion method based on improved GOOSE-SVM, which aims to solve the above technical problems.

[0006] This invention is implemented as follows: a soil organic matter content inversion method based on an improved GOOSE-SVM, comprising the following steps:

[0007] Collect hyperspectral data from soil samples;

[0008] The hyperspectral data is preprocessed to obtain the preprocessed data;

[0009] Based on the CARS method, the preprocessed data is dimensionality reduced to obtain dimensionality-reduced data.

[0010] Based on the dimensionality-reduced data, an SVM model is constructed, and the parameters of the SVM model are optimized based on a pre-defined improved goose flock optimization algorithm.

[0011] The optimized SVM model was used to retrieve the soil organic matter content.

[0012] Furthermore, the preprocessing method includes outlier removal and spectral transformation.

[0013] Furthermore, the outlier removal method is as follows: the Mahalanobis distance algorithm is used to identify and remove outlier samples.

[0014] Furthermore, the method for spectral transformation is as follows: the hyperspectral data is transformed using the first derivative.

[0015] Furthermore, the steps for optimizing the SVM model parameters based on the preset improved goose flock optimization algorithm specifically include:

[0016] Initialize the SVM hyperparameters as the initial goose flock positions;

[0017] Dynamically adjust population size;

[0018] The search strategy is executed in stages based on the current iteration number:

[0019] In each iteration, individuals are sorted by fitness, elite individuals are selected and the current generation's best solution is recorded in the history set. After updating the positions of all individuals, the elite individuals are put back into the population to preserve the information of excellent individuals and guide the population to evolve in a better direction.

[0020] After reaching the maximum number of iterations, output the historically optimal combination of SVM hyperparameters.

[0021] Furthermore, the formula for dynamically adjusting the population size is:

[0022] ;

[0023] In the formula, t represents the current iteration number; T represents the total number of iterations, i.e. the maximum number of iterations; N(t) represents the population size at the current iteration number; and N0 represents the total population size.

[0024] Furthermore, based on the current iteration number, the search strategy is executed in stages, specifically including:

[0025] When t < 0.3T, by increasing the population size and combining it with the Levy flight strategy, more promising areas can be discovered, as shown in the following formula:

[0026] ;

[0027] In the formula, The exploration intensity coefficient is represented by r1~U(0,1), which are random numbers in the distribution. L(λ) represents the Levy flight disturbance. i (t) represents an individual in the current generation of the population; X i (t+1) represents the updated population individual; X best (t) represents the best individual in the t-th generation of the population.

[0028] When 0.3T≤t≤0.7T, the original search strategy of the goose flock optimization algorithm is combined with Levy flight and random search strategies to perform a fine search for the potential optimal region, as shown in the following formula:

[0029] ;

[0030] In the formula, ω represents the inertia weight; c1r2 is the acceleration constant; c2r3~U(0,1) is a random number; X pbest,it X represents the historical best position of individual i; gbest The global historical optimal position; β is the Levy flight weight; γ is the random search weight; R~U(X) min ,X max () represents a random point within the search space;

[0031] When t > 0.3T, the population size is reduced and a hill-climbing search strategy is adopted. By continuously moving towards better solutions in the neighborhood of the current solution, the local optimum is gradually approached, as shown in the following formula:

[0032] ;

[0033] In the formula, δ represents the climbing step length; Indicates the objective function in X i Gradient estimation at point ε; ε is the neighborhood search strength; This represents a mixed neighborhood disturbance.

[0034] Furthermore, the elite retention strategy is as follows:

[0035] ;

[0036] In the formula, S(t) represents the top k% of individuals with the best fitness in the current generation; H(t) represents the historical best solution set; H(t+1) represents the updated historical best solution set; and f(X) represents the fitness function, which is determined by the mean squared error.

[0037] Furthermore, the SVM hyperparameters include a penalty factor and kernel function parameters.

[0038] Another objective of this invention is to provide a soil organic matter content inversion system based on an improved GOOSE-SVM, for implementing the aforementioned soil organic matter content inversion method, comprising:

[0039] The data acquisition module is used to collect hyperspectral data from soil samples;

[0040] The data preprocessing module is used to preprocess the hyperspectral data to obtain preprocessed data;

[0041] The data dimensionality reduction module is used to perform dimensionality reduction on preprocessed data based on the CARS method to obtain dimensionality-reduced data.

[0042] The model optimization module is used to construct an SVM model based on the dimensionality-reduced data and optimize the SVM model parameters based on a preset improved goose flock optimization algorithm.

[0043] The inversion module is used to invert the soil organic matter content using the optimized SVM model.

