CN110334869A - A training method for mangrove ecological health prediction based on dynamic group optimization algorithm - Google Patents

Abstract

The invention discloses a mangrove ecological health prediction training method based on dynamic group optimization algorithm, comprising step A, data preprocessing; step A1, data generation module based on PSO algorithm; step A2, based on relative entropy and cosine similarity Similarity evaluation module; step B, model training based on DGO algorithm; step B1, intra-group cooperation; step B2, the centroid of each group communicates with the centroids of other groups; step B3, the members of each group Other members perform random crossover; step B4, during the communication between each centroid and the random crossover process of each member, search for the mutation situation, find out the corresponding member based on the mutation situation, and transfer the member randomly Go to other groups; step C, combine BP algorithm and DGO algorithm with step B to train the feedforward neural network. The invention improves the accuracy of the prediction data and the integrity of the training model.

Description

A training method for mangrove ecological health prediction based on dynamic group optimization algorithm

technical field

The invention relates to the technical field of mangrove ecological protection and data mining, in particular to a mangrove ecological health prediction training method based on a dynamic group optimization algorithm.

Background technique

Mangrove is a woody plant community that grows in the intertidal zone of tropical and subtropical coasts and has the characteristics of both terrestrial and marine ecosystems. It plays a vital role in berms, purification of pollution, etc. Mangrove wetland ecosystems are known as "species gene pools", and protecting the integrity of mangrove ecosystems is one of the important foundations for maintaining biodiversity.

The monitoring, protection and pest control of mangrove wetland ecological environment are of great strategic significance to support the coastal marine ecological security and sustainable development along the southeast coast of my country. However, in recent years, due to the destruction of human activities, mangrove resources have decreased sharply. A reliable mangrove ecological health monitoring and evaluation system has been established to speed up the construction of ecological environment monitoring network, which in turn provides important information for the formulation of environmental protection standards and government decision-making. It is necessary to promote the orderly and efficient progress of environmental protection work, thereby improving the quality of the environment.

For this reason, ecological monitoring mainly through neural network algorithms has appeared in the prior art. The structure construction based on neural network includes a learning and training method. The training method is mainly divided into two categories: gradient method and heuristic method. The minimum value method is to find the minimum value according to the direction of gradient descent, so it can quickly converge to the nearest minimum value, that is, the target optimal value. The most widely used gradient method is the back propagation (BP) algorithm. , which is a concept proposed by scientists led by Rumelhart and McClelland in 1986. It is a multi-layer feedforward neural network trained according to the error back-propagation algorithm; heuristic algorithm is an algorithm based on intuition or experience. A feasible solution for each instance of the combinatorial optimization problem to be solved is given under the cost (referring to computing time and space), and the degree of deviation of the feasible solution from the optimal solution cannot generally be predicted.

At present, the heuristic algorithm is mainly based on the natural body algorithm, mainly including genetic algorithm, simulated annealing method, particle swarm optimization algorithm and so on. Compared with the heuristic algorithm, the gradient method has strong local search ability and can quickly converge to the optimal value. However, it is difficult to find the global optimal value. Similarly, under the principle of the gradient method, the position of the initial point is very important. As shown in Figure 2, if the position of the red point is the initial value, then the algorithm will Take it to the direction marked by the black line to find the minimum value, and then you will find the local minimum on the right and stagnate. It can be seen that this point is not the target global optimal value. The gradient algorithm has the following problems in the specific training: 1) The algorithm is sensitive to the initial value, and different initial values will get different results; 2) It is easy to fall into the local optimal value; 3) The learning rate is a very important superimposition in the gradient method. If the learning rate is too large, the algorithm is difficult to converge, and if the learning rate is too small, it may converge prematurely, and it is difficult to adjust the parameters; 4) If the search space is very complex, it is prone to oscillation.

