CN114398951A - Land use change driving factor mining method based on random forest and crowd-sourced geographic information - Google Patents

Abstract

The invention provides a land use change driving factor mining method based on random forest and crowd-sourced geographic information, which comprises the following steps: firstly, constructing a multi-element potential driving factor data set influencing land use change by using multi-source geographic data mainly comprising POI points, and carrying out data spatialization processing; then, constructing a random forest classifier model by taking the multivariate potential driving factors as characteristic variables and taking the land type of the land use thematic map as a prediction variable, and performing model training; secondly, performing K times of random replacement on a single variable by using a trained model so as to calculate the importance score of the variable, and ranking the importance of the driving factors according to the score; and finally, screening the core driving force influencing the land use change by utilizing a recursive characteristic elimination principle. The method can carry out importance quantitative evaluation and core factor screening on the driving factors influencing urban land use change, thereby revealing a microscopic driving mechanism of urban evolution.

Description

A land-use change driving factor mining method based on random forest and crowd-source geographic information

technical field

The invention belongs to the field of geographic big data analysis and mining, in particular to a land use change driving factor mining method based on random forest and crowd source geographic information.

Background technique

As an intuitive manifestation of urban development, the urban land use pattern is driven by human consciousness and under the comprehensive influence of multiple factors such as nature, economy, culture and policy, it has experienced the most complex evolution process on the surface, and has had a profound impact on nature and ecosystems. influences. As the largest developing country in the world, my country is currently in the stage of rapid urbanization. Under the dual pressures of population growth and economic development, land space resources have been greatly developed and utilized, and urban land use has experienced frequent and drastic changes. Mastering the evolution law of urban spatial structure and revealing the microscopic driving mechanism of land use change can provide a scientific reference for government departments to optimize the allocation of urban land resources, and is of great significance to the sustainable development of cities.

The driving factor mining of land use change is the basis for revealing the mechanism, evolution rules and future trend simulation of land use change, and has always been an important direction of land use research. Scholars at home and abroad have carried out a lot of research on the driving forces of land use change. Early research mainly used the form of empirical analysis to reveal the macro-level driving mechanism of a certain land use type in a characteristic area. Moran, for example, points out that forest degradation in the Brazilian Amazon from 1975 to 1987 was largely the result of changes in local government policies on livestock farms, rather than population growth. By comparing grassland changes in China, Russia, and Mongolia between 1992 and 1995, Sneath found that modern animal husbandry was the main cause of grassland degradation. Pulido and Bocco took farmers in developing countries as research objects, and proved that farmers' subjective consciousness and traditional culture play a decisive role in local land degradation. Although these qualitative analyses provide a good foundation for the identification of land-use drivers, it is difficult to assess the extent to which different types of human or natural factors affect land-use change. To this end, statistical methods have been used in subsequent research to conduct quantitative research on factor driving forces, with multiple driving factors as independent variables and land use change as dependent variables to build linear equation models, such as correlation analysis, logistic regression, and linear regression. , principal component analysis, etc.

Considering that my country's urban development is changing from an outward-oriented spatial expansion to a small-scale urban space redevelopment such as land function renewal and old city renovation, it is urgent to study the microscopic driving mechanism of land use change. However, there are two main deficiencies in the existing researches, which are difficult to meet the research on the microscopic driving forces of land use change: first, most of the existing researches focus on the exploration of the macroscopic driving factors of land use change on a large scale. Problems such as inaccurate classification of factors make it difficult to support the discovery of laws and mechanisms for the transformation of urban construction land functions; secondly, urban land use changes are influenced by multiple factors such as natural environment and social economy, and land use functions and structures are affected by high-intensity human activities Under the influence of more complex, traditional statistical models are mostly based on linear equations, which simplifies the relationship model between multiple driving factors and land use change, and cannot truly and comprehensively reflect the complex and nonlinear mapping relationship between the two. .

With the development of Web 2.0 technology, crowdsourced geographic information (Crowedsourcing geographic information), the massive data generated actively or passively by people in daily life, has become an important supplement to professional geographic information data. The depth and breadth of geoscience research has been further enhanced by utilizing the human activities and socio-economic micro-characteristics reflected in crowdsourced geographic information. Point of Interest (POI) data is the most widely used type of crowd-sourced geographic information. A large amount of dynamic and refined socioeconomic information contained in POI tags is applied to urban land use research, in order to explore the microcosm of urban land use. The drive mechanism offers the possibility. However, in practical applications, problems such as high data redundancy and strong information correlation brought about by abundant crowd-sourced geographic information are bound to bring serious interference to the accurate identification and screening of core driving forces. Therefore, it is necessary to establish a driving force analysis method of land use change that is less affected by the correlation between variables, which can not only build a nonlinear model between driving factors and land use change, but also avoid the interference of data redundancy. Accurately identify the dominant and contributing driving factors from the multiple driving factors, and explore the core driving factors affecting urban land use change in a more in-depth and detailed manner. Random forest has a natural advantage in feature screening, and evaluates the importance of feature variables according to their contribution to the model, as the result of analyzing the driving force factors of land use change. Random forest can measure the importance of characteristic variables according to the contribution of variables to the prediction results. Applying the random forest model to build the relationship between land use categories and spatial variables, the importance of variables can be calculated to analyze the effect of land use change. core driver. The random forest model provides a variable importance evaluation method, such as calculating the variable importance MDI (mean decrease in impurity) index according to Gini impurity, which calculates the impact of each variable on the heterogeneity of observations on each node of the classification tree, Therefore, the importance of variables is compared. However, this algorithm may lead to serious deviations in the evaluation of variable importance, mainly because the MDI indicator is a statistical value calculated based on model training data, which cannot fully reflect the contribution of variables to model prediction. Relevant studies have shown that in the case of uneven sample data distribution, when the MDI indicator overfits the model, indicators that do not contribute significantly will be misjudged as important factors. In addition, variable importance evaluations based on Gini impurity tend to assign high values to continuous variables and underestimate the importance of discrete variables. To this end, based on the random forest model and crowd-sourced geographic information, this study introduces a variable permutation test method to evaluate the importance of characteristic variables. Through random permutation of a single variable, the relationship between the variables already established by the model and the prediction target is broken, and then The index calculation is carried out for the change of model error caused by the change of the variable, so as to obtain the importance of a single variable. The advantage of this method is that there is no obvious bias in the importance assessment of discrete and continuous variables, and it can more accurately quantify the importance of driving factors affecting urban land use change and screen core factors.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is: in order to reveal the microscopic driving mechanism of land use change, a land use change driving factor mining method based on random forest and crowd source geographic information is proposed, which accurately reflects the characteristic factors affecting urban land use change. importance, and realize the screening of the core driving factors of land use change.

The present invention solves its technical problem and adopts following technical scheme:

The invention provides a land use change driving factor mining method based on random forest and crowd source geographic information. The method is as follows: firstly, using the crowd source geographic data mainly based on POI points to construct multivariate potential driving factor data affecting land use change Then, using the multivariate potential driving factors as the feature variables and the land use type of the land use thematic map as the predictor variable, build a random forest classifier model, and conduct model training; then, use the trained model , perform K random permutations of a single variable to calculate the importance score of the variable, and sort the importance of driving factors according to the score; finally, use the principle of recursive feature elimination to screen the core driving forces that affect land use change.

The present invention can use the following methods to perform data spatialization processing: according to the data type characteristics of different factors, using various methods such as kernel density estimation, buffer creation, Euclidean distance calculation, partition statistics and slope, slope aspect calculation, etc. It is processed to generate a spatial variable set with consistent resolution and continuous surface shape.

The present invention can adopt the following method to eliminate the difference in the data dimension of the spatial variable set: using the Fuzzy tool in the Arcmap10.2 software to standardize the dispersion of the variable, realize the normalization of the pixel value of the spatial variable, and map the numerical range of the variable to Between 0 and 1.

The present invention can use the following method to construct a random forest classifier model: take multiple potential driving factors as characteristic variables and the land use type of the land use thematic map as a predictor variable to construct a mapping relationship between the two.

The construction of the random forest classifier model described in the present invention can adopt the following methods to collect model training samples:

(1) Import the land use thematic maps of the study area in two different years into ArcMap software, the data space classification rate is 30m, and the land use types are divided into water, forest and grassland, cultivated land, unused land, residential land, industrial land, commercial land There are nine categories of land, public management land and mixed land, and the category codes are numbers from 1 to 9;

(2) Use the Raster calculator in ArcMap software to detect changes in land use types in different years. The algebraic expression is Con("landuse1.tif"!="landuse2.tif", 1, 0) to generate a new raster dataset landuse_difference.tif, a pixel value of 1 represents an area where the land use type has changed, and a pixel value of 0 represents no change;

(3) For the pixels with changes in land use, according to the spatial position index of the pixels, the random traversal sampling method is used to set the corresponding traversal step size for different types of land, perform global search and sampling, and form a training sample set D = [ (x ₁ , y ₁ ), , ..., (x _n , y _n )].