[0044] The soil organic matter content inversion method based on the improved GOOSE-SVM provided by this invention optimizes the SVM model parameters by using an improved goose flock optimization algorithm, which can effectively balance global exploration and local development and avoid getting trapped in local optima. This method can achieve high-precision inversion of soil organic matter content, providing a new approach for rapid detection of soil organic matter, and has practical significance for the assessment of organic matter status and soil resource protection in black soil areas. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the process for retrieving soil organic matter content based on an improved GOOSE-SVM method, provided in an embodiment of the present invention.

[0046] Figure 2 These are the original soil spectra and those obtained after various transformation processes.

[0047] Figure 3 The distribution maps of characteristic bands obtained by different dimensionality reduction methods are shown.

[0048] Figure 4 This is a flowchart illustrating the improved goose flock optimization algorithm provided in an embodiment of the present invention.

[0049] Figure 5 A radar chart comparing the CEC2022 test function results of IGOOSE and GOOSE.

[0050] Figure 6 This is a scatter plot of the true and predicted values ​​of the IGOOSE-optimized SVM model provided in an embodiment of the present invention.

[0051] Figure 7 This is a schematic diagram of the structure of the soil organic matter content inversion system based on the improved GOOSE-SVM provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] This invention uses soil samples from Baoquanling Farm and Youyi Farm in Heilongjiang Province as research objects, and proposes an improved goose flock optimization algorithm strategy (IGOOSE) to construct a black soil organic matter inversion model. By comparing various spectral processing and modeling methods, the global exploration stage is strengthened on the basis of the goose flock optimization algorithm, and the dynamic population mechanism is adjusted to improve the accuracy of the black soil organic matter inversion model, providing more scientific and accurate technical support for improving and monitoring black soil fertility.

[0054] Specifically, such as Figure 1 As shown, in one embodiment of the present invention, a method for inverting soil organic matter content based on an improved GOOSE-SVM is provided, comprising the following steps:

[0055] S100, Collect hyperspectral data of soil samples;

[0056] S200. Preprocess the hyperspectral data to obtain preprocessed data;

[0057] S300: Based on the CARS method, the preprocessed data is dimensionality reduced to obtain dimensionality-reduced data.

[0058] S400. Based on the dimensionality-reduced data, construct an SVM model and optimize the SVM model parameters based on a preset improved goose flock optimization algorithm.

[0059] S500. The optimized SVM model was used to invert the soil organic matter content.

[0060] In practical application, the soil samples in step S100 were collected from Youyi Farm and Baoquanling Farm in Heilongjiang Province. Based on different terrains and crop planting types, a combination of random sampling and grid-based sampling was used to collect a total of 387 soil samples. Of these, 215 samples were collected from Youyi Farm and 172 from Baoquanling Farm. During the sampling process, after removing surface debris, soil samples were taken from a depth of 0-20 cm using a soil drill. Five sub-samples were thoroughly mixed into one sample at each sampling point. After removing abnormal soil containing large amounts of gravel, plant and animal remains, etc., each sample was retained at (200±2) g, sealed, and clearly labeled. To minimize the impact of environmental factors on the samples, soil samples were placed in a constant temperature and humidity environment in the laboratory (temperature 25±2℃, humidity 40-60%) for 48 hours to allow their physicochemical properties to stabilize. After this period, the soil organic matter content was accurately determined using the potassium dichromate oxidation method. Simultaneously, hyperspectral data acquisition was performed using a FieldSpec 4 spectrometer manufactured by ASD (Americas), with a measurement wavelength of 350nm-2500nm and a resolution of 3nm. (200±2)g of the settled soil sample was taken and evenly filled into the instrument's cylindrical sample well (8cm in diameter, 5cm in depth). The sample well was gently pressed to ensure a smooth surface without gaps, and a custom-made standard white board was used to eliminate edge light leakage and ambient light interference. The sample well was then placed in the center of the instrument's stage. Each sample was scanned 10 times, and the average spectrum was taken as the effective hyperspectral data.

[0061] In a preferred embodiment of the present invention, the preprocessing method includes outlier removal and spectral transformation. During data acquisition, errors introduced by human factors or equipment may lead to outliers, thereby reducing data quality. Therefore, this embodiment of the present invention employs the existing Mahalanobis distance algorithm to identify and remove outlier samples, and performs spectral data transformation processing in subsequent steps.

[0062] In practical applications, after using the Mahalanobis distance algorithm (MA) to remove outlier samples, the remaining 366 soil samples were randomly divided into a training set and a test set in a ratio of 8:2. The training set had 292 samples and the test set had 74 samples, which were used for subsequent model building and accuracy evaluation. The dataset sample statistics are shown in Table 1.