Compared with the gradient method, the heuristic algorithm does not update in a specific direction, it allows random mutation, and there are three types of algorithms: algorithm based on a single solution, algorithm based on population and other related algorithms. Most of the heuristic algorithms mainly form a search method to find the global optimal solution by simulating the behavior of animals. For example, the early genetic algorithm solves problems by simulating the process of biological evolution in nature. After a population in nature has undergone several generations of natural selection, the remaining populations must be adapted to the environment. All solutions of a problem are regarded as a population, and after several natural selections, the remaining solutions are the optimal solutions of the problem. The general process of genetic algorithm is divided into initialization coding, individual evaluation, selection, crossover and mutation. The advantage of the heuristic algorithm is that it overcomes the defect that the gradient method is sensitive to the initial value and easy to fall into the local optimal value, and has more advantages in global exploration. But also because of its randomness, its convergence speed is slower, and there are certain limitations in local search.

For the mangrove ecosystem, the monitoring data involves time, space and its numerical changes. At the same time, the prediction model is a nonlinear model with high complexity and incompleteness. The model obtained by using the neural network structure and the above training method is consistent with the The reality is quite different, and the prediction accuracy is low. Therefore, a new method or model is used for the prediction of mangrove ecosystem health.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a mangrove ecological health prediction training method based on a dynamic group optimization algorithm, provide a scientific training model, and improve the accuracy of the prediction data and the integrity of the training model.

The technical scheme adopted by the present invention to solve the technical problem is: a mangrove ecological health prediction training method based on a dynamic group optimization algorithm, and the method comprises the following steps:

Step X, establishing a mangrove ecological data set;

Step A, data preprocessing, data generation based on the algorithm of particle swarm optimization and information entropy;

Step A1, based on the data generation module of PSO algorithm;

Step A2, a similarity evaluation module based on relative entropy and cosine similarity;

In step B, model training is performed based on the DGO algorithm, and its model includes a population, and the population includes several groups consisting of centroids and members;

Step B1, intra-group cooperation, which is used for member update to combine members with the greatest fitness and obtain an optimal solution, and this member update includes normal update and mutation;

Step B2, the centroid of each group communicates with the centroids of other groups;

Step B3, the members of each group randomly cross with other members in the population;

Step B4, in steps B2 and B3, search for the mutation situation, find out the member corresponding to this based on the mutation situation, and transfer the member to other groups at random;

Step C, combine the BP algorithm and the DGO algorithm with the step B to train the feedforward neural network.

Preferably, the training process in the step C includes the following steps:

Step C1, based on the DGO algorithm, calculate the fitness between individual members in each group in the population through inter-group communication and intra-group cooperation. The calculation is based on the BP algorithm, which includes first parameter initialization, and then gradient descent. training, and finally the fitness is obtained by calculating the loss value of the function;

Step C2, update the centroid of each group based on step B4;

Step C3, whether the iteration termination condition is satisfied, if yes, obtain relevant parameters and fitness according to the current data information; if not, return to Step C1 and repeat the foregoing steps in turn until the iteration termination condition is satisfied.

Preferably, the step A2 also includes calculating the fitness between each individual member in the population, and updating the individuals in the population according to the fitness, obtaining the individual with the largest fitness, and adding the result to the data set to This dataset is evaluated, and the evaluation process includes:

S1, encode the dataset into populations

P=(g ₁ , g ₂ , ..., g _n )

g=(x ₁ , x ₂ , . . . , x _m )

Among them, P is the population, g is the group in the population, n is the set number of population groups, and m is the length of the individuals in the group;

S2, the population is expanded by the PSO algorithm. The PSO includes a particle corresponding to each individual, and the particle exists in the d-dimensional space. During the population expansion process, it includes the position, velocity and The historical optimal position that it has passed through, and the optimal position that the population has passed through, the above expressions are as follows:

the position of particle i, x _i = (x _i1 , x _i2 ,..., x _id ), i=1, 2,..., m;

Velocity of particle i, v _i =(v _i1 , v _i2 , . . . , v _id ), i=1, 2, . . . , m;

The position of the optimal point passed by the particle _i , pi =( _pi1 , _pi2 ,..., _pid ), i=1, 2,...,m;