The construction of the random forest classifier model according to the present invention can adopt the following methods for random forest model training: input the training sample D into the random forest model for model training, and set the maximum number of features as the square root of the number N of potential driving factors; The iterative growth of the number of trees is used for model training, and the model training effect is measured by the average interval of model errors:

In the formula: MG _avg represents the average interval of all samples, n represents the number of samples, and mg( _xi , y _i ) represents the interval of a single sample. If mg(x _i , y _i ) is greater than zero, then the correct category occupies the most votes, and the final classification result is correct under voting; otherwise, the final classification result is wrong.

计算新的样本间隔MG_j，对每一个变量重复50次随机置换，以平均值作为变量重要性的最终结果：The present invention can use the following method to calculate and sort the importance score of the driving factor: for each feature variable j in the sample D, randomly replace the value in the variable to generate a new and destroyed training sample

Calculate a new sample interval MG _j , repeating 50 random permutations for each variable, taking the mean as the final result of variable importance:

In the formula: i _j represents the importance score of variable j, MG represents the average interval of the model before random permutation, K is the number of random permutations, MG _{k, j} represents the average interval of the model after the k-th random permutation of variable j;

Using the feature factor importance score obtained in this step, the feature variables are sorted in descending order, that is, the importance ranking of the driving factors of land use change.

The present invention can use the following method to screen the core driving force: using the principle of recursive feature elimination, according to the importance of driving factors, starting from the most important driving factor, adding one factor at a time, forming a new feature subset, and inputting the random In the forest model, the cross-validation method is used to train the model and obtain the new model classification accuracy; the above steps are repeated until all driving factors are included in the feature subset.

The core driving force screening method can use the following method to determine the number of core driving forces: draw a curve that changes the classification accuracy of the model with the decrease in the number of characteristic variables, and find the corresponding point in the curve where the classification accuracy tends to converge, which is the value of the core driving factor. number.

The above-mentioned land use change driving factor mining method based on random forest and crowd-source geographic information provided by the present invention is used to accurately evaluate and screen the microscopic driving factors affecting urban land use change.

Compared with the prior art, the present invention has the following main technical effects:

(1) Aiming at the problems that the research scale is too large and the classification of driving factors is not precise in the existing research on the driving force of land use change, the present invention finds that the introduction of crowd-sourced geographic data containing rich social and economic information, through the method of spatial statistics, the abstract urban The development driving mechanism is transformed into quantitative feature expression in two-dimensional space, and the complete unification of driving factors and land use status is realized in the form of data. The degree of refinement of the land use driving force analysis has laid a good data foundation for revealing the microscopic driving mechanism of urban evolution.

(2) Considering that the traditional statistical model oversimplifies the complex relationship between multiple driving factors and land use change, and it is difficult to truly reflect the microscopic driving mechanism, the present invention designs a land use change driving factor based on random forest when mining the microscopic driving factors of land use. Importance evaluation method, this method reconstructs and changes the random forest model through random permutation of a single variable, and independently examines the influence of each driving factor on the predictive ability of the model, thus effectively avoiding redundant information among multiple factors. Identify adverse effects of core drivers. Compared with the traditional variable importance MDI index in the random forest model, the present invention has the advantage that there is no obvious bias in the importance evaluation of discrete and continuous variables, and the core driving factors can be more accurately screened.

Description of drawings

FIG. 1 is a flow chart of the method of the present invention.

Figure 2 is a graph of spatialization results of potential drivers based on crowdsourced geographic information.

Figure 3 is the iterative training result of the random forest model.

Figure 4 is a ranking diagram of the importance of driving factors.

Figure 5 is a graph of the screening results of core drivers.

Detailed ways

Aiming at the problem that the traditional statistical model oversimplifies the complex relationship between multiple driving factors and land use change, and it is difficult to truly reflect the microscopic driving mechanism, the present invention proposes a land use change driving factor importance evaluation method based on random forest model and crowd-source geographic information , using the crowd-sourced geographic information with rich socioeconomic factors to construct a multivariate latent driver data set of changes, and design a classification interval index to measure the generalization error of the reconstructed model. The importance of eigenfactors. This method independently examines the influence of each driving factor on the predictive ability of the model, thereby effectively avoiding the adverse effect of information redundancy among multiple factors on identifying core driving factors, and solving the traditional random forest algorithm's insufficient response to small variable disturbances. Sensitive puzzles that improve the accuracy of identifying microscopic drivers of land-use change.

The present invention will be further described below with reference to application examples and accompanying drawings, but is not limited to the content described below.

The invention provides a land use change driving factor mining method based on random forest and crowd source geographic information, specifically: first, using crowd source geographic information mainly based on POI point data (such as education, public services and transportation) Construct a dataset of multivariate latent driving factors affecting land use change, and generate data spatial processing; then, with multivariate latent driving factors as feature variables and land use type as predictor variables, build a random forest classifier model, and perform model iterative training ; Next, use the trained model to perform K random permutations of a single variable to calculate the importance score of the variable and rank the importance of the driving factors; finally, use the principle of recursive feature elimination to screen out the core that affects land use change driving force. The invention can accurately carry out quantitative evaluation of importance and screening of core factors for the driving factors affecting urban land use change, so as to excavate the microscopic evolution mechanism of land use change.

In the above method, the following methods can be used for data spatial processing: for the multivariate potential driving factors represented by POI point data, line data, area data and raster data, using kernel density estimation, buffer creation, Euclidean distance calculation, Partition statistics and spatialization methods for slope and aspect calculation are used to spatialize the data (the spatialization methods corresponding to different driving factors are shown in Table 1) to generate a spatial variable set with a resolution of 30m and a continuous surface shape. The implementation of the specific spatialization method adopts the kernel density, Multiple Ring Buffer, euclideanallocation, zonal statistics, slope and aspect tools in the Arcmap10.2 software, import the original feature data into the corresponding tools, and set the output uniformly to a resolution of 30m and tif The resulting graph in the format forms a spatial variable set, that is, a multivariate latent driver data set.

In the above method, the following methods can be used to eliminate the difference in the data dimension of the spatial variable set: using the Fuzzy membership tool in the Arcmap10.2 software to standardize the dispersion of the variables, and import the tif format data of all spatial variables into the tool in turn, Keep the default settings, the exported result image is still in tif format, and the pixel value is converted into a floating-point decimal between 0 and 1, which realizes the normalization of the pixel value of the spatial variable, and eliminates the dimension and data between different variables. level differences.

In the above method, the random forest classifier model can be constructed by the following methods: 20 normalized spatial variables are used as characteristic variables (independent variables), and 9 land use types of the land use thematic map are used as predictor variables (dependent variables), The random forest model is used to construct the mapping relationship between the two.

In the above method, the following methods can be used to collect model training samples:

(1) Import the land use thematic maps of the study area in two different years into ArcMap software, the data space classification rate is 30m, and the land use types are divided into water, forest and grassland, cultivated land, unused land, residential land, industrial land, commercial land There are nine categories of land, public management land and mixed land, and the category codes are numbers from 1 to 9 in sequence.

(2) Use the Raster calculator in ArcMap software to detect changes in land use types in different years. The algebraic expression is Con("landuse1.tif"!="landuse2.tif", 1,0) to generate a new raster dataset landuse_difference.tif, a pixel value of 1 represents an area where the land use type has changed, and a pixel value of 0 represents no change.

(3) Write a Python language program, according to the pixels with changes in land use, according to the spatial position index of the pixels, adopt the random traversal sampling method, set the corresponding traversal step size for different types of land, perform global search and sampling, and form training. Sample set D=[(x ₁ , y ₁ ),,...,(x _n , y _n )], where: x ₁ ...x _n represents the independent variable of n samples in the random forest model, y ₁ ...y _n represents the dependent variable for n samples.

In the above method, the random forest model can be iteratively trained by the following methods: write a Python language program, input the training sample D into the random forest model for model training, and set the maximum number of features as the square root of the number of potential driving factors N; The iterative growth of the number of models is used for model training, and the measurement of the model training effect adopts the average interval of model errors:

In the formula: MG _avg represents the average interval of all samples, n represents the number of samples, and mg( _xi , y _i ) represents the interval of a single sample. If mg(x _i , y _i ) is greater than zero, then the correct category occupies the most votes, and the final classification result is correct under voting; otherwise, the final classification result is wrong.