[0063] Table 1. Dataset Sample Statistics

[0064]

[0065] In addition, to address issues such as stray light, instrument noise, and baseline drift in the original hyperspectral data, this invention employs methods such as Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), Savitzky-Golay Filter (SG), First-order Derivative (D1), and Second-order Derivative (D2) to transform the spectral data. The optimal method is selected by comparing and comparing different methods using a Partial Least Squares Regression (PLSR) model. Preferably, the first-order derivative is used to transform the hyperspectral data.

[0066] In practical applications, the original soil spectra and those obtained from various transformation processes are as follows: Figure 2 As shown.

[0067] Depend on Figure 2 As can be seen from (1), the original soil spectrum generally exhibits a typical reflectance spectrum, with its reflectance typically increasing gradually with increasing wavelength. In specific wavelength bands (such as around 1400 nm, 1900 nm, and 2200 nm), significant absorption valleys are formed due to the absorption of components such as moisture and organic matter. However, the original is easily affected by factors such as particle size, surface roughness, and measurement conditions, resulting in baseline drift and stray light noise as shown in the figure. Figure 2 As can be seen from (2) and (3), multivariate scattering correction (MSC) and standard normal variable transformation (SNV) can effectively suppress path-dependent scattering effects, making the spectral curves more concentrated in the longitudinal direction, highlighting the absorption characteristics of chemical components, and thus improving the comparability of spectral shapes. Figure 2 As can be seen from (4), Convolutional Smoothing Spectrum (SG) performs convolution operations on the spectrum through local polynomial fitting, effectively smoothing subtle fluctuations while preserving the key features of the Origin. Figure 2 As can be seen from (5) and (6), the first derivative spectrum (D1) can highlight the slope change point of Origin, accurately identify the position and inflection point of absorption features, and clearly distinguish the absorption positions of organic matter-related functional groups such as 1400nm (OH stretching vibration), 1900nm (OH overtone absorption), and 2200-2300nm (CH vibration); the second derivative (D2) further responds to the rate of change of D1, is more sensitive to determining the center position of absorption valley, and can effectively distinguish the broad and gentle absorption bands that overlap in Origin, but the noise is more prominent than that of D1.

[0068] In addition, PLSR model indices were established using data preprocessed using the different methods described above, as shown in Table 2.

[0069] Table 2. Indicators for building PLSR models from data preprocessed using different methods.

[0070]

[0071] As shown in Table 2, in the performance comparison of PLSR models built based on different preprocessing methods, the first derivative (D1) preprocessing shows a significant advantage. The model built from the unprocessed raw data has the weakest performance, with its test set R... 2 The value was only 0.2007, indicating that the model could hardly explain the changes in soil organic matter. While pretreatment methods such as SNV, MSC, and SG slightly improved the model performance, the R value on the test set was still insufficient. 2 None of the above exceeded 0.44, indicating limited improvement. In contrast, spectral derivative processing, especially first derivative (D1) preprocessing, significantly improved model performance. The D1-preprocessed PLSR model achieved best or near-best metrics on five-fold cross-validation, the training set, and the test set, with its test set R... 2 The R² value reached 0.6377, while the RMSE decreased to 9.5918, demonstrating that this processing method is highly effective in eliminating background interference and enhancing the effective signal, significantly improving the model's prediction accuracy and robustness. Although the second derivative (D²) processing showed a high fit on the training set (R²...),... 2 =0.8190), but its performance on the test set (R) 2 The value of D1 (0.5497) is significantly lower than that of D1, indicating that the model may be overfitting. Overall, D1 is the most effective at extracting key information related to organic matter from the original soil spectrum; therefore, in this embodiment of the invention, D1 preprocessed data is preferred for subsequent model construction.

[0072] In a preferred embodiment of the present invention, to improve model computational efficiency and generalization ability, and reduce the interference of redundant information in spectral data, the present invention employs three feature selection methods for data dimensionality reduction: Competitive Adaptive Reweighted Sampling (CARS), Random Frog (RFROG), and Successive Projection Algorithm (SPA). These methods systematically compare the impact of different feature dimensions on modeling performance. SPA and RFROG tend to select the feature subset with the strongest discriminative power, making them suitable for scenarios with high requirements for model simplification and interpretability. CARS, on the other hand, retains more potentially effective information by simulating a competition mechanism between variables, making it more suitable for complex modeling tasks with high prediction accuracy requirements.

[0073] In practical applications, three models—Support Vector Machine (SVM), Random Forest (RF), and Extreme Learning Machine (ELM)—were constructed based on the data preprocessed using D1. The accuracy comparisons of the different models are shown in Table 3.

[0074] Table 3 Comparison of accuracy of different models

[0075]

[0076] As can be seen from the data in Table 3, the average R-value of the SVM model in 5-fold cross-validation is... 2 The R² value is 0.7624, higher than the 0.5189 of the RF model and the 0.5351 of the ELM model; on both the training and test sets, the R² value of the SVM model is higher. 2 The values ​​were also relatively high, at 0.99085 and 0.43179 respectively. Furthermore, as shown in Table 3, the SVM model built based on the D1 preprocessed data exhibited the best performance in both prediction accuracy and generalization ability. In the five-fold cross-validation, a key indicator for evaluating model robustness, the SVM model's average R-value was [missing value]. 2 The highest R² value was 0.7624, and the lowest RMSE was 7.8735, indicating that it has the strongest adaptability and stability on different data subsets. Further observation of the performance on the training and test sets reveals that while RF and ELM show high goodness of fit on the training set, their test set results show a significant decline, especially ELM, where the training set R² value drops considerably. 2 The value is as high as 0.9532, while the test set R... 2 The R² value is only 0.5759, reflecting a certain risk of overfitting. In contrast, the SVM model achieves an excellent fit on the training set (R² = 0.5759). 2 =0.9906), while still maintaining the highest R on the test set. 2 The low RMSE (0.8195) and low generalization ability (5.2458) demonstrate its excellent generalization ability and practical value, indicating that the optimal kernel function type is the Gaussian kernel function. In summary, when retrieving soil organic matter content based on D1 preprocessed spectral data, SVM is the most reliable and effective modeling choice among the three models. Therefore, the SVM model using the Gaussian kernel function will be the basis for subsequent experiments.

[0077] Analysis of data modeling results after dimensionality reduction: The distribution map of characteristic bands after dimensionality reduction of the spectrum is shown in the following figure, obtained by using competitive adaptive reweighted sampling (CARS), random frog jumping (RFROG), and continuous projection (SPA) algorithms. Figure 3As shown, the distribution of the selected bands in the 350-2500 nm spectral range reveals that the bands selected by each method are concentrated in several specific regions, particularly in the 2000-2400 nm range. This reflects the strong spectral response of these wavelength positions to soil organic matter. The commonly selected band regions may correspond to the vibrational absorption bands of functional groups such as CH, NH, and OH in the organic matter. The screening results in the visible light region may be related to the color characteristics of soil organic matter, while the near-infrared bands are related to the combination and overtone absorption of hydrogen-containing groups in the soil.

[0078] In addition, as shown in Table 4, CARS, RFROG, and SPA selected 16, 10, and 44 feature bands, respectively, with dimensionality reduction ratios all exceeding 97%. Due to the differences in their basic principles, these methods exhibit their own characteristics in feature selection strategies. SPA and RFROG selected fewer bands with a concentrated distribution, demonstrating strong feature compression capabilities, which helps to build simple and efficient prediction models. CARS retained a larger number of bands with a wider distribution, giving it an advantage in preserving detailed information related to organic matter.

[0079] Table 4 Number of Feature Bands Selected

[0080]

[0081] The SVM model was reconstructed for the data processed by the three dimensionality reduction algorithms to determine the optimal dimensionality reduction method. The accuracy of each model is shown in Table 5.

[0082] Table 5. Comparison of the accuracy of SVM models built from dimensionality-reduced data.

[0083]

[0084] The analysis of Table 5, comparing the accuracy of the three dimensionality reduction methods combined with the SVM model, shows that the CARS algorithm performs best in the SVM model constructed after dimensionality reduction. Its R-values ​​on five-fold cross-validation, training set, and test set are significantly higher. 2 All scores were the highest, and RMSE scores were the lowest, especially on the test set. 2 The RSI reached 0.7775, and the RMSE was 6.6407, indicating that the model has excellent prediction accuracy and good generalization ability, effectively preserving key features in spectral data. The model built using the RFROG method showed consistent performance across various datasets, with an RSI of 0.7775. 2 The R² value is approximately between 0.62 and 0.65. Although no significant overfitting was observed, the overall accuracy was only average, indicating that its dimensionality reduction effect was relatively conservative and its information extraction capability was limited. The model constructed using the SPA method underwent five-fold cross-validation with R². 2The R² value is only 0.2311, far lower than its 0.8199 on the training set, indicating that the feature set selected by this method has poor generalization ability and the model is severely unstable. Although its R² value on the test set is much higher... 2 The dimensionality reduction efficiency was 0.6799, but considering its cross-validation performance, the reliability and practicality of the SPA dimensionality reduction results were significantly insufficient. In summary, CARS is the most effective dimensionality reduction method. While significantly compressing the data dimensionality, it retains the key spectral information used for modeling to the greatest extent, providing strong support for building a high-precision, highly generalizable soil organic matter inversion model. Therefore, this embodiment of the invention preferentially selects CARS method for dimensionality reduction data for subsequent model optimization.

[0085] In a preferred embodiment of the present invention, three regression models—Support Vector Machine (SVM), Random Forest (RF), and Extreme Learning Machine (ELM)—are used to compare their performance in soil organic matter content retrieval. The above models are constructed based on optimally preprocessed data to evaluate their specific performance under experimental data conditions. SVM achieves data regression by finding the optimal classification hyperplane. This embodiment employs a grid search method to dynamically optimize Gaussian, multinomial, and linear kernel functions to determine the optimal kernel configuration. RF, as an ensemble learning algorithm, constructs multiple decision trees by randomly selecting samples and features, and integrates the decision results to improve model performance. ELM quickly determines the network output weights by randomly initializing input weights and hidden layer biases, thereby achieving efficient learning. This embodiment preferably uses the SVM model, and the specific construction method is as follows:

[0086] (1) Model input parameters and dataset partitioning: Soil sample data after removing outliers are randomly divided into training and test sets in an 8:2 ratio. All models in this embodiment of the invention are constructed and their accuracy evaluated based on this partitioning.

[0087] (2) Model optimization: The model parameters were optimized using the Goose Optimization Algorithm (GOOSE), and the GOOSE algorithm was improved.

[0088] Among them, the Goose Flock Optimization Algorithm (GOOSE) is a novel metaheuristic optimization algorithm inspired by the alert behavior of geese. After initializing the population, the algorithm determines the current optimal position by comparing individual fitness values. Its core lies in using random numbers to balance the exploration and development phases. In the development phase, the algorithm simulates the physical processes of "falling stones" and "sound propagation" to update individual positions. This process accurately simulates the intelligent behavior of geese adjusting their positions based on the direction and distance of the sound source after receiving an alarm. This unique mechanism enables the GOOSE algorithm to establish an effective balance between global exploration and local development, thus exhibiting excellent convergence performance. The specific mathematical model is described below:

[0089] First, the weight of the stones carried by the guard goose is randomly generated, and this weight determines the step size of the local search:

[0090] Sw=rand([5,25]) (1;

[0091] Then, the distance the sound of the stone hitting the ground is calculated, which defines the search range:

[0092] DST = SS × TAS (2);

[0093] The distance between the guard goose and other individuals in the group is calculated using the following formula:

[0094] DG = 0.5 × DST (3);

[0095] Finally, the individual's position is updated based on the aforementioned physical quantities. A core position update formula is:

[0096] X(t+1) = FFS + DG × TA 2 (4);

[0097] In the formula, SS = 343.2 m / s is the speed of sound in air, TAS is the sound propagation time, FFS is the free fall velocity of the stone, TA is the average sound propagation time, t is the current iteration number, and X(t+1) is the updated population.

[0098] (3) Model accuracy evaluation index: During the model construction process, the average index was calculated using 5-fold cross-validation to evaluate the generalization of the model, and the root mean square error (RMSE) and coefficient of determination (R²) of the three models on the training and test sets were compared. 2 Select the optimal model.

[0099] like Figure 4 As shown, in a preferred embodiment of the present invention, the step of optimizing the SVM model parameters based on a preset improved goose flock optimization algorithm specifically includes:

[0100] Initialize the SVM hyperparameters as the initial flock positions; the SVM hyperparameters include the penalty factor and kernel function parameters.

[0101] Dynamically adjust population size;

[0102] The search strategy is executed in stages based on the current iteration number:

[0103] In each iteration, individuals are sorted by fitness, elite individuals are selected and the current generation's best solution is recorded in the history set. After updating the positions of all individuals, the elite individuals are put back into the population to preserve the information of excellent individuals and guide the population to evolve in a better direction.

[0104] After reaching the maximum number of iterations, output the historically optimal combination of SVM hyperparameters.

[0105] In practical applications, because the goose optimization algorithm controls the exploration and development behavior of the goose flock by controlling parameters throughout the entire iteration cycle, it cannot specifically strengthen the global search in the early stage and refine the local exploration in the later stage. This makes the algorithm prone to getting trapped in local optima, with low search efficiency in the later stage and difficulty in accurately approximating the global optimum.

[0106] To address the above issues, the goose population size is dynamically adjusted using formula (5) during the algorithm iteration process:

[0107] (5);

[0108] In the formula, t represents the current iteration number; T represents the total number of iterations, i.e. the maximum number of iterations; N(t) represents the population size at the current iteration number; and N0 represents the total population size.

[0109] In addition, the steps for executing the search strategy in stages, based on the current iteration number, specifically include:

[0110] When t < 0.3T, that is, in the first 30% stage, by increasing the population size and combining the Levy flight strategy for searching, the algorithm can explore a larger range in the search space, which helps to escape local optima and discover more promising regions, as shown in formula (6):

[0111] (6);

[0112] In the formula, The exploration intensity coefficient is represented by r1~U(0,1), which are random numbers in the distribution. L(λ) represents the Levy flight disturbance. i (t) represents an individual in the current generation of the population; X i (t+1) represents the updated population individual; X best (t) represents the best individual in the t-th generation of the population.

[0113] When 0.3T≤t≤0.7T, that is, in the 30%-70% stage, the original search strategy of the goose flock optimization algorithm is combined with Levy flight and random search strategies to balance the global search and local search capabilities of the algorithm. While expanding the search range, a fine search is performed on the potential optimal region as shown in equation (7):

[0114] (7);

[0115] In the formula, ω represents the inertia weight; c1r2 is the acceleration constant; c2r3~U(0,1) is a random number; X pbest,it X represents the historical best position of individual i; gbest The global historical optimal position; β is the Levy flight weight; γ is the random search weight; R~U(X) min ,X max () represents a random point within the search space;

[0116] When t > 0.3T, that is, in the last 30% stage, the population size is reduced and a hill-climbing search strategy is adopted. By continuously moving to better solutions in the neighborhood of the current solution, the local optimum is gradually approached, as shown in the following equation (8):

[0117] (8);

[0118] In the formula, δ represents the climbing step length; Indicates the objective function in X i Gradient estimation at point ε; ε is the neighborhood search strength; This represents a mixed neighborhood disturbance.

[0119] Meanwhile, for individuals in the latter 30% of the search, a hybrid neighborhood search strategy is adopted, which combines single-dimensional and multi-dimensional perturbations. Single-dimensional perturbations can fine-tune the individual's position in each dimension, uncovering subtle differences in local areas; multi-dimensional perturbations, on the other hand, make a larger-scale adjustment to the individual's position as a whole, enhancing the algorithm's search capability in local areas and more accurately approximating the optimal solution.

[0120] In each iteration, the algorithm sorts individuals by fitness, selects elite individuals, records the current generation's best solution in the history set, updates the positions of all individuals, and then puts the elite individuals back into the population to preserve the information of superior individuals and guide the population to evolve in a better direction. The elite preservation strategy is shown in equation (9):

[0121] (9);

[0122] In the formula, S(t) represents the top k% of individuals with the best fitness in the current generation; H(t) represents the historical best solution set; H(t+1) represents the updated historical best solution set; and f(X) represents the fitness function, which is determined by the mean squared error.

[0123] In practical applications, the performance comparison between the improved goose flock optimization algorithm (IGOOSE) and the existing goose flock optimization algorithm (GOOSE) is as follows:

[0124] IGOOSE and GOOSE were compared on the CEC2022 test function. The population iteration count was set to 500 generations, the population size to 40, and the test function dimension to 20. Each algorithm was run 30 times to calculate the mean, standard deviation, best, and worst values ​​of the 12 test functions. The CEC2022 test function results for IGOOSE and GOOSE are compared in Table 6, and the comparison radar chart is shown below. Figure 5 As shown.

[0125] Table 6 Comparison of CEC2022 test results for IGOOSE and GOOSE

[0126]

[0127] From Table 6 and Figure 5 The CEC2022 test function results show that IGOOSE outperforms the original algorithm GOOSE in most functions. On functions F1, F3, F5, and F9, both algorithms approach the theoretical optimum, but IGOOSE generally has a lower standard deviation, demonstrating more stable convergence characteristics. For example, in F3, the standard deviation of IGOOSE is 2.4413E-04, significantly lower than GOOSE's 4.3647E-02, indicating less fluctuation in the quality of its solutions.

[0128] In more complex multimodal and composite functions, IGOOSE's advantages are even more pronounced. In functions F2, F4, F6, F7, F10, and F11, IGOOSE not only significantly outperforms GOOSE in terms of average value, but its optimal and worst solutions are also generally superior. For example, in F6, IGOOSE's average value of 3.2264E+03 is significantly lower than GOOSE's 4.2149E+03, and its worst solution of 7.8070E+03 is also better than GOOSE's 8.0652E+03, demonstrating stronger global exploration capability and robustness.

[0129] In some functions, such as F8 and F12, the performance of the two algorithms is quite similar, but IGOOSE still has a slight advantage in average value and stability. Overall, the improved algorithm shows better convergence accuracy and stability in the vast majority of test functions, verifying the effectiveness of its improvement strategy in enhancing algorithm performance. When applied to SVM hyperparameter optimization, it can obtain models with higher accuracy and stability.

[0130] The SVM model was optimized using the IGOOSE algorithm described above: Using the mean squared error of the training set as the fitness function, the two hyperparameters of the SVM were optimized using the IGOOSE algorithm for soil organic matter regression. The specific optimization steps are illustrated in the flowchart below. Figure 4 As shown.

