CN101599138A - Land evaluation method based on artificial neural network - Google Patents

Abstract

The invention discloses a land evaluation method based on an artificial neural network. The method is based on a self-learning method based on actual survey samples or on the basis of existing knowledge, and performs self-learning corrections based on samples to construct a self-learning and self-adaptive A Neural Network Approach to Land Evaluation. According to the problem that the neural network model converges too slowly or even diverges due to the existence of non-differentiable excitation functions such as step functions in the model structure, genetic optimization is introduced, and a land evaluation method based on genetic optimization is constructed to realize the neural network based on genetic optimization. methods of land evaluation. The invention uses genetic algorithm to optimize the connection weight of neural network and genetic algorithm to optimize neural network structure to improve the accuracy and practicability of neural network model.

Description

Land Evaluation Method Based on Artificial Neural Network

Technical field:

The invention relates to a land evaluation method based on genetically optimized neural network self-learning technology, which belongs to the field of land investigation and evaluation.

Background technique

Land evaluation is the basis of land use planning and an important prerequisite for rational use of land. Since the 1960s, land evaluation has been widely concerned, and its theory, technology and application have been developed rapidly. The land evaluation model is the core of land evaluation, and it has always been a hot spot in land evaluation research, and has been widely concerned at home and abroad. In general, its development has gone through a process from qualitative to quantitative, from individual to comprehensive, from mathematical statistical analysis to intelligent computing, complex geographic computing and expert systems, forming qualitative models, statistical methods, parameterized systems, expert systems, and hybrid methods and other evaluation techniques. Most of the traditional land evaluation methods rely on empirical knowledge. According to the use of empirical knowledge and the complexity of reasoning, they can basically be divided into three categories: ① simple fitting method. Through empirical analysis, a function form is given in advance to simulate the relationship between the participating factors and land quality, and then regression analysis is carried out based on a small number of survey samples to obtain the fitting formula. This method is generally called regression analysis. ② empirical rule reasoning. According to the experience of experts, the judgment rules are formulated, including the importance degree of factors or restrictive factors and reasoning methods, etc., and then the restriction judgment or weighted judgment is carried out. Such as limit condition method, empirical index sum method, etc. ③ Land evaluation expert system method. Establish a knowledge base based on expert experience, and transform it into a rule representation, establish a reasoning rule base, and use a comprehensive and more reasonable reasoning method to conduct a comprehensive evaluation of land quality. The methods used usually include fuzzy comprehensive evaluation method, gray system method and so on. The main problems of the above methods are: the simple fitting method is too simplistic in approximating the relationship between the participating factors and land quality, and the given functional form is difficult to describe the complex relationship; Accuracy is extremely sensitive, and inaccurate knowledge often leads to biased results. Therefore, the accuracy and reliability of the evaluation results of these two types of methods are poor. The evaluation system depends on the existing knowledge and rules, and cannot make adjustments to the incompleteness of knowledge, and has no self-learning ability; the system is generally only suitable for a specific area at a specific time, and is not adaptive and universal; the use of forms In generalized and strict deterministic reasoning, the misestimation of individual parameters may lead to serious deviations in the results, which is not fault-tolerant. Artificial neural network is a method system for constructing an information processing system closer to human intelligence by imitating the organizational structure and operating mechanism of the brain and designing a new computer processing structure model. It adopts a data-driven mechanism. These methods have the characteristics of self-learning, self-organization, self-adaptation, versatility and robustness. In solving the problem of traditional land evaluation relying on expert experience, the evaluation results are greatly affected by subjective factors, and the reliability is poor. The advantages.

Contents of the invention

In view of the traditional land evaluation method, the approximation of the relationship between the participating factors and the land quality is oversimplified by the simple fitting method, and the given functional form is difficult to describe the complex relationship between the land quality and its influencing factors; the reasoning method of empirical rules The process is simple, but it is extremely sensitive to the accuracy of empirical knowledge, and inaccurate knowledge often brings defects such as results with large deviations. The invention introduces the theory of computational intelligence into land evaluation, and changes the land evaluation method from the traditional multi-factor evaluation method relying solely on empirical knowledge to a self-learning method based on actual survey samples or self-learning based on existing knowledge and samples. Learning and correcting methods, constructing a self-learning and self-adaptive land evaluation neural network method, and introducing genetic optimization algorithms to promote the quantification and intelligence of land evaluation.

