A method and device for determining a type of forest disturbance, and an electronic device

Abstract

The application discloses a kind of method and device for determining forest disturbance type, electronic equipment, wherein the method comprises: obtaining forest disturbance data from land cover data set;Forest disturbance data is analyzed to feature, and the forest disturbance data under each classification is obtained;With the forest disturbance data under each classification as basic data, the way of mutual induction of random forest classifier and support vector machine classifier is used to carry out sample data volume expansion and classifier training multiple iterations, and the random forest classifier and support vector machine classifier satisfying preset condition are obtained;Based on the random forest classifier and support vector machine classifier satisfying preset condition, the forest disturbance data to be predicted is predicted, and the forest disturbance type is obtained.The scheme for determining forest disturbance type provided in the application can save the time and labor cost consumed by sample labeling, and can accurately identify the forest disturbance type.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of small sample neighborhood ecological environment remote sensing monitoring, and in particular relates to a method and device for determining forest disturbance type and electronic equipment. BACKGROUND

[0002] Disturbance is an important part of the dynamic change process of forest ecosystems, which is heterogeneous in space and time and is a relatively discrete event in time. It can change the composition and structure of forests, weaken the ecological service capacity, and reduce the availability of forest resources. Disturbance not only affects the evolution of forest ecosystems itself, but also has a strong impact on forest carbon storage and carbon sinks. In particular, subtropical evergreen forests are important global carbon sinks. However, due to frequent natural disasters (such as fires) and human activities (such as reclamation and deforestation), subtropical forests have become a research hotspot for forest disturbance. In order to better quantify forest loss and regional and global carbon dynamics, and maintain the sustainability of forest ecosystems, it is essential to understand the characteristics of different disturbance types in subtropical evergreen forests and their impacts.

[0003] Currently, the methods for identifying forest disturbance types can be mainly summarized into two categories, namely direct method and two-stage method. The direct method directly uses the spectral and temporal measurement indicators obtained from remote sensing time series data and prior knowledge of disturbance types, and simultaneously performs disturbance detection and attribution at the pixel scale with the aid of machine learning algorithms (such as random forest, decision tree, etc.). The two-stage method is divided into two steps. First, the disturbance event is determined at the pixel scale based on change detection algorithms. Then, after merging the disturbance pixels into disturbance patches, the spectral and spatial features of the disturbance events are extracted. Finally, machine learning algorithms (such as random forest, deep learning algorithms, etc.) are used for disturbance event attribution. The essence of the above two methods is to use machine learning for classification. However, this requires a large amount of sample data (prior knowledge) to train the model. Many studies obtain sample data through various historical records, high-resolution images, field surveys, and other means, and manually interpret a large number of sample data. In order to improve the generalization performance of the classification model, at least 200 labeled samples are required for each type of disturbance. The labeling cost is very high, and a lot of time is needed. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a method and device for determining forest disturbance type and electronic equipment, which can solve the above-mentioned problems existing in the prior art.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] In one aspect of the present application, a method for determining forest disturbance type is provided, comprising:

[0007] obtaining forest disturbance data from the land cover dataset; wherein the disturbance data comprises: disturbance time and disturbance location;

[0008] performing feature analysis on the forest disturbance data to obtain forest disturbance data under each category;

[0009] based on the forest disturbance data under each category as the basic data, using the mutual induction mode of random forest classifier and support vector machine classifier to perform sample data volume expansion and classifier training multiple times to obtain random forest classifier and support vector machine classifier satisfying the preset condition;

[0010] based on the random forest classifier and support vector machine classifier satisfying the preset condition, predicting the forest disturbance data to be predicted to obtain the forest disturbance type.

[0011] Optionally, the step of obtaining forest disturbance data from the land cover dataset comprises:

[0012] performing forest mask on the initial land cover dataset to obtain forest related data;

[0013] using continuous change detection and classification algorithm (CCDC) to detect the forest related data to obtain a detection result, wherein the detection result includes: forest change to other land disturbance and other land change to forest disturbance;

[0014] based on the land cover dataset, performing disturbance pixel filtering on the forest change to other land disturbance to obtain forest disturbance data.

[0015] Optionally, the step of performing feature analysis on the forest disturbance data to obtain forest disturbance data under each category comprises:

[0016] based on the time feature, spatial feature, spectral feature and topographic feature of subtropical evergreen forest disturbance classification, analyzing the forest disturbance data to obtain the classification corresponding to each forest disturbance;

[0017] forest disturbance data of the same category is divided into the same category to obtain forest disturbance data under each category.

[0018] Optionally, the step of using the forest disturbance data under each category as the basic data, using the mutual induction mode of random forest classifier and support vector machine classifier to perform sample data volume expansion and classifier training multiple times to obtain random forest classifier and support vector machine classifier satisfying the preset condition comprises:

[0019] for the forest disturbance data of each category, a preset number of forest disturbance data is taken as sample data;

[0020] Divide the sample data of each category into training samples and test samples according to a preset ratio;

[0021] Train the random forest classifier and the support vector machine classifier respectively by using the training samples;

[0022] According to the trained random forest classifier and the support vector machine classifier, the test samples corresponding to each category are predicted respectively to determine the test accuracy of the random forest classifier and the support vector machine classifier;

[0023] If the test accuracy does not meet a preset condition, the un-labeled forest disturbance data are classified according to the trained random forest classifier and the support vector machine classifier respectively to obtain a first classification result and a second classification result;

[0024] According to the first classification result and the second classification result, the training sample increment samples of the random forest classifier and the support vector machine classifier are determined;

[0025] The random forest classifier is trained by using the increment training samples corresponding to the random forest classifier, and the support vector machine classifier is trained by using the increment training samples corresponding to the support vector machine classifier, so as to complete one mutual training of the random forest classifier and the support vector machine classifier.

[0026] Optionally, the step of determining the training sample increment samples of the random forest classifier and the support vector machine classifier according to the first classification result and the second classification result comprises:

[0027] By comparing the first classification result and the second classification result, the classification result with differences is determined as pseudo-label data;

[0028] First high-confidence pseudo-label data are screened out from the pseudo-label data contained in the first classification result, and the first high-confidence pseudo-label data are used as the training sample increment samples of the support vector machine classifier;

[0029] Second high-confidence pseudo-label data are screened out from the pseudo-label data contained in the second classification result, and the second high-confidence pseudo-label data are used as the training sample increment samples of the random forest classifier.

[0030] Optionally, the step of screening first high-confidence pseudo-label data from the pseudo-label data contained in the first classification result comprises:

[0031] From the pseudo-label data contained in the first classification result, first pseudo-label data with a classification probability value greater than a first preset probability threshold value are screened out;

[0032] Filtering out false pseudo-label data in the first pseudo-label data by using a confidence learning algorithm to obtain first high-confidence pseudo-label data.

[0033] Optionally, the step of taking the first high-confidence pseudo-label data as the training sample incremental sample of the support vector machine classifier comprises:

[0034] For each classification contained in the first classification result, determining first high-confidence pseudo-label data corresponding to the classification;

[0035] Performing inter-class data balancing processing on the first high-confidence pseudo-label data corresponding to each classification;

[0036] Taking the first high-confidence pseudo-label data after the inter-class balancing processing as the training sample incremental sample of the support vector machine classifier.

[0037] In another aspect of the embodiment of the present application, a device for determining the type of forest disturbance is provided, wherein,

[0038] An acquisition module is configured to acquire forest disturbance data from a land cover data set; wherein the disturbance data comprises: disturbance time and disturbance location;

[0039] A classification module is configured to perform feature analysis on the forest disturbance data to obtain forest disturbance data under each classification;

[0040] A mutual guidance training module is configured to use the forest disturbance data under each classification as basic data, and use a random forest classifier and a support vector machine classifier in a mutual guidance manner to perform sample data volume expansion and classifier training multiple times to obtain a random forest classifier and a support vector machine classifier that meet a preset condition;

[0041] A prediction module is configured to predict forest disturbance data to be predicted based on the random forest classifier and the support vector machine classifier that meet the preset condition to obtain the type of forest disturbance.

[0042] Optionally, the acquisition module comprises:

[0043] A first sub-module is configured to perform forest mask on the initial land cover data set to obtain forest-related data;

[0044] A second sub-module is configured to use a continuous change detection and classification algorithm (CCDC) to detect the forest-related data to obtain a detection result, wherein the detection result comprises: forest change to other land disturbance and other land change to forest disturbance;

[0045] A third sub-module is configured to perform disturbance pixel filtering on the forest change to other land disturbance based on the land cover data set to obtain forest disturbance data.

[0046] In still another aspect of the embodiments of the present application, an electronic device is also provided, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is configured to store a computer program; and the processor is configured to execute the program stored in the memory to implement the method for determining the forest disturbance type.

[0047] The embodiments of the present application provide a scheme for determining a forest disturbance type, obtaining forest disturbance data from a land cover data set; performing feature analysis on the forest disturbance data to obtain forest disturbance data under each category; taking the forest disturbance data under each category as basic data, using a mutual guidance method of a random forest classifier and a support vector machine classifier to perform sample data volume expansion and classifier training multiple times, to obtain a random forest classifier and a support vector machine classifier satisfying a preset condition; and predicting the forest disturbance data to be predicted based on the random forest classifier and the support vector machine classifier satisfying the preset condition, to obtain the forest disturbance type. Through the scheme for determining the forest disturbance type provided by the embodiments of the present application, the model is trained by combining small sample learning and the mutual guidance learning method of the two base classifiers of the random forest and the support vector machine, and the user only needs to manually mark a small amount of samples, and the subsequent iteration of the mutual guidance to expand the sample data volume can save the time and labor cost consumed by sample marking. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is a step flow chart of a method for determining a forest disturbance type according to an embodiment of the present application;

[0049] Figure 2 is a flow chart of another method for determining a forest disturbance type according to an embodiment of the present application;

[0050] Figure 3 is a classifier mutual guidance framework based on confidence learning according to an embodiment of the present application;

[0051] Figure 4 is a structural block diagram of a device for determining a forest disturbance type according to an embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to make the technical problems, technical solutions and advantages of the present application clearer, specific embodiments will be described in detail below with reference to the accompanying drawings.

[0053] Therefore, to reduce the labeling cost of sample data and address the issue of a small proportion of a particular type of change region compared to other types, leading to a small sample size for that type of change, researching forest disturbance type identification methods under small sample conditions is of great significance. Classification under small sample conditions, also known as small-sample learning, is currently applied in the hyperspectral field and belongs to semi-supervised learning. Semi-supervised learning uses a small amount of sample data to train a model, and then uses the trained model to predict a large amount of unlabeled data. The prediction results are called pseudo-labels. High-quality pseudo-labels are selected from the prediction results and added to a small amount of labeled data to expand the sample data, thereby reducing human intervention from a mechanistic perspective. Thus, how to select high-quality pseudo-labels to add to the training set is a key point of semi-supervised learning. Existing research mainly expands the sample data in two ways. One is to select high-quality pseudo-labels based on similarity measures such as cosine distance within a certain neighborhood. This method has the problem of a limited number of selected pseudo-labels and the inability to effectively handle noise. The second is selection based on probability uncertainty and confidence. This method can integrate multiple classification models to expand the training set, solving the problem of the limited number of pseudo-labels. However, it relies on its own model for iterative prediction, and the presence of noisy pseudo-labels can easily lead to error accumulation. To address this, some scholars have proposed a mutual guidance framework. This framework adds the high-confidence pseudo-label data obtained by each classifier to the other's small sample set, thereby eliminating the error accumulation problem that may be caused by noisy iterative training of the same classification model. However, noise issues still exist in the mutual guidance framework, and the selection of a fixed number of high-confidence pseudo-labels in each iteration does not consider the balance of different types of pseudo-labels, leading to some data waste. Based on this, this invention designs a scheme for determining forest disturbance types, which is a forest disturbance attribution method based on a confidence-learning mutual guidance framework under small sample conditions.

[0054] The following description, in conjunction with the accompanying drawings, details the scheme for determining forest disturbance types provided in this application through specific embodiments and application scenarios.

[0055] As attached Figure 1 As shown, the method for determining the type of forest disturbance in this application includes the following steps:

[0056] Step 101: Obtain forest disturbance data from the land cover dataset.

[0057] The disturbance data includes: the disturbance time and the location where the disturbance occurred.

[0058] Land cover dataset: Forest mask is extracted from the existing land cover classification data. The data is derived from Wuhan University, which uses all available Landsat data on GEE to construct spatiotemporal features, combine random forest classifier classification, and propose a post-processing method containing spatiotemporal filtering and logical reasoning to further improve the spatiotemporal consistency of CLCD. An annual China Land Cover Dataset (CLCD) is produced, including China's annual land cover information from 1985, 1990-2020. Finally, based on 5463 visual interpretation samples, the classification accuracy is verified, and the overall accuracy of the CLCD can reach 80%.

[0059] Continuous Change Detection and Classification (CCDC): It is a disturbance detection algorithm for all land cover types based on time series data proposed by Zhu and Woodcock, which can identify all three types of changes: seasonal change, gradual change and sudden change, mainly for identifying sudden change. The change detection principle is: use ordinary least squares (OLS) to model the seasonal change and gradual change of the pixel value of each band in the image over time. Then calculate the difference between the predicted modeled pixel value and the real pixel value. When the difference between the values is greater than three times the root mean square error (RMSE), the pixel will be marked as a possible land cover change. Then according to the number of continuous observations, evaluate whether the potential land cover change is a real change.

[0060] In the embodiments of the present application, the forest disturbance data is obtained based on the land cover dataset and the CCDC. In an optional embodiment, the way to obtain the forest disturbance data from the land cover dataset can include the following sub-steps:

[0061] Sub-step 1: Perform forest mask on the initial land cover dataset to obtain forest-related data;

[0062] Sub-step 2: Use the continuous change detection and classification algorithm CCDC to detect the forest-related data to obtain a detection result;

[0063] Among them, the detection result includes: forest change to other land class disturbance and other land class change to forest disturbance.

[0064] Sub-step 3: Based on the land cover dataset, filter the disturbance pixels of the forest change to other land class disturbance to obtain forest disturbance data.

[0065] Step 102: Perform feature analysis on the forest disturbance data to obtain forest disturbance data under each classification.

[0066] In this step, the time-space spectrum and terrain features suitable for the classification of subtropical evergreen forest disturbance are found out to distinguish different types of forest disturbance.

[0067] An optional way of feature analysis of forest disturbance data to obtain forest disturbance data under each classification is as follows:

[0068] Based on the time feature, spatial feature, spectral feature and terrain feature of the subtropical evergreen forest disturbance classification, the forest disturbance data is analyzed to obtain the classification corresponding to each forest disturbance; the forest disturbance data of the same classification is divided into the same classification to obtain the forest disturbance data under each classification.

[0069] Time feature: the time feature is the disturbance amplitude, disturbance duration, disturbance rate, etc. of different forest disturbance types;

[0070] Spatial feature: the spatial feature represents the disturbance area and texture difference of different forest disturbance types;

[0071] Spectral feature: the spectral feature represents the difference in spectral value or vegetation index value before and after the disturbance of different forest disturbance types;

[0072] Terrain feature: the terrain feature represents the difference in spatial position of different forest disturbance types.

[0073] Step 103: Based on the forest disturbance data under each classification, the random forest classifier and the support vector machine classifier are iterated multiple times to expand the sample data and train the classifier, and the random forest classifier and the support vector machine classifier satisfying the preset condition are obtained.

[0074] Random forest classifier: the random forest algorithm (Random Forest, RF) uses a random way to establish a forest, and the forest is composed of a plurality of unassociated decision trees. After obtaining the forest, when a new input sample enters, each decision tree in the forest respectively judges the class of the sample (for classification algorithm), and finally votes the final class of the sample.

[0075] Support vector machine classifier: support vector machine (Support Vector Machine, SVM) is a kind of efficient supervised learning algorithm. Its goal is to maximize the class edge by finding the optimal classification interface equation, so as to realize accurate classification.

[0076] The mutual guidance of the RF and the SVM is used to provide pseudo labels for each other, guide each other to learn, and avoid accumulation of errors. The RF and the SVM are used to preliminarily classify a large amount of unlabeled data by using a small amount of sample data for the first time, and two classification results with differences, that is, pseudo labels, are obtained. The pseudo labels obtained from the RF are selected, and high-confidence pseudo labels are added to the small amount of sample data of the SVM, so as to expand the sample set of the SVM. Meanwhile, the pseudo labels obtained from the SVM are selected, and high-confidence pseudo labels are added to the small amount of sample data of the RF, so as to expand the sample set of the RF, and the mutual guidance of the RF and the SVM is implemented.

[0077] In the embodiments of the present application, the mutual guidance of the random forest and the support vector machine classifier is used to reduce the cost of manually labeling samples, realize mutual teaching of each other in a small sample condition, and provide optional high-confidence pseudo sample data. The high-confidence pseudo labels are selected based on classification probability, confidence learning is introduced to remove false pseudo labels in the high-confidence pseudo label data, the number of pseudo labels of each category is balanced, the sample data amount is expanded, and high-precision forest disturbance attribution can be realized in a small sample condition.

[0078] In step 104, the forest disturbance data to be predicted is predicted based on the random forest classifier and the support vector machine classifier that meet the preset condition, and the forest disturbance type is obtained.

[0079] The preset condition can be flexibly set by a person skilled in the art, for example, the prediction accuracy is less than a preset threshold after two iterations of training; for example, the prediction accuracy of the classifier reaches a preset accuracy after one iteration of training; for example, the number of iterations reaches a preset number of iterations, and the like. The preset condition is not limited in the embodiments of the present application.

[0080] The method for determining the forest disturbance type provided in the embodiments of the present application obtains forest disturbance data from a land cover data set, performs feature analysis on the forest disturbance data, obtains forest disturbance data under each category, uses the forest disturbance data under each category as basic data, and uses the mutual guidance of the random forest classifier and the support vector machine classifier to iteratively expand the sample data amount and train the classifier multiple times, so as to obtain the random forest classifier and the support vector machine classifier that meet the preset condition. The forest disturbance data to be predicted is predicted based on the random forest classifier and the support vector machine classifier that meet the preset condition, and the forest disturbance type is obtained. Through the scheme for determining the forest disturbance type provided in the embodiments of the present application, the model is trained by combining small sample learning and the mutual guidance learning method of the two base classifiers of the random forest and the support vector machine. The user only needs to manually mark a small amount of samples, and the sample data amount is expanded through subsequent iteration and mutual guidance, so that the time and labor cost of sample labeling can be saved.

[0081] In an alternative embodiment, based on the forest disturbance data under each category, the random forest classifier and the support vector machine classifier are iterated multiple times for sample data augmentation and classifier training in a mutual induction manner, and the manner of obtaining the random forest classifier and the support vector machine classifier satisfying the preset condition can include the following sub-steps:

[0082] Sub-step 1: For the forest disturbance data of each category, a preset number of forest disturbance data is taken as sample data.

[0083] Sub-step 2: The sample data of each category is divided into training samples and test samples according to a preset ratio.

[0084] An exemplary sample data acquisition and division manner is that the sample data of each category is one hundred, which is randomly divided into training samples and test samples according to a ratio of 7:3. It should be noted that the preset number and the preset ratio can be flexibly set by those skilled in the art, and the present embodiment does not make specific limitations thereon. For example, the preset ratio can be set to 8:2, 9:1, etc.

[0085] Sub-step 3: The training samples are used to train the random forest classifier and the support vector machine classifier respectively.

[0086] Sub-step 4: The test samples corresponding to each category are predicted according to the trained random forest classifier and support vector machine classifier respectively, and the test accuracy of the random forest classifier and the support vector machine classifier is determined.

[0087] If the test accuracy meets the preset condition, the iteration training is stopped, and it is determined that the random forest classifier and the support vector machine classifier have been trained.

[0088] Sub-step 5: If the test accuracy does not meet the preset condition, the unmarked forest disturbance data is classified according to the trained random forest classifier and support vector machine classifier respectively, and the first classification result and the second classification result are obtained.

[0089] Sub-step 6: The training sample increment sample of the random forest classifier and the support vector machine classifier is determined according to the first classification result and the second classification result.

[0090] More specifically, the manner of determining the training sample increment samples of the random forest classifier and the support vector machine classifier according to the first classification result and the second classification result can be: comparing the first classification result and the second classification result, determining the classification result with differences as pseudo-label data (i.e. pseudo-labels); screening first high-confidence pseudo-label data from the pseudo-label data contained in the first classification result, and taking the first high-confidence pseudo-label data as the training sample increment samples of the support vector machine classifier; screening second high-confidence pseudo-label data from the pseudo-label data contