Forest fire risk prediction non-fire-point sample point selection method and system based on semi-supervised clustering membership

Abstract

The invention discloses a forest fire risk prediction non-fire-point sample point selection method and system based on a semi-supervised clustering membership degree. The method comprises the following steps: step 1, obtaining forest fire point data; step 2, fire point environment feature extraction; 3, candidate non-fire point sample points are generated, and environment features of the candidate non-fire point sample points are extracted; step 4, membership degree calculation based on a semi-supervised clustering idea; step 5, non-fire point sample point selection based on the membership degree; and step 6, using the non-fire point prediction sample for forest fire risk prediction model training. According to the forest fire risk prediction method, by means of the semi-supervised clustering thought, the membership degree of the environment feature of each position to the forest fire point environment feature is calculated, the representativeness of non-fire point samples is improved, the influence of other variables (such as meteorological changes and human activities) on forest fire risks is researched and predicted, the sample data quality of forest fire risk prediction is effectively improved, and the forest fire risk prediction efficiency is improved. And the model precision of forest fire risk prediction is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of data processing, and particularly relates to a forest fire risk prediction non-fire point sample point selection method and system based on semi-supervised clustering membership. BACKGROUND

[0002] Forest fire is one of the major natural disasters threatening the ecological system and people's life and property safety, and accurate prediction of forest fire risk is the premise of scientific disaster prevention and control. In recent years, data-driven risk prediction models have been widely used because they can learn disaster occurrence rules from data, and have become the mainstream technology for forest fire risk prediction. High-quality samples (fire point samples and non-fire point samples) are the basis for ensuring the accuracy of forest fire risk prediction. Since fire point samples can be obtained from satellite images and have high quality, the core of affecting the accuracy of forest fire risk prediction in practical application is the quality of non-fire point samples.

[0003] Therefore, how to design a scientific and reasonable non-fire point sample selection method to help the prediction model find the key factors affecting forest fires and conduct risk modeling and prediction is still a key problem that needs to be solved in the field of forest fire risk prediction research and application.

[0004] Currently, there are two methods for non-fire point sample selection. One is the spatial constraint + random sampling method. The other is the non-fire point sample selection based on geographical similarity and confidence. (1) Non-fire point sample selection method based on spatial constraint + random sampling

[0005] The non-fire point sample selection method based on spatial constraint + random sampling first limits the selection area of non-fire point samples through rule constraint, and then performs random sampling. This method reduces the probability of collecting invalid samples to a certain extent. However, since the random sampling method does not consider the influence of environmental characteristics such as terrain, weather, vegetation and human disturbance on forest fires when collecting non-fire point samples, it is easy to delimit the areas similar to the environmental characteristics of fire points as non-fire point sampling areas, which reduces the representativeness of non-fire point samples and the discrimination between fire point samples and non-fire point samples, and further cannot guarantee the representativeness of non-fire point samples. (2) Non-fire point sample selection method based on geographical similarity

[0006] The method for selecting non-fire point samples based on geographical similarity selects non-fire point samples based on factors such as meteorology, vegetation, terrain and human activities, calculates the comprehensive geographical similarity between a position and a forest fire position, and calculates the confidence of a position as a non-fire point sample by using the comprehensive geographical similarity. The higher the comprehensive geographical similarity is, the lower the confidence of the position as a non-fire point sample is. That is, when considering non-fire point samples, the area with high comprehensive geographical similarity with the forest fire position is excluded. However, in practical application, this method has obvious theoretical shortcomings and cannot select representative non-fire point samples. First, regardless of human ignition, forest fires are natural disasters controlled by natural factors. For a fixed position, the forest fire risk prediction focuses on the forest fire risk prediction under the change of natural factors, especially the change of meteorological conditions, and the forest and terrain features are the characteristics of the combustible itself and the site conditions and are relatively stable, so meteorological factors and human factors do not need to be introduced when measuring geographical similarity. Second, assuming that the vegetation and terrain features of two positions (one position has a forest fire and one position does not have a forest fire) have high similarity, if non-fire point samples are selected according to the principle of geographical similarity, the position without forest fire but with high similarity will be excluded from the non-fire point samples, at this time, the contrast data cannot be formed, so that the model cannot capture the key factors affecting the fire risk in this situation and make accurate prediction. In other words, when calculating the geographical similarity principle, the similarity of the relatively stable environmental characteristics needs to be considered, including vegetation characteristics and terrain characteristics. Therefore, when considering the candidate area of non-fire point samples, the position with high environmental feature similarity should be considered, rather than the position with low environmental feature similarity.

[0007] Therefore, it is necessary to improve the existing non-fire point sample selection method. SUMMARY

[0008] To solve the above problems, the present application discloses a forest fire risk prediction non-fire point sample point selection method and system based on semi-supervised clustering membership degree.

[0009] To achieve the above purpose, the technical scheme of the present application is as follows: A forest fire risk prediction non-fire point sample point selection method based on semi-supervised clustering membership degree, comprising the following steps: Step one, forest fire point data acquisition; extract historical forest fire points as fire point sample points, a total of M fire point sample points.

[0010] Step two, fire point environmental feature extraction: extract the environmental features of the fire point sample points, the environmental features including vegetation features and terrain features and excluding meteorological features; Step three, candidate non-fire point sample point generation and environmental feature extraction: The grid center without a fire point is taken as a candidate non-fire point sample point, and the corresponding environmental features are extracted; Step four, membership calculation based on semi-supervised clustering idea: The fire point sample points are taken as labeled samples, and the candidate non-fire point sample points are taken as unlabeled samples to construct a data set; the final fire point clustering center of the data set is calculated according to the environmental features, and the final membership of each labeled sample and unlabeled sample to the final fire point clustering center is calculated; Step five, non-fire point sample point selection based on membership: The minimum membership value of the final membership in the labeled sample is taken as a boundary threshold for non-fire point selection; the range from the boundary threshold to 1 is taken as a sampling interval, and M final selected non-fire point sample points are sampled from the candidate non-fire point sample points in the sampling interval; Step six, non-fire point prediction sample is used for forest fire risk prediction model training: the effective feature factors of the M fire point sample points and the effective feature factors of the M non-fire point sample points are taken as positive samples and negative samples respectively, and are input into the forest fire risk prediction model to obtain a trained forest fire risk prediction model; the effective feature factors include vegetation features, screened terrain features, screened meteorological features and human disturbance features.

[0011] Further improvement, the vegetation features include vegetation coverage and normalized difference vegetation index NDVI; the terrain features include elevation, slope and slope direction; the screened terrain features include slope and slope direction; the meteorological features include air temperature, relative humidity, consecutive rainless days, wind speed, soil moisture and potential evapotranspiration; the screened meteorological features include air temperature, relative humidity, consecutive rainless days, wind speed and soil moisture; the human disturbance features include road number, distance from road, distance from water system and distance from residential area.

[0012] Further improvement, the specific steps of step one are as follows: MODIS Collection 6 fire product data is downloaded, fire points with a confidence lower than 95% are removed according to the confidence information provided by the product, the reliability of the fire points is improved, and the fire points on non-forest land are removed by taking land use data as mask data, and the obtained historical forest fire points are taken as fire point sample points.

[0013] Further improvement, in step three, the forest region is divided into 1 km×1 km grids corresponding to the forest fire point data, and the center of the 1 km×1 km grid without a fire point is taken as a candidate non-fire point sample point.

[0014] Further improvement, in step four, the semi-supervised clustering idea is used to calculate the membership, and a target function for constructing an optimal clustering center is needed: (1) Where, is the total number of samples; is the number of cluster centers; For the Sample pair The membership degree of each cluster center; is a weighted index used to control the fuzziness of clustering, with a value of 2; For samples and cluster centers The Mahalanobis distance between is a balancing factor used to adjust the balance between unsupervised information and supervised information; is the membership of the labeled sample, indicating that the labeled sample Cluster centers The preset membership of is a Boolean binary vector. When the sample is an unlabeled sample , when the sample is a labeled sample, ; min means taking the minimum value; , represents the membership of each sample to each cluster center; represents the set of cluster centers, ; Objective function, used to measure the clustering effect under the current cluster center and membership distribution. The smaller the value, the better the clustering effect. In satisfying the constraints , that is, under the premise that the membership of each sample on each cluster center is 1, the objective function is iteratively updated to obtain the cluster center and the corresponding membership , the iterative update formula is as follows: (2) (3) In each iteration, firstly according to the cluster center of the previous round Calculate the membership of all samples according to formula (4) , and then update the cluster center according to formula (3) When satisfied or When , the iteration stops, and the is the final membership degree, The corresponding cluster center is the final fire point cluster center; Where: Represents a sample and cluster centers The Mahalanobis distance between Indicates the labeled samples; (4) wherein, is a covariance matrix of sample features, denotes matrix transposition.