[0131] The penalty factor C, kernel function type, and kernel function parameter γ of the SVM model were optimized using GOOSE and IGOOSE. Table 7 shows a comparison of the accuracy of the GOOSE-optimized SVM model (GOOSE-SVM model) and the IGOOSE-optimized SVM model (IGOOSE-SVM model). Additionally, a scatter plot of the actual and predicted values ​​of the IGOOSE-SVM model is shown below. Figure 6 As shown.

[0132] Table 7 Comparison of accuracy between GOOSE-SVM and IGOOSE-SVM models

[0133]

[0134] Based on the model accuracy comparison results in Table 7, IGOOSE demonstrates a significant advantage in improving SVM model performance. Compared to the original GOOSE algorithm, IGOOSE finds a different set of SVM parameter combinations with a smaller penalty coefficient C and a larger kernel function parameter γ. This configuration helps the model achieve a better balance between fitting and generalization. This indicates that in dealing with complex soil spectral data modeling problems, the multi-stage search strategy and dynamic population adjustment mechanism enable it to better adapt to the complex feature space of the data. Through Levy fly-through and random search strategies, the IGOOSE algorithm can explore the solution space more extensively during the global search process, avoiding the algorithm from getting trapped in local optima; while in the local search stage, the combination of hill climbing and hybrid neighborhood search strategies allows the algorithm to perform fine-grained searches within local regions, improving the accuracy of approximating the optimal solution.

[0135] In terms of model robustness, IGOOSE-SVM achieved a higher average R-value in five-fold cross-validation. 2 The lower RMSE indicates more stable predictive performance across different data subsets. Further analysis of its performance on the training and test sets reveals that IGOOSE-SVM exhibits extremely high goodness of fit on the training set, with R... 2 The R-value reached 0.9252, and this excellent fitting ability was effectively transferred to the test set, with its R-value being [missing value]. 2 The performance improved to 0.8416, and the RMSE decreased to 7.0407 g / kg. In comparison, the GOOSE-SVM model performed relatively poorly on the training set and showed weak generalization ability to the test set.

[0136] In summary, the IGOOSE algorithm, through a more effective parameter search strategy, not only significantly improves the model's ability to capture data features but also ensures excellent generalization performance, successfully constructing a soil organic matter content prediction model that combines high accuracy and strong robustness.

[0137] This invention, based on hyperspectral technology, enables rapid detection of soil organic matter content. Analysis and processing of hyperspectral data from soil samples from Baoquanling Farm and Youyi Farm in Heilongjiang Province yielded the following conclusions:

[0138] (1) Based on the Mahalanobis distance algorithm to remove abnormal samples, the first derivative (D1) is used to preprocess the soil organic matter hyperspectral data, which can better preserve and highlight the effective spectral feature information while suppressing noise interference.

[0139] (2) Among the three machine learning models SVM, RF and ELM, SVM shows better fitting effect and generalization ability on both the training set and the test set, and is suitable for hyperspectral inversion modeling of soil organic matter content.

[0140] (3) Comparing the three feature selection methods CARS, RFROOG and SPA, the CARS algorithm retains a higher feature dimension than the other methods, but the SVM model built on it has the best prediction performance. It can more effectively retain key information during the dimensionality reduction process.

[0141] (4) To further improve model performance, a multi-strategy improved goose flocking algorithm (IGOOSE) is proposed. Its optimized SVM model significantly outperforms the original GOOSE algorithm in generalization ability, and the R-values ​​of the training set and test set are significantly higher. 2 The RMSE values ​​were increased by 0.1066 and 0.0411 respectively, and decreased by 2.5413 g / kg and 0.5287 g / kg respectively, indicating that the model performance meets the requirements of practical applications.

[0142] In summary, the IGOOSE-SVM soil organic matter hyperspectral inversion method proposed in this invention provides a new approach for rapid detection of soil organic matter and has practical significance for assessing the organic matter status and protecting soil resources in black soil regions.

[0143] like Figure 7 As shown, in another embodiment of the present invention, a soil organic matter content inversion system based on an improved GOOSE-SVM is also provided to implement the above-mentioned soil organic matter content inversion method, comprising:

[0144] Data acquisition module 10 is used to acquire hyperspectral data of soil samples;

[0145] Data preprocessing module 20 is used to preprocess hyperspectral data to obtain preprocessed data;

[0146] The data dimensionality reduction module 30 is used to perform dimensionality reduction on the preprocessed data based on the CARS method to obtain dimensionality-reduced data.

[0147] The model optimization module 40 is used to construct an SVM model based on the dimensionality-reduced data and optimize the SVM model parameters based on a preset improved goose flock optimization algorithm.

[0148] Inversion module 50 is used to invert soil organic matter content using the optimized SVM model.

[0149] It should be noted that each of the above modules can be implemented as a computer program, which can run on a computer device. The computer device's memory can store the computer program that makes up each module, enabling the processor to execute each step of the above method.