The optimal point position passed by the population, p _g = (p _g1 , p _g2 , . . . , p _gd );

The position of each particle is updated with the current position and velocity, and the update formula is as follows:

The velocity from time t to time t+1,

The position from time t to time t+1,

in, is the optimal point position passed by the current particle, is the optimal point position passed by the current particle, ω is the inertia weight, c ₁ c ₂ is the learning factor, r ₁ r ₂ is a random number between [0, 1];

S3, evaluate the expanded new individual, which includes the calculation of relative entropy and cosine similarity, and the calculation formula is as follows:

f(PO, _PG ) _{= fE} (PO, _PG ) _{+ fK} (PO, _PG ) _{+ fC} ( _PO , _PG )

Among them, PO and _PG represent the original mangrove ecological monitoring data and generated data, respectively, and f( _PO , _{PG )} represents the similarity;

f _E (P _O , P _G )=|H(P _G )-H(P _o )|

Among them, H(P _G ) represents the information entropy of the generated data, and H(P _o ) represents the information entropy of the original data;

and respectively represent the probability value of PO and _PG at the i _- th position;

P _G (X _i ) and P _O (X _i ) represent the values of P _O and P _G at the i-th position, respectively;

Steps S2 and S3 are sequentially repeated to cooperate with the step C3 until the iteration termination condition is satisfied.

Preferably, the step B1 includes updating the formula:

Among them, x _{i, j, k} represent the j-th member of the i-th group of the k-th generation, w is the weight that determines the moving direction, r is a random number between 0 and 1, C _i is the centroid of the i-th group, G _best is the global optimal value, μ is a random number obeying μ～(0, s ² ), and s is the step size.

Preferably, in the step B2, the communication between the centroid of each group and the centroids of other groups includes the movement of the centroid, and the movement of the centroid adopts a Lévy random walk, and the formula of the Lévy random walk is as follows:

where α is the step size, represents the centroid of the i-th group of the k-th generation, Representing an entry-wise multiplication, Lévy(λ) is a random number that obeys the Lévy distribution.

Preferably, the step B3 also includes randomly selecting a crossover operator and a biased random walk operator, and selects it with the following formula:

random selection operator,

Biased random walk operator,

where r is the random number generator and Cr is the crossover probability.

Preferably, in the step C, MSE is used as the learning error function in the BP algorithm, and its calculation formula is:

The fitness calculation formula in the DGO algorithm is:

in, represents the jth output value of the ith training sample, y _{i, j} represents the jth expected output value of the ith training sample, n is the number of training samples, and m is the output dimension.

The beneficial effects of the invention are as follows: through the combination of the BP algorithm and the DGO algorithm, the prediction model training of the mangrove ecological health data with high complexity is realized; The original data generates simulation data with a controllable scale, thus providing reliable data support for the training model; through data preprocessing, the data generation efficiency is improved, and at the same time, the incompleteness of traditional data sets after collection and the inaccuracy of prediction and classification are eliminated. It effectively solves the problem and enhances the anti-overfitting ability of the training model; the feedforward neural network is trained by the BP algorithm and the dynamic group optimization algorithm, and the advantages of the two algorithms are combined, which not only avoids falling into the local optimum point during global exploration Under the conditions of the local search, the convergence speed is faster, and the complex training model can be effectively trained, so that it can also be well applied in the field of data mining technology outside the field of mangrove ecological protection technology, which enhances the practicality. sex.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the present invention will be further described below with reference to the accompanying drawings and embodiments. Ordinary technicians can also obtain other drawings based on these drawings without creative labor:

Fig. 1 is a kind of training flow chart of training model in the mangrove ecological health prediction training method based on dynamic group optimization algorithm of the present invention;

Fig. 2 is a kind of global best quality search schematic diagram in a kind of mangrove ecological health prediction training method based on dynamic group optimization algorithm of the present invention;