[0015] Further improvement, the forest fire risk prediction model is a forest fire risk prediction model built by random forest, logistic regression, gradient boosting decision tree or support vector machine.

[0016] Further improvement, the number of positive samples and negative samples is the same.

[0017] Further improvement, in step six, the effectiveness of the trained forest fire risk prediction model is verified by accuracy, recall rate, precision, F1 score and area under the curve.

[0018] Further improvement, in step five, the sampling interval is divided into N subintervals in an equal division manner; M / N unlabeled sample points are randomly selected in each subinterval as non-fire point sample points, and finally M non-fire point sample points covering the membership sampling interval are formed, which are the finally selected M non-fire point sample points.

[0019] A forest fire risk prediction non-fire point sample point selection system based on semi-supervised clustering membership, the forest fire risk prediction non-fire point sample point selection based on semi-supervised clustering membership is used to run the above forest fire risk prediction non-fire point sample point selection method based on semi-supervised clustering membership.

[0020] Advantages of the present application: The present application calculates the membership of the environmental characteristics of each position to the environmental characteristics of the forest fire point by means of the semi-supervised clustering idea, the higher the membership, the more similar the environmental characteristics to the environmental characteristics of the forest fire point, and the greater the possibility of being a non-fire point sample point; in other words, when the vegetation and terrain characteristics of the fire point and the non-fire point sample are similar, the comparative data similar to the controlled experiment can be formed, that is, the representativeness of the non-fire point sample is improved, and then the influence of other variables (such as meteorological changes and human activities) on forest fire risk is studied and predicted. This method effectively improves the sample data quality of forest fire risk prediction, and further improves the model accuracy of forest fire risk prediction. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The technical flowchart of the present application.

[0022] Figure 2 The step schematic diagram of the semi-supervised clustering membership method of the present application.

[0023] Figure 3aAUC evaluation index result graph of each forest fire risk prediction model application of the non-fire point sample constructed by the method.

[0024] Figure 3b AUC evaluation index result graph of each forest fire risk prediction model application of the non-fire point sample constructed by the method. DETAILED DESCRIPTION

[0025] The application will be further described below in combination with the drawings and examples. Example 1

[0026] As shown in a forest fire risk prediction non-fire point sample point selection method based on semi-supervised clustering membership, comprising the following steps: Figure 1 Step 1: Obtain forest fire point data. First, download MODIS Collection6 fire product data of Hunan Province from 2004 to 2022, and remove fire points with confidence lower than 95% according to the confidence information provided by the product to improve the reliability of the fire points. Second, use the land use data released by Wuhan University as mask data to remove fire points in non-forest land, and finally obtain 7939 historical forest fire points, which are used as fire point sample points. Step 2: Extract fire point environmental feature. Extract the environmental feature parameters (labeled sample) of the corresponding forest fire points in step 1, including relatively stable vegetation parameters and terrain parameters. Vegetation parameters include vegetation coverage and NDVI, and terrain features include elevation, slope and aspect. It should be noted that the core of forest fire risk prediction is to study how meteorological conditions change affect forest fire risk under the condition of relatively stable forest vegetation and terrain features, so the forest fire point environmental features should exclude meteorological parameters, otherwise it is impossible to form effective comparison data.

[0027] Step 3: Generate candidate non-fire point sample points and extract their environmental features. Divide the forest area of Hunan Province into 1 km×1 km grids corresponding to the fire point data, and take the center of the 1 km×1 km grid without fire points as the candidate non-fire point sample point; extract the vegetation coverage, NDVI, elevation, slope and aspect of the sample point as the environmental features of the candidate non-fire point (unlabeled sample).

[0028] Step 4: Optimize the membership calculation model based on the semi-supervised clustering idea.