[0150] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0151] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0152] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for inverting soil organic matter content based on an improved GOOSE-SVM, characterized in that, Includes the following steps: Collect hyperspectral data from soil samples; after removing abnormal soils, place the soil samples in a constant temperature and humidity environment in the laboratory until their physicochemical properties stabilize, then determine the soil organic matter content and take the average spectrum as the effective hyperspectral data. The hyperspectral data is preprocessed to obtain the preprocessed data; Based on the CARS method, the preprocessed data is dimensionality reduced to obtain dimensionality-reduced data. Based on the dimensionality-reduced data, an SVM model is constructed, and the SVM model parameters are optimized using a pre-defined improved goose flock optimization algorithm. The improved goose flock optimization algorithm includes dynamically adjusting the population size; the formula for dynamically adjusting the population size is: ； In the formula, t represents the current iteration number; T represents the total number of iterations, i.e. the maximum number of iterations; N(t) represents the population size at the current iteration number, and N0 represents the total population size; When t < 0.3T, by increasing the population size and combining it with the Levy flight strategy, more promising areas can be discovered, as shown in the following formula: ； In the formula, The exploration intensity coefficient is represented by r1~U(0,1), which are random numbers in the distribution. L(λ) represents the Levy flight disturbance. i (t) represents an individual in the current generation of the population; X i (t+1) represents the updated population individual; X best (t) represents the best individual in the population of generation t; When 0.3T≤t≤0.7T, the original search strategy of the goose flock optimization algorithm is combined with Levy flight and random search strategies to perform a fine search for the potential optimal region, as shown in the following formula: ； In the formula, ω represents the inertia weight; c1r2 is the acceleration constant; c2r3~U(0,1) is a random number; X pbest,it X represents the historical best position of individual i; gbest The global historical optimal position; β is the Levy flight weight; γ is the random search weight; R~U(X) min ,X max () represents a random point within the search space; When t > 0.3T, the population size is reduced and a hill-climbing search strategy is adopted. By continuously moving towards better solutions in the neighborhood of the current solution, the local optimum is gradually approached, as shown in the following formula: ； In the formula, δ represents the climbing step length; Indicates the objective function in X i Gradient estimation at point ε; ε is the neighborhood search strength; Represents mixed neighborhood perturbation; The optimized SVM model was used to retrieve the soil organic matter content.

2. The soil organic matter content inversion method based on the improved GOOSE-SVM according to claim 1, characterized in that, The preprocessing methods include outlier removal and spectral transformation.

3. The soil organic matter content inversion method based on the improved GOOSE-SVM according to claim 2, characterized in that, The outlier removal method is as follows: the Mahalanobis distance algorithm is used to identify and remove outlier samples.

4. The method for inverting soil organic matter content based on the improved GOOSE-SVM according to claim 2, characterized in that, The method for spectral transformation is as follows: the hyperspectral data is transformed using the first derivative.

5. The soil organic matter content inversion method based on the improved GOOSE-SVM according to claim 4, characterized in that, The steps for optimizing SVM model parameters based on a pre-defined improved goose flock optimization algorithm specifically include: Initialize the SVM hyperparameters as the initial goose flock positions; Dynamically adjust population size; The search strategy is executed in stages based on the current iteration number: In each iteration, individuals are sorted by fitness, elite individuals are selected and the current generation's best solution is recorded in the history set. After updating the positions of all individuals, the elite individuals are put back into the population to preserve the information of excellent individuals and guide the population to evolve in a better direction. After reaching the maximum number of iterations, output the historically optimal combination of SVM hyperparameters.

6. The soil organic matter content inversion method based on the improved GOOSE-SVM according to claim 1, characterized in that, The elite retention strategy is shown in the following formula: ； In the formula, S(t) represents the top k% of individuals with the best fitness in the current generation; H(t) represents the historical best solution set up to generation t; H(t+1) represents the updated historical best solution set; and f(X) represents the fitness function, which is determined by the mean squared error.

7. The method for inverting soil organic matter content based on the improved GOOSE-SVM according to claim 5, characterized in that, The SVM hyperparameters include the penalty factor and kernel function parameters.

8. A soil organic matter content inversion system based on an improved GOOSE-SVM, used to implement the soil organic matter content inversion method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect hyperspectral data from soil samples; The data preprocessing module is used to preprocess the hyperspectral data to obtain preprocessed data; The data dimensionality reduction module is used to perform dimensionality reduction on preprocessed data based on the CARS method to obtain dimensionality-reduced data. The model optimization module is used to construct an SVM model based on the dimensionality-reduced data and optimize the SVM model parameters based on a preset improved goose flock optimization algorithm. The inversion module is used to invert the soil organic matter content using the optimized SVM model.

Images