计算新的样本间隔MG_j，对每一个变量重复K次(本例中为50次)随机置换，以平均值作为变量重要性的最终结果：In the above method, the following methods can be used to calculate and sort the importance scores of the driving factors: write a Python language program, for each feature variable j in sample D, randomly replace the value in the variable to generate a new, destroyed training sample

Compute a new sample interval MG _j , repeating K (50 in this case) random permutations for each variable, taking the mean as the final result of variable importance:

where i _j represents the importance score of variable j, MG represents the average interval of the model before random permutation, K is the number of random permutations, and MG _{k, j} represents the average interval of the model after the k-th random permutation of variable j.

Using the feature factor importance score obtained in this step, the feature variables are sorted in descending order, that is, the importance ranking of the driving factors of land use change.

In the above method, the following methods can be used to screen the core driving force: using the principle of recursive feature elimination, sorting according to the importance of driving factors, starting from the most important driving factor, adding one factor at a time to form a new feature subset, inputting In the random forest model, use the cross-validation method to train the model and obtain the new model classification accuracy; repeat the above steps until all driving factors are included in the feature subset.

In the above method, the following methods can be used to determine the number of core driving factors: draw a curve of the variation of the classification accuracy of the model with the decrease in the number of characteristic variables, and find the corresponding point in the curve where the classification accuracy tends to converge, which is the number of core driving factors.

The above-mentioned land use change driving factor mining method based on random forest and crowd-source geographic information provided by the present invention is used to reveal the microscopic driving mechanism of urban evolution.

Applications:

This case takes the central area of Wuhan as the research area, which covers an area of 2724.228 square kilometers, accounting for 31.79% of the total area of Wuhan, and is the area with the highest degree of urbanization. Taking the driving force analysis of land use change in this region from 2015 to 2020 as an example, the present invention will be further described with reference to the accompanying drawings and the attached table.

The specific processing steps (Figure 1) are as follows:

Step 1: From the perspectives of social economy and natural ecological volume, construct a dataset of multiple potential driving factors affecting land use change based on crowd-sourced geographic information, and perform data spatial processing, including:

(1) The factors affecting urban land use change mainly include 20 factors in two categories: natural ecology and social economy. Among them, natural ecological factors include 3 terrain factors including elevation, slope and slope aspect, soil and water conservation function, soil organic matter and water system. There are 3 ecological factors; socioeconomic factors are derived from POI and other crowdsourced geographic information, including 14 types of population, economy, education, public services, and transportation (Table 1).

(2) Obtain raster (tif format) or vector data (shapefile format) representing all natural ecological and socioeconomic factors in the study area, and then import them into professional geographic information data processing and analysis software ArcMap10.2 in turn, and use the toolbox The Project function converts data from multiple sources into coordinate projections to keep the coordinate projections of the data consistent, as follows: WGS_1984_UTM_Zone_49N; then use the Clip function to cut all the data into a uniform shape with the boundary of the study area as the clipping range.

(3) Different spatial processing methods are adopted for different types of factors, including four methods of kernel density estimation, buffer creation, Euclidean distance calculation and partition statistics, respectively using Kernel Density, Multiple Ring Buffer in Arcmap10.2 software, Euclidean Distance and Zonal Statistic tools are implemented, and the spatial processing methods corresponding to the 20 factors are shown in Table 1. After the processing is completed, a spatial variable dataset with a resolution of 30m, a number of pixels of 3366990 (2090*1611) and a data format of tif type will be generated (Figure 2).

(4) In the first step, in order to eliminate the difference in the data dimension between different factors, use the Fuzzy tool in the Arcmap10.2 software to standardize the dispersion of the variables to achieve the normalization of the pixel values of the spatial variables, and the numerical values of the variables. The range maps to between 0 and 1.

The second step is to build a random forest classifier model with multiple potential driving factors as characteristic variables and the land use type of the land use thematic map as a predictor variable, and conduct model training, including:

(1) Import the 2015 and 2020 land use thematic maps in the central area of Wuhan into ArcMap software, the data space classification rate is 30m, and the land use types are divided into water, forest and grassland, cultivated land, unused land, residential land, industrial land, There are nine categories of commercial land, public management land and mixed land, and the category codes are numbers from 1 to 9.