The present invention uses the artificial neural network method to evaluate the land, which is a completely different idea from the traditional method based on knowledge and mathematical logic reasoning. The artificial neural network method of land evaluation is based on sample data rather than experience and knowledge, and adopts machine learning methods to automatically fit the nonlinear relationship between participating factors and land quality, which is self-learning. At the same time, according to the convergence of the neural network model is too slow or even divergent due to the existence of non-differentiable excitation functions such as step functions in the model structure, genetic optimization is introduced, and a land evaluation method based on genetic optimization is constructed to realize the neural network land evaluation method based on genetic optimization. evaluation method. The main content of the invention includes optimizing the connection weight of the neural network with the genetic algorithm and optimizing the neural network structure with the genetic algorithm to improve the accuracy and practicability of the neural network model, wherein optimizing the connection weight of the neural network with the genetic algorithm includes the following steps:

(1) Arrange the connection weights and thresholds of the neural network in a certain order (such as descending order or ascending order), and use a binary coding scheme to encode, randomly generate a set of distributions, and then construct a set of code chains, each code chain represents A weight distribution of a neural network, that is, a neural network corresponding to a specific value of weight and threshold;

(2) Calculate its mean square error on the training sample set for the generated neural network, thereby determining its fitness. The larger the mean square error, the smaller the fitness;

(3) Select a number of individuals with the largest fitness value to directly inherit to the next generation;

(4) Use the genetic operator of crossover and mutation to process the current generation population to generate the next generation population;

(5) Repeat steps (2) (3) (4) to make the initially determined set of weight distributions evolve continuously until the training objective is met.

The steps to train and optimize the neural network structure with genetic algorithm are:

(1) Randomly generate N structures, encode each structure, and each coded individual corresponds to a neural network structure;

(2) train the structure in the individual set with many different initial weight distributions;

(3) Determine the fitness of each individual according to the training results or other strategies;

(4) Select a number of individuals with the largest fitness values and directly pass them on to the next generation;

(5) Perform crossover and mutation genetic operations on the current generation population to generate the next generation population;

(6) Repeat (2)-(5) until an individual in the current generation group can meet the requirements.

Compared with the traditional land evaluation method, the present invention has the following advantages:

(1) The various evaluation models based on the theory of computational intelligence proposed by the present invention have completely changed the shortcomings of traditional models such as relying on empirical knowledge, being greatly influenced by human beings, and not having wide applicability, so that the land evaluation model has self-learning The accuracy and practicality of the model are greatly improved. This type of model conforms to people's logical habits and is easy to understand.

(2) The construction of the genetic neural network model for land evaluation is simple; the initial rules are corrected by self-learning; the convergence of genetic training is good, the accuracy of the results is high, and the rules obtained by self-learning can be completely extracted.

(3) The neural network method based on genetic optimization will also improve the degree of automation, result reliability, robustness and wide adaptability of the evaluation system.

Description of drawings

Fig.1 Structure diagram of genetic neural network for land evaluation

Fig. 2 The change curve of the optimal individual fitness value and the average value of group fitness (the maximum number of iterations is 100).

Fig. 3 is the curve diagram of the optimal individual fitness value and the average value of group fitness (the maximum number of iterations is 500).

Figure 4 Distribution of sample points and distribution of evaluation results before and after genetic training.

Figure 5 and Figure 6 are schematic diagrams of the genetic training of the fuzzy neural network from two improved initial states.

Detailed ways

Utilize the land evaluation method based on genetic optimization constructed by the present invention, its structure is as shown in accompanying drawing 1, the evaluation method proposed carries out evaluation practice to the paddy field suitability of a certain area, and its initial evaluation index is as follows:

Evaluation factor index system of initial suitable paddy fields

Optimizing the connection weight of neural network with genetic algorithm includes the following steps:

(1) Arrange the connection weights and thresholds of the neural network in a certain order and encode them with a binary coding scheme, randomly generate a set of distributions, and then construct a set of code chains, each code chain represents a weight of the neural network Distribution, that is, a neural network corresponding to a specific value of weight and threshold;

(2) Calculate its mean square error on the training sample set for the generated neural network, thereby determining its fitness. The larger the mean square error, the smaller the fitness;

(3) Select a number of individuals with the largest fitness value to directly inherit to the next generation;

(4) Use the genetic operator of crossover and mutation to process the current generation population to generate the next generation population;

(5) Repeat steps (2) (3) (4) to make the initially determined set of weight distributions evolve continuously until the training objective is met.