Fig. 3 is the transfer schematic diagram of group master variation in a kind of mangrove ecological health prediction training method based on dynamic group optimization algorithm of the present invention;

Fig. 4 is a list of minimum values and mean values of LF and CA on the training data set in a kind of mangrove ecological health prediction training method based on dynamic group optimization algorithm of the present invention;

Fig. 5 is the list of the minimum value and the mean value of LF and CA on the test data set in a kind of mangrove ecological health prediction training method based on dynamic group optimization algorithm of the present invention;

6 is a schematic diagram of a fitness convergence curve reflecting how fitness changes with the number of iterations in a mangrove ecological health prediction training method based on a dynamic group optimization algorithm of the present invention;

Fig. 7 is a kind of data set characteristic table in the mangrove ecological health prediction training method based on dynamic group optimization algorithm of the present invention;

Fig. 8 is a data set label table in a mangrove ecological health prediction training method based on the dynamic group optimization algorithm of the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the following will be described clearly and completely in combination with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, and Not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

It should be noted that "step A1" and "step A2" described in this application belong to the two parts of "step A", and there is no directional relationship between them. The specific process relationship is subject to the embodiments in the description. In addition, the order in which they are implemented does not affect data preprocessing.

In embodiment 1, as shown in Figure 1, a kind of mangrove ecological health prediction training method based on dynamic group optimization algorithm, described method comprises the following steps:

Step X, establishing a mangrove ecological data set;

Step A, data preprocessing, data generation based on the algorithm of particle swarm optimization and information entropy;

Step A1, based on the data generation module of PSO algorithm;

Step A2, a similarity evaluation module based on relative entropy and cosine similarity;

In step B, model training is performed based on the DGO algorithm, and its model includes a population, and the population includes several groups consisting of centroids and members;

Step B1, intra-group cooperation, which is used for member update, to combine members with the greatest fitness and obtain an optimal solution, and this update includes normal update and mutation;

Step B2, the centroid of each group communicates with the centroids of other groups;

Step B3, the members of each group are randomly crossed with other members in the population;

Step B4, in the communication between each centroid and the random crossover process of each member, search for the mutation situation, find out the corresponding member based on the mutation situation, and randomly transfer the member to other groups;

Step C, combine the BP algorithm and the DGO algorithm with the step B to train the feedforward neural network.

First, in step X, a mangrove data set is established. As shown in the figure, the mangrove data set includes different features, each of which corresponds to a predictor: annual change in salinity, pH value, active phosphate/(μg/L) , inorganic nitrogen/(μg/L), mercury/(μg/g), cadmium/(μg/g), lead/(μg/g), arsenic/(μg/g), oil/(μg/g) , sulfide/(μg/g), organic carbon (%), phytoplankton density/(x10 ⁵ /m ³ ), zooplankton density/(x10 ³ /m ³ ), zooplankton biomass/(mg/ m ³ ), benthic density/(individuals/m ² ) and benthic biomass/(g/m ² ), the data corresponding to the aforementioned indicators are 1.534, 9.25, 11.833, 298.333, 0.12, 0.119, 1.6667 , 1.0067, 48.333, 44.7, 0.96, 2.39, 0.813, 27.68, 360, and 365.35. Create a feature set X from this data:

Among them, n is the number of samples, m is the number of features, and a sample contains more than one feature. As shown in Figure 8, labels are divided according to the index data. The index below the data value is classified as the label "0", the index equal to the data value is classified as the label "1", and the index above the data value is classified as the label: "2", The level corresponding to the label "0" is healthy, the level corresponding to the label "1" is sub-health, and the level corresponding to the label "2" is unhealthy. Finally, the output of the prediction result is performed according to the label table, and the output label set is Y:

Specifically, in the step A, the algorithm of particle swarm optimization and information entropy is used to generate data, so as to generate a large-scale data set similar to the original data distribution by simulating a small amount of original mangrove ecological monitoring data, and form A trained model containing a d-dimensional space. In the two algorithms combined in step A1 and step A2, all the original data are in a data population, and the data population includes the individual corresponding to each unit of data. Among them, all individuals in the group are randomly initialized, and each individual constitutes a piece of generated data. In its algorithm, it includes an iterative mechanism. In each iteration process, the mutual fitness of each individual is first calculated, including the setting of the fitness function, which refers to the similarity evaluation function; Update the population individuals with the size of the fitness, and obtain the individual with the highest fitness and the fitness; finally, the individual with the highest fitness is used as the result and added to the data set.