[0029] ​Theoretically, many locations may have similar vegetation and topographical characteristics, but significant differences in meteorological conditions can lead to different forest fire risk levels at these locations. To further analyze the impact of meteorological conditions on forest fire risk, this paper uses the concept of semi-supervised clustering to calculate membership. It identifies locations with environmental characteristics similar to forest fire sites but where no forest fire has occurred. These locations are treated as non-fire sites, thereby generating comparative and representative positive and negative sample data.

[0030] Membership calculation based on semi-supervised clustering uses labeled samples of known types as guiding information, imposing constraints or adjustments on the clustering process so that it gradually approaches the optimal fire cluster center during iterative optimization. The membership of each unlabeled sample point to the fire cluster center is calculated based on the optimal fire cluster center. The membership degree is used as a measure of the similarity between the candidate sample and the fire cluster center. The closer the membership degree is to 1, the more similar the environmental characteristics of the sample are to the fire sample, and the closer the membership degree is to 0, the greater the difference between the environmental characteristics of the sample and the fire sample. Based on the membership degree, locations with similar environmental characteristics to the fire sample are selected as non-fire sample points.

[0031] Using the semi-supervised clustering idea to calculate the membership degree requires constructing the objective function of the optimal cluster center and finding the optimal cluster center and the membership degree of each unlabeled sample point pair with the optimal cluster center through multiple iterations. The objective function of the optimal cluster center is shown in formula (1): (1) Where, is the total number of samples; is the number of cluster centers; For the Sample pair The membership degree of each cluster center; is a weighted index used to control the fuzziness of clustering, with a value of 2; For samples and cluster centers The Mahalanobis distance between is a balancing factor used to adjust the balance between unsupervised information and supervised information; is the membership of the labeled sample, indicating that the labeled sample Cluster centers The preset membership of is a Boolean binary vector. When the sample is an unlabeled sample , when the sample is a labeled sample, ; min means taking the minimum value; , represents the membership of each sample to each cluster center; represents the set of cluster centers, ; Objective function, for measuring the current clustering center and membership distribution under the clustering effect, the smaller the value indicates the better clustering effect.

[0032] In order to meet the constraint condition , that is, the membership of each sample on each cluster center is 1, the objective function is updated to obtain the cluster center and the corresponding membership , the iterative update formula is as follows: (2) (3) In the formula, , the Mahalanobis distance between the sample and the cluster center , the first Marked sample. In each iteration, first, according to the cluster center of the last round, the membership of all samples is calculated according to formula (4) , then the cluster center is updated according to formula (3) . When or is satisfied, the iteration is stopped, and the at this time is the final membership.

[0033] In the above objective function for seeking the optimal class center, the Euclidean distance is usually used to measure the closeness between the sample and the cluster center, but the Euclidean distance only calculates the geometric straight line distance in each feature dimension when measuring the difference between samples, and assumes that each feature is independent and has the same dimension. However, Mahalanobis distance can consider the overall distribution characteristics of data, and scale normalization and correlation correction are introduced to effectively eliminate the influence of dimension difference and correlation between features. Therefore, the objective function for the optimal class center is improved, that is, the Mahalanobis distance is used instead of the Euclidean distance, so as to select a more reasonable class center. The method of Mahalanobis distance is shown in formula (4): (4) In the formula, , the covariance matrix of the sample features, represents the Mahalanobis distance between the sample and the cluster center .

[0034] The specific steps of the semi-supervised clustering method are shown in Figure 2 ​As shown, first input the sample set composed of labeled samples and unlabeled samples, initialize the cluster centers; calculate the Mahalanobis distance between all samples in the sample set and the cluster centers according to the initial cluster centers, and update the membership matrix according to formula (4); calculate the objective function value according to formula (2), and judge whether the iteration stopping condition is met: when the change of the objective function value is less than a threshold , or the maximum displacement of the cluster center is lower than a threshold , the iteration stops and the final membership matrix and cluster center are output, otherwise, update the cluster center according to formula (3) and continue iteration until the stopping condition is met. Step 5: Selection of non-fire point sample points based on membership

[0035] In the application of non-fire point sample point selection of the present application, the sample set is composed of labeled samples (environmental characteristics of forest fire points) and unlabeled samples (environmental characteristics of candidate non-fire point sample points), denoted as , indicating the dimension of the environmental characteristics. The final membership of each sample belonging to the fire point cluster center is output, the smallest membership value in the labeled sample is taken as the boundary threshold for non-fire point selection, and 7939 non-fire point sample points are selected from the unlabeled samples with membership higher than the threshold (i.e. samples with similar environmental characteristics to fire points) in a ratio of 1:1 of positive and negative samples. Step 6: Application and evaluation of non-fire point sample point selection based on semi-supervised clustering membership

[0036] Step 6-1: Selection of forest fire driving factors. Four types of 15 forest fire driving factors that have a significant impact on the occurrence of forest fires are obtained, including elevation, slope and aspect in the terrain category; air temperature, relative humidity, consecutive rainless days, wind speed, potential evapotranspiration factor and soil moisture in the meteorological category; vegetation coverage and NDVI in the vegetation category; population, distance to road, distance to water system and distance to residential area in the human disturbance category; to reduce the redundant interference between variables, Pearson correlation analysis and variance inflation factor diagnosis method are used to select forest fire driving factors, and the slope factor and potential evapotranspiration factor are removed, and after removal, the factors are tested again, and there is no multicollinearity between the factors, and finally 13 effective characteristic factors are obtained: vegetation coverage, NDVI, elevation, aspect, air temperature, relative humidity, consecutive rainless days, wind speed, soil moisture, population, distance to road, distance to water system and distance to residential area.

[0037] Step 6-2: Forest fire risk prediction sample set construction. Using the 7939 forest fire points (positive samples) obtained in step 1 and the 7939 non-fire point sample points (negative samples) selected based on membership in step 5, the factor values of the corresponding 13 feature factors are extracted as the input data for the construction of the forest fire risk prediction model, and they are divided into forest fire risk prediction model training data and training effect test data in the ratio of 7:3.

[0038] Step 6-3: Forest fire risk prediction model construction. Four machine learning methods, including random forest (RF), logistic regression (LR), gradient boosting decision tree (GBDT), and support vector machine (SVM), are used to construct forest fire risk prediction models. The Optune hyperparameter optimization method is introduced to optimize the parameters of each model. The forest fire risk prediction sample set constructed in step 6-2 is input into the four machine learning methods for training and testing, and the probability values of each sample belonging to forest fire points are output, ranging from 0 to 1. The discrimination threshold is set to 0.5, and samples with probability values greater than 0.5 are identified as forest fire points, and samples with probability values less than 0.5 are identified as non-fire points.

[0039] Step 6-4: Forest fire risk prediction model accuracy evaluation. To verify the effectiveness of the non-fire point sample point selection method proposed in the present application in model construction, it is compared with the commonly used random sampling method for modeling. The prediction performance of different sample data on the test set of each model is quantitatively evaluated through five commonly used evaluation indexes: accuracy, recall rate, precision, F1 score, and area under the curve (AUC). The evaluation results are shown in Table 1, and the AUC result graph is shown in Figure 3a and Figure 3b The results show that among the four machine learning models, including RF, LR, GBDT, and SVM, the non-fire point sample point selection method based on semi-supervised clustering membership is superior to the random sampling method.

[0040] Table 1 Performance of forest fire risk prediction model under different non-fire point sample selection methods

[0041] Although the embodiments of the present application have been disclosed as above, they are not limited to the use listed in the specification and embodiments, and can be fully applied to various fields suitable for the present application. For those skilled in the art, other modifications can be easily realized, and therefore the present application is not limited to specific details and examples shown herein without departing from the general concept defined by the claims and their equivalent scope.

Claims

1. A method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership, characterized in that: The steps include: Step 1: Obtain forest fire point data: extract historical forest fire points as fire point sample points, with a total of M fire point sample points; Step 2: Extracting fire point environmental characteristics: extracting environmental characteristics of the fire point sample points, including vegetation characteristics and terrain characteristics, and excluding meteorological characteristics; Step 3: Generation of candidate non-fire point sample points and extraction of their environmental features: The grid center without fire points is used as a candidate non-fire point sample point, and the corresponding environmental features are extracted; Step 4: Membership calculation based on semi-supervised clustering: The fire point sample points are used as labeled samples, and the candidate non-fire point sample points are used as unlabeled samples to construct a data set. The final fire point cluster center of the data set and the final membership of each labeled sample and unlabeled sample to the final fire point cluster center are calculated based on the environmental characteristics. Step 5: Selection of non-fire point sample points based on membership degree: The minimum membership value of the final membership in the marked sample is used as the boundary threshold for non-fire point selection; the range from the boundary threshold to 1 is used as the sampling interval, and the M non-fire point sample points are finally selected by sampling from the candidate non-fire point sample points whose membership values ​​are within the sampling interval; Step 6. Non-fire point prediction samples are used for forest fire risk prediction model training: effective feature factors of M fire point sample points and effective feature factors of M non-fire point sample points are extracted as positive samples and negative samples respectively, and input into the forest fire risk prediction model to obtain a trained forest fire risk prediction model; the effective feature factors include vegetation characteristics, screened terrain characteristics, screened meteorological characteristics and human interference characteristics.

2. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: The vegetation characteristics include vegetation coverage and normalized difference vegetation index NDVI; Topographic characteristics include altitude, slope and aspect; the topographic characteristics after screening include slope and aspect; meteorological characteristics include temperature, relative humidity, consecutive rainless days, wind speed, soil moisture and potential evapotranspiration; the meteorological characteristics after screening include temperature, relative humidity, consecutive rainless days, wind speed and soil moisture; human interference characteristics include number of populations, distance to roads, distance to water systems and distance to residential areas.

3. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: The specific steps of step one are as follows: Download MODIS Collection 6 fire product data and remove fire points with a confidence level below 95% based on the built-in confidence information to improve the reliability of the fire points. Then, use land use data as mask data to remove fire points in non-forest land and use the historical forest fire points as fire point sample points.

4. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: In the step 3, the forest area is divided into 1 km×1 km grids corresponding to the forest fire point data, and the center of the 1 km×1 km grid without a fire point is used as a candidate non-fire point sample point.

5. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: In step 4, the objective function of constructing the optimal cluster center is required to calculate the membership degree using the semi-supervised clustering idea: (1) Where, is the total number of samples; is the number of cluster centers; For the Sample pair The membership degree of each cluster center; is a weighted index used to control the fuzziness of clustering, with a value of 2; For samples and cluster centers The Mahalanobis distance between is a balancing factor used to adjust the balance between unsupervised information and supervised information; is the membership of the labeled sample, indicating that the labeled sample Cluster centers The preset membership of is a Boolean binary vector. When the sample is an unlabeled sample , when the sample is a labeled sample, ; min means taking the minimum value; , represents the membership of each sample to each cluster center; represents the set of cluster centers, ; Objective function, used to measure the clustering effect under the current cluster center and membership distribution. The smaller the value, the better the clustering effect. In satisfying the constraints , that is, under the premise that the membership of each sample on each cluster center is 1, the objective function is iteratively updated to obtain the cluster center and the corresponding membership , the iterative update formula is as follows: (2) (3) In each iteration, firstly according to the cluster center of the previous round Calculate the membership of all samples according to formula (4) , and then update the cluster center according to formula (3) When satisfied or When , the iteration stops, and the is the final membership degree, The corresponding cluster center is the final fire point cluster center; Where: Represents a sample and cluster centers The Mahalanobis distance between Indicates the labeled samples; (4) Where, is the covariance matrix of sample features, Represents matrix transpose.

6. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: The forest fire risk prediction model is a forest fire risk prediction model constructed by random forest, logistic regression, gradient boosting decision tree or support vector machine.

7. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: The number of positive samples and negative samples is the same.

8. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: In step six, the effectiveness of the trained forest fire risk prediction model is verified by accuracy, recall, precision, F1 score and area under the curve.

9. The method for selecting non-fire point sample points for forest fire risk prediction based on semi-supervised clustering membership as claimed in claim 1, characterized in that: In step five, the sampling interval is divided into N sub-intervals in an equal manner; M / N unlabeled sample points are randomly selected in each sub-interval as non-fire point sample points, and finally M non-fire point sample points covering the membership sampling interval are formed, which are the M non-fire point sample points finally selected.

10. A forest fire risk prediction non-fire point sample point selection system based on semi-supervised clustering membership, characterized by: The forest fire risk prediction non-fire point sample point selection based on semi-supervised clustering membership is used to run the forest fire risk prediction non-fire point sample point selection method based on semi-supervised clustering membership described in any one of claims 1-9.

Images