In the above step (1), the weights and thresholds in the neural network are encoded using a binary coding scheme. Each connection weight of the neural network is represented by a certain length of 0/1 string, and the threshold is regarded as the input is -1 The connection weight of each connection weight has the following relationship between the string representation value and the actual weight:

${\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; bin</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>$

Among them, W _i (i, j) represents the actual weight, bin (t) is a binary integer represented by an l-bit string, [W _min (i, j), W _max (i, j)] is the weight of each connection The change range of the weight, all the corresponding 0/1 strings are cascaded together, and the obtained binary string represents a weight distribution of the network.

The fitness function of the above neural network is F( _xi )=CE( _xi ) or F(xi ₎ =1/E( _xi ), where _xi represents the chromosome individual, and F(xi ₎ represents the fitness function, E( _xi ) is the mean square error, and C is a constant; for the neural network described by a certain chromosome individual, the mean square error of the training sample is expressed as:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

In the formula, _xi represents the chromosome individual, E( _xi ) is the mean square error, d _m ⁽ⁿ⁾ and z _m ⁽ⁿ⁾ represent the expected output and actual output of the nth sample in the mth output neuron, respectively, M is the number of neurons in the output layer, and N is the number of samples.

The steps to train and optimize the neural network structure with genetic algorithm are:

(1) Randomly generate N structures, encode each structure, and each coded individual (chromosome in the genetic algorithm) corresponds to a neural network structure.

(2) Train the structure in the individual set with many different initial weight distributions (using genetic algorithm or BP algorithm).

(3) Determine the fitness of each individual according to the training results or other strategies.

(4) Select a number of individuals with the highest fitness value and directly pass them on to the next generation.

(5) Perform genetic operations such as crossover and mutation on the current generation population to generate the next generation population.

(6) Repeat (2)-(5) until an individual in the current generation group (corresponding to a network structure) can meet the requirements.

The total number of samples collected in this area is 200, including 140 training set samples and 60 test set samples. The aforementioned genetic algorithm is used for training. The initial population is generated by adding random values based on the initial state of the fuzzy neural network, and the population size is 20. The exchange probability is taken as 0.7, and the mutation operation is performed on the unexchanged individuals with a probability of 1.0. In order to narrow the search space and speed up the convergence, the crossover and mutation operations are performed with the logical subset as the basic unit according to the aforementioned method. Accompanying figure 2 and figure 3 are the change curves of the optimal individual fitness value and the group average fitness value with evolution algebra in training. The upper curve is the change curve of the optimal individual fitness value, and the lower curve is the change curve of the group average fitness value; Figure 2 and Figure 3 are the cases where the maximum number of iterations is set to 100 and 500, respectively.

After the training is completed, the adjusted rule base can be derived from the fuzzy neural network (optimal chromosome). Since the rule base is obtained by "learning" and adjusting from the actual survey samples on the basis of the initial rule base, it is more in line with The actual situation. The adjusted factor index system is shown in the table below. Attached Figure 3 is an example of sample distribution in a typical region and a comparison of results before and after training.

Adjusted suitable paddy field evaluation factors and index system

It can be seen from Figure 3 that in the evaluation results before training (using the initial index system evaluation), many samples fall in the level area inconsistent with their level. There are two main situations, one is at the junction of level transformation , and the second is that there are heterogeneous samples with sporadic distribution in the continuous level area. This reflects the differences in the indicator system and survey results, that is, the incompleteness and inaccuracy of empirical knowledge. This situation is improved and corrected after training, and the evaluation results and survey samples maintain a high consistency. In order to study the influence of the initial rules on the genetic training algorithm, the initial factor index system is slightly improved according to the previous training results, and then the genetic training is carried out. It can be seen that the convergence speed of genetic training is accelerated, and accompanying drawing 4 shows the situation that the fuzzy neural network starts genetic training from two improved initial states. This shows that although genetic training does not depend on the initial rules and has global convergence, the quality of the initial rules has an impact on the training speed. The closer the initial rules are to the objective laws (the mapping represented by the sample set), the faster the training speed . Therefore, obtaining more objective initial rules based on empirical knowledge in practice is beneficial to speed up the convergence speed.

Claims (9)

1. land evaluation method based on artificial neural network, it is characterized in that: based on the self-learning method of factual survey sample or based on existing knowledge, carry out the self study correction according to sample, make up land evaluation method based on self study, adaptive neural network.