The PSO algorithm is a particle swarm optimization algorithm. The PSO algorithm includes particle clusters. The particle clusters are composed of large-scale particles, and each particle corresponds to an individual in the population. Among them, each particle searches for the optimal point position in the d-dimensional space, and each particle matches the fitness function to determine the fitness value to judge the proximity of the current position to the optimal point. The particle has an optimal point memory, and during its search, it records the location of the optimal point that it passes through. The speed of each particle in the search process will determine the distance and direction of its motion, and the speed is dynamically adjusted according to its own motion habits and the motion habits of accompanying particles. The motion habit is the speed that the particle has taken to reach the optimal point position in the past, and the corresponding relationship with the optimal point position that the population has passed through in the past, and the training experience formed by simulating its inertial parameters.

The DGO algorithm is a dynamic swarm optimization algorithm, which is mainly calculated by combining the swarm algorithm and the evolutionary algorithm. Among them, the swarm algorithm is used for global particle exploration; the evolutionary algorithm is used for local particle optimal point development. In the training model of this calculation process, including the centroid and members of each group consisting of two parts, several groups can form a population, wherein the centroid of each group records the optimal solution found by the group, that is, the relationship with the individual. The corresponding maximum fitness.

In the step B4, as shown in Fig. 3, the mutation situation belongs to the group mutation, and the group mutation is set to save computing resources and avoid falling into a local optimum. Among them, if a member of a certain group cannot improve itself within a given number of attempts, it is randomly transferred to other groups. Compared with other heuristic algorithms, the DGO algorithm is more efficient for training neural networks because it considers both global exploration and local exploitation. The above four-step process ensures these good performances, effectively avoiding the problems of slow convergence, sensitivity to initial values, and easy to fall into local optimum when solving complex problems.

Further, the training process in the step C includes the following steps:

Step C1, based on the DGO algorithm, calculate the fitness between individual members in each group in the population through inter-group communication and intra-group cooperation. The calculation is based on the BP algorithm, which includes first parameter initialization, and then gradient descent. training, and finally the fitness is obtained by calculating the loss value of the function;

Step C2, update the centroid of each group based on step B4;

Step C3, whether the iteration termination condition is satisfied, if yes, obtain relevant parameters and fitness according to the current data information; if not, return to Step C1 and repeat the foregoing steps in turn until the iteration termination condition is satisfied.

Among them, in the initialization phase, the parameters will be initialized. Some parameters are preset with thresholds, such as population size and the number of members in each group; while other parameters are randomly generated, such as the starting point of the population, that is, the initial state of all individuals. In each iteration, the individual obtained by the DGO algorithm is used as the initial value of the parameters of the BP algorithm, and then the fitness is obtained through the training of the BP algorithm, and then the DGO algorithm will update the population according to the fitness of all individuals. Iterates in this way to get the final optimal parameters and optimal fitness. The iterative termination condition is to combine the members with the largest fitness and obtain the optimal solution, that is, the completion of the update of members in the group.

Further, the step A2 also includes calculating the fitness between each individual member in the population, and updating the individuals in the population according to the fitness, obtaining the individual with the largest fitness at the same time, and adding the result to the data set. This dataset is evaluated, and the evaluation process includes:

S1, encode the dataset into populations

P=(g ₁ , g ₂ , ..., g _n )

g=(x ₁ , x ₂ , . . . , x _m )

Among them, P is the population, g is the group in the population, n is the set number of population groups, and m is the length of the individuals in the group;

S2, the population is expanded by the PSO algorithm. The PSO includes a particle corresponding to each individual, and the particle exists in the d-dimensional space. During the population expansion process, it includes the position, velocity and The historical optimal position that it has passed through, and the optimal position that the population has passed through, the above expressions are as follows:

the position of particle i, x _i = (x _i1 , x _i2 ,..., x _id ), i=1, 2,..., m;

Velocity of particle i, v _i =(v _i1 , v _i2 , . . . , v _id ), i=1, 2, . . . , m;

The position of the optimal point passed by the particle _i , pi =( _pi1 , _pi2 ,..., _pid ), i=1, 2,...,m;

The optimal point position passed by the population, p _g = (p _g1 , p _g2 , . . . , p _gd );

The position of each particle is updated with the current position and velocity, and the update formula is as follows:

The velocity from time t to time t+1,

The position from time t to time t+1,

in, is the optimal point position passed by the current particle, is the optimal point position passed by the current particle, ω is the inertia weight, c ₁ c ₂ is the learning factor, r ₁ r ₂ is a random number between [0, 1];

S3, evaluate the expanded new individual, which includes the calculation of relative entropy and cosine similarity, and the calculation formula is as follows:

f(PO, _PG ) _{= fE} (PO, _PG ) _{+ fK} (PO, _PG ) _{+ fC} ( _PO , _PG )

Among them, PO and _PG represent the original mangrove ecological monitoring data and generated data, respectively, and f( _PO , _{PG )} represents the similarity;

f _E (P _O , P _G )=|H(P _G )-H(P _o )|

Among them, H(P _G ) represents the information entropy of the generated data, and H(P _o ) represents the information entropy of the original data;

and respectively represent the probability value of PO and _PG at the i _- th position;

P _G (X _i ) and P _O (X _i ) represent the values of P _O and P _G at the i-th position, respectively;

Steps S2 and S3 are sequentially repeated to cooperate with the step C3 until the iteration termination condition is satisfied.

Further, the step B1 includes updating the formula:

Among them, x _{i, j, k} represent the j-th member of the i-th group of the k-th generation, w is the weight that determines the moving direction, r is a random number between 0 and 1, C _i is the centroid of the i-th group, G _best is the global optimal value, μ is a random number obeying μ～(0, s ² ), and s is the step size.

The above calculation process is used for member update, including normal update and mutation. The update process is calculated based on each set of optimal solutions and the global optimal solution. Mutation can ensure the diversity of the population, thus avoiding falling into a local optimum.

Further, the communication between the centroid of each group and the centroids of other groups in the step B2 includes the movement of the centroid, and the movement of the centroid adopts a Lévy random walk, and the formula of the Lévy random walk is as follows:

where α is the step size, represents the centroid of the i-th group of the k-th generation, Representing an entry-wise multiplication, Lévy(λ) is a random number that obeys the Lévy distribution.

The Lévy random walk is suitable for the above calculation process, which provides population diversity and the ability to expand the search range. The number of Lévy walks is updated. During this process, the centroids of each group communicate with the centroids of other groups, thereby improving the It improves communication efficiency and saves communication costs.

Further, the step B3 also includes randomly selecting a crossover operator and a biased random walk operator, and selecting them according to the following formula:

random selection operator,

Biased random walk operator,

where r is the random number generator and Cr is the crossover probability.

In the above process, each member crosses with other members of the population with random probability. This crossover action is an efficient global optimization method that aims to quickly solve problems stuck in local optima.

Further, described step C uses MSE as the learning error function in the BP algorithm, and its calculation formula is:

The fitness calculation formula in the DGO algorithm is:

in, represents the jth output value of the ith training sample, y _{i, j} represents the jth expected output value of the ith training sample, n is the number of training samples, and m is the output dimension.

In the above training process, the BP and DGO algorithms are combined as the training method of the feedforward neural network. The DGO algorithm is added to the traditional BP neural network for optimization training. Since the BP neural network has the defects of being sensitive to the initial value and easy to fall into the local optimum, the high randomness and diversity of the DGO algorithm can better make up for the above. defect. The feedforward neural network includes various neurons, and each neuron is arranged in layers, and the neurons have a connection relationship between the upper and lower layers, and the optimization of the feedforward neural network in this embodiment is mainly to find a set of weights for each layer connection and bias to enrich the training parameters, thereby improving the accuracy of the results predicted by the model. Among them, the accuracy of the prediction results is related to the learning loss. Therefore, by setting the learning loss function, the data target is optimized to minimize the loss function. Among them, the weights and biases in the feedforward neural network are converted into a vector, and each vector represents the position of an individual.

Specifically, in this embodiment, four different training models are used for comparison, wherein the DGBPNN model is the training model provided by this application. At the same time, it includes two indicators to evaluate the technical effect, as follows:

1) Quantitative statistical performance:

A. LF: the minimum and average value of the loss function (LF) for 20 independent runs, which is calculated as the fitness formula.

B.CA: The maximum and average value of the health status classification accuracy for 20 independent runs, calculated as

where is the number of correctly classified samples and is the total number of samples used for classification.

As shown in Figures 4 and 5, it can be seen from these two tables that DGBPNN outperforms the other three algorithms in terms of fitting and generalization ability. The average and minimum values of the loss function are the smallest among the three algorithms, and in terms of classification accuracy, the average and optimal values are also the best.

2) The fitness convergence curve that reflects how the fitness changes with the number of iterations: as shown in Figure 6, in contrast, the convergence speed of DGBPNN is the fastest, and its convergence speed is close to that of DGEFNN. Although the convergence speed of the two algorithms is similar, the fitness of DGBPNN is smaller. In the curve position, the curve of DGBPNN is close to PSBNN, but the convergence speed is faster than PSBNN. These results show that DGBPNN has the best performance in solving the prediction training model of mangrove ecosystem health.

Claims (8)

1. a mangrove ecological health prediction training method based on dynamic group optimization algorithm, is characterized in that, described method may further comprise the steps: Step X, establishing a mangrove ecological data set; Step A, data preprocessing, data generation based on the algorithm of particle swarm optimization and information entropy; Step A1, based on the data generation module of PSO algorithm; Step A2, a similarity evaluation module based on relative entropy and cosine similarity; In step B, model training is performed based on the DGO algorithm, and the model includes a population, and the population includes several groups composed of centroids and members; Step B1, intra-group cooperation, which is used for member update to combine members with the greatest fitness and obtain an optimal solution, and the member update includes normal update and mutation; Step B2, the centroid of each group communicates with the centroids of other groups; Step B3, the members of each group randomly cross other members in the population; Step B4, in steps B2 and B3, search for the mutation, find out the member corresponding to the mutation based on the mutation, and randomly transfer the member to other groups; Step C, combine the BP algorithm and the DGO algorithm with the step B to train the feedforward neural network. 2. the mangrove ecological health prediction training method based on dynamic group optimization algorithm according to claim 1, is characterized in that, the training process in described step C comprises the following steps: Step C1, based on the DGO algorithm, calculate the fitness between individual members in each group in the population through inter-group communication and intra-group cooperation. The calculation is based on the BP algorithm, which includes first parameter initialization, and then gradient descent. training, and finally the fitness is obtained by calculating the loss value of the function; Step C2, update the centroid of each group based on step B4; Step C3, whether the iteration termination condition is satisfied, if so, obtain the relevant parameters and fitness according to the current data information; if not, return to Step C1 and repeat the foregoing steps in sequence until the iteration termination condition is satisfied. 3. The mangrove ecological health prediction training method based on the dynamic group optimization algorithm according to claim 1 or 2, is characterized in that, described step A2 also comprises calculating the fitness between each individual member in the population, and according to The fitness updates the individuals in the population, and at the same time, the individual with the maximum fitness is obtained, and the result is added to the data set to evaluate the data set. The evaluation process includes: S1, encode the dataset into populations P=(g ₁ , g ₂ , ..., g _n ) g=(x ₁ , x ₂ , . . . , x _m ) Among them, P is the population, g is the group in the population, n is the set number of population groups, and m is the length of the individuals in the group; S2, the population is expanded by the PSO algorithm. The PSO includes a particle corresponding to each individual, and the particle exists in the d-dimensional space. During the population expansion process, it includes the position, velocity and The historical optimal position that it has passed through, and the optimal position that the population has passed through, the above expressions are as follows: the position of particle i, x _i = (x _i1 , x _i2 ,..., x _id ), i=1, 2,..., m; Velocity of particle i, v _i =(v _i1 , v _i2 , . . . , v _id ), i=1, 2, . . . , n; The position of the optimal point passed by the particle _i , pi =( _pi1 , _pi2 ,..., _pid ), i=1, 2,...,m; The optimal point position passed by the population, p _g = (p _g1 , p _g2 , . . . , p _gd ); The position of each particle is updated with the current position and velocity, and the update formula is as follows: The velocity from time t to time t+1, The position from time t to time t+1, in, is the optimal point position passed by the current particle, is the optimal point position passed by the current particle, ω is the inertia weight, c ₁ c ₂ is the learning factor, r ₁ r ₂ is a random number between [0, 1]; S3, evaluate the expanded new individual, which includes the calculation of relative entropy and cosine similarity, and the calculation formula is as follows: f(PO, _PG ) _{= fE} (PO, _PG ) _{+ fK} (PO, _PG ) _{+ fC} ( _PO , _PG ) Among them, PO and _PG represent the original mangrove ecological monitoring data and generated data, respectively, and f( _PO , _{PG )} represents the similarity; f _E (P _O , P _G )=|H(P _G )-H(P _o )| Among them, H(P _G ) represents the information entropy of the generated data, and H(P _o ) represents the information entropy of the original data; and respectively represent the probability value of PO and _PG at the i _- th position; P _G (X _i ) and P _O (X _i ) represent the values of P _O and P _G at the i-th position, respectively; Steps S2 and S3 are sequentially repeated to cooperate with the step C3 until the iteration termination condition is satisfied. 4. the mangrove ecological health prediction training method based on dynamic group optimization algorithm according to claim 1 and 2, is characterized in that, described step B1 comprises update formula: Among them, x _{i, j, k} represent the j-th member of the i-th group of the k-th generation, w is the weight that determines the moving direction, r is a random number between 0 and 1, C _i is the centroid of the i-th group, G _best is the global optimal value, μ is a random number obeying μ～(0, s ² ), and s is the step size. 5. the mangrove ecological health prediction training method based on dynamic group optimization algorithm according to claim 4, is characterized in that, the calculation formula of described step size is as follows: Among them, Γ is the gamma function, and β is a constant. 6. The mangrove ecological health prediction training method based on dynamic group optimization algorithm according to claim 1 or 2, is characterized in that, in described step B2, the centroid of each group communicates with the centroid of other groups and includes the movement of centroid , its center of mass moves using a Lévy random walk, and the formula for the Lévy random walk is as follows: where α is the step size, represents the centroid of the i-th group of the k-th generation, Representing an entry-wise multiplication, Lévy(λ) is a random number that obeys the Lévy distribution. 7. The mangrove ecological health prediction training method based on the dynamic group optimization algorithm according to claim 1 or 2, wherein the step B3 further comprises randomly selecting a crossover operator and a biased random walk operator, and Select with the following formula: random selection operator, Biased random walk operator, where r is the random number generator and Cr is the crossover probability. 8. the mangrove ecological health prediction training method based on dynamic group optimization algorithm according to claim 1 and 2, is characterized in that, described step C uses MSE as the learning error function in the BP algorithm, and its calculation formula is: The fitness calculation formula in the DGO algorithm is: in, represents the jth output value of the ith training sample, y _{i, j} represents the jth expected output value of the ith training sample, n is the number of training samples, and m is the output dimension.