(2) Use the Raster calculator in ArcMap software to detect the change of land use type from 2015 to 2020. The algebraic expression is Con("landuse2015.tif"!="landuse2020.tif", 1, 0) to generate a new grid In the grid dataset landuse_difference.tif, a pixel value of 1 represents a change in land use type from 2015 to 2020, and a pixel value of 0 represents no change. The pixels with the pixel value of 1 are selected, which is the area where the land use type has changed from 2015 to 2020, with a total of 823,636 pixels.

(3) For the 823,636 pixels with changes in land use, according to the spatial position index of the pixels, the random traversal sampling method is used to set the corresponding traversal step size for different types of land, and perform global search and sampling. The factors are characteristic variables, and the type of land used in 2020 is used as a predictor variable to form a training sample set D=[(x ₁ , y ₁ ), ,..., (x _n , y _n )], including 5717 water samples, woodland and There were 3283 grassland samples, 10583 cultivated land samples, 6509 unused land samples, 6852 residential land samples, 6791 industrial land samples, 4985 commercial land samples, 2416 public management land samples and 2570 mixed land samples.

The specific implementation code is as follows:

通过决策树个数的迭代增长进行模型训练，分析模型训练效果采用模型误差的平均间隔：(4) Input the training sample D into the random forest model for model training, and set the maximum number of features as the square root of the number of potential driving factors N. In this case, if 20 driving factors are used, the maximum number of features is

Model training is performed by iterative growth of the number of decision trees, and the average interval of model errors is used to analyze the model training effect:

In the formula: MG _avg represents the average interval of all samples, n represents the number of samples, and mg( _xi , y _i ) represents the interval of a single sample. If mg(x _i , y _i ) is greater than zero, then the correct category occupies the most votes, and the final classification result is correct under voting; otherwise, the final classification result is wrong. The results show that when the number of decision trees is 60, the average interval of the model tends to be stable (Figure 3). The specific implementation code is as follows:

计算新的样本间隔MG_j，考虑到随机置换的不稳定性，对每一个变量重复50次随机置换，以平均值作为变量重要性的最终结果：The third step is to evaluate the importance of variables. For each feature variable j in sample D, randomly replace the value in the variable to generate new and destroyed training samples.

Calculate the new sample interval MG _j , taking into account the instability of random permutations, repeat 50 random permutations for each variable, and take the mean as the final result of variable importance:

where i _j represents the importance score of variable j, MG represents the average interval of the model before random permutation, K is the number of random permutations, and MG _{k, j} represents the average interval of the model after the k-th random permutation of variable j.

Using the feature factor importance score obtained in this step, the feature variables are sorted in descending order, which is the driving force ranking of land use change (Figure 4). In this case, the importance of the 20 driving factors is ranked from high to low: population , K12 schools, industrial facilities, cultural and leisure venues, parks, green spaces, bus stations, colleges and universities, stadiums, hospitals, water systems, residential prices, subway stations, soil and water conservation, commercial facilities, main roads, elevation, soil organic matter, long-distance stations , slope and aspect. The specific implementation code is as follows:

The fourth step is to use the principle of recursive feature elimination to screen the core driving forces. According to the importance of the driving factors, start from the most important driving factor, and add one factor at a time to form a new feature subset, which is input to the random forest model. , use the cross-validation method to train the model and obtain the new model classification accuracy; repeat the above steps until all driving factors are included in the feature subset; draw the curve of the model classification accuracy as the number of feature variables decreases, and find that the classification accuracy in the curve tends to be The corresponding point of convergence is located at the 15th factor (Figure 5). After that, the classification accuracy of the model remains basically unchanged. The 15 core driving factors screened out are: population, K12 schools, industrial facilities, cultural and leisure venues, parks and green spaces , bus stations, colleges and universities, sports venues, hospitals, water systems, residential prices, subway stations, soil and water conservation, commercial facilities and main roads.

Table 1 Classification and spatialization of potential drivers of land use

Claims (10)

1. A land-use change driving factor mining method based on random forests and crowd-source geographic information, characterized in that: first, a multivariate potential driving factor dataset affecting land-use change is constructed by using crowd-source geographic data mainly based on POI points, Then, using multiple potential driving factors as feature variables and the land use type of land use thematic map as predictor variables, build a random forest classifier model, and conduct model training; then, use the trained model to conduct K random permutations of a single variable are used to calculate the importance score of the variable, and the importance of driving factors is ranked according to the score. Finally, the principle of recursive feature elimination is used to screen the core driving forces affecting land use change. 2. The land-use change driving factor mining method based on random forest and crowd-source geographic information according to claim 1, characterized in that, the following method is used to carry out data spatialization processing: according to the data type characteristics of different factors, using kernel density Various methods such as estimation, buffer creation, Euclidean distance calculation, zonal statistics, and slope and aspect calculation are used to spatially process the data to generate a spatial variable set with consistent resolution and continuous surface shape. 3. The land-use change driving factor mining method based on random forest and crowd-source geographic information according to claim 1, is characterized in that, adopt the following method to eliminate the difference of spatial variable set data dimension: utilize the Arcmap10.2 software in the The Fuzzy tool normalizes the dispersion of the variables to achieve the normalization of the pixel values of the spatial variables, and the numerical range of the variables is mapped between 0 and 1. 4. The land-use change driving factor mining method based on random forest and crowd-source geographic information according to claim 1, wherein the random forest classifier model is constructed by adopting the following method: Using the land use type of the thematic map as a predictor variable, the mapping relationship between the two is constructed. 5. The land-use change driving factor mining method based on random forest and crowd source geographic information according to claim 4, is characterized in that, constructing random forest classifier model adopts following method to carry out the collection of random forest model training sample: (1) Import the land use thematic maps of the study area in two different years into ArcMap software, the data space classification rate is 30m, and the land use types are divided into water, forest and grassland, cultivated land, unused land, residential land, industrial land, commercial land There are nine categories of land, public management land and mixed land, and the category codes are numbers from 1 to 9; (2) Use the Raster calculator in ArcMap software to detect changes in land use types in different years. The algebraic expression is Con("landuse1.tif"!="landuse2.tif", 1,0) to generate a new raster dataset landuse_difference.tif, a pixel value of 1 represents an area where the land use type has changed, and a pixel value of 0 represents no change; (3) For the pixels with changes in land use, according to the spatial position index of the pixels, the random traversal sampling method is used to set the corresponding traversal step size for different types of land, perform global search and sampling, and form a training sample set D = [ (x ₁ , y ₁ ), , ..., (x _n , y _n )] 6. The land-use change driving factor mining method based on random forest and crowd-source geographic information according to claim 4, characterized in that, constructing a random forest classifier model adopts the following method to perform random forest model training: The training sample D is input The random forest model is used for model training, and the maximum number of features is set as the square root of the number of potential driving factors N; model training is performed through the iterative growth of the number of decision trees, and the model training effect is measured by the average interval of model errors:

In the formula: MG _avg represents the average interval of all samples, n represents the number of samples, and mg( _xi , y _i ) represents the interval of a single sample. If mg(x _i , y _i ) is greater than zero, then the correct category occupies the most votes, and the final classification result is correct under voting; otherwise, the final classification result is wrong. 7. The land-use change driving factor mining method based on random forest and crowd-source geographic information according to claim 1, wherein the following method is used to calculate and sort the importance score of the driving factors: Feature variable j, randomly replace the values in this variable to generate new, destroyed training samples

Calculate a new sample interval MG _j , repeating 50 random permutations for each variable, taking the mean as the final result of variable importance:

In the formula: i _j represents the importance score of variable j, MG represents the average interval of the model before random permutation, K is the number of random permutations, MG _{k, j} represents the average interval of the model after the k-th random permutation of variable j; Using the feature factor importance score obtained in this step, the feature variables are sorted in descending order, that is, the importance ranking of the driving factors of land use change. 8. The land-use change driving factor mining method based on random forest and crowd-source geographic information according to claim 1, characterized in that, the following method is used to screen the core driving forces: sorting according to the importance of the driving factors, from the most important Starting from the driving factor of , one factor is added each time to form a new feature subset, which is input into the random forest model, and the model is trained using the cross-validation method to obtain the new model classification accuracy; repeat the above steps until all drivers are included in the feature subset. factor. 9. The land-use change driving factor mining method based on random forest and crowd-source geographic information according to claim 8, wherein the core driving force screening method adopts the following method to determine the number of core driving forces: The number of characteristic variables decreases and the curve changes. Find the corresponding point in the curve where the classification accuracy tends to converge, which is the number of core driving factors. 10. The land-use change driving factor mining method based on random forests and crowd-source geographic information according to any one of claims 1 to 9, characterized in that it is used to accurately assess and screen microscopic drivers that affect urban land use change factor.

Images