2. according to the described land evaluation method of claim 1, it is characterized in that: with the connection weight of genetic algorithm optimization neural network and the accuracy and the practicality of genetic algorithm optimization neural network structure raising neural network model based on artificial neural network.

3. according to the described land evaluation method of claim 2, it is characterized in that may further comprise the steps with the connection weight of genetic algorithm optimization neural network based on artificial neural network:

(1) each connects weights and threshold value according to certain series arrangement and adopt the binary coding scheme to encode to neural network, produce a component cloth at random, and then construct a group code chain, on behalf of a kind of weights of neural network, each yard chain distribute, i.e. neural network of getting particular value corresponding to weights and threshold value;

(2) neural network that is produced is calculated its square error on training sample set, thereby determine its fitness, square error is big more, and fitness is more little;

(3) select the individual directly heredity of some fitness value maximums to of future generation;

(4) utilize to intersect and the genetic operation operator of variation is handled current generation colony, produce colony of future generation;

(5) repeating step (2) (3) (4) distributes initial one group of weights determining and is constantly evolved, till training objective is met.

4. according to the described land evaluation method of claim 3 based on artificial neural network, it is characterized in that: the weights in the neural network and threshold value are encoded adopt the binary coding scheme, each of neural network connects weights and all uses 0/1 string list of a fixed length to show, threshold value is to be input as-1 connection weights, and following relation is arranged between the string table indicating value of each connection weight and the actual weights:

$W_i(i,j)=W_{\min }(i,j)+\frac{\mathrm{bin}\left(t\right)}{2^l-1}\&times;[W_{\max }(i,j)-W_{\min }(i,j)+1]$

Wherein, W _i(i j) represents actual weights, and bin (t) is by the represented bigit of l position character string, [W _Min(i, j), W _Max(i, j)] be the variation range of each connection weight, 0/1 tandem of all correspondences is associated in together, on behalf of a kind of weights of network, the string of binary characters that obtains just distribute.

5. according to the described land evaluation method based on artificial neural network of claim 3, it is characterized in that: the fitness function of neural network is F (x _i)=C-E (x _i) or F (x _i)=1/E (x _i), wherein, x _iExpression chromosome individuality, F (x _i) the expression fitness function, E (x _i) be square error, C is a constant; For the individual described neural network of certain chromosome, the square error of training sample is expressed as:

$E\left(x_i\right)=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}\frac{1}{2}{(d_m^{\left(n\right)}-z_m^{\left(n\right)})}^2$

In the formula, x _iExpression chromosome individuality, E (x _i) be square error, d _m ⁽ⁿ⁾And z _m ⁽ⁿ⁾Represent desired output and the actual output of n sample at m output neuron respectively, M is the output layer neuron number, and N is a number of samples.

6. according to the described land evaluation method of claim 2, it is characterized in that with the step of genetic algorithm optimization neural network structure being based on artificial neural network:

(1) produce N structure at random, to each structured coding, each encodes individual, i.e. the corresponding neural network structure of chromosome in the genetic algorithm;

(2) structure of individuality being concentrated with different initial weight distributions is trained;

(3) result according to training determines the fitness that each is individual;

(4) individuality of the some fitness value maximums of selection directly entails the next generation;

(5) genetic manipulation that current generation colony is intersected and makes a variation is to produce colony of future generation;

(6) repeat (2)～(5), till certain physical efficiency in current generation colony meets the demands.

7. according to the described land evaluation method of claim 6, it is characterized in that: calculate the selected probability of a certain chromosome with following formula based on artificial neural network

P _c＝f(x _i)/∑f(x _i)

In the formula, P _cRepresent the probability that a certain chromosome is selected, x _iI chromosome in the expression population, f (x _i) be i chromosomal fitness value, ∑ f (x _i) be all chromosomal fitness value sums in the population.

8. according to the described land evaluation method of claim 6 based on artificial neural network, it is characterized in that: interlace operation is the exchange to chromosomal logical subsets, promptly select two former generation's chromosomes, with of the logical subsets exchange of one or more exchanged forms with correspondence.

9. according to the described land evaluation method of claim 6 based on artificial neural network, it is characterized in that: the mode that inclined to one side variation weights are arranged is adopted in variation, promptly from initialization probability distributes, get on each numerical value that a series of values are added to logical subsets respectively, guarantee that variation back logical subsets also should satisfy the original numerical value condition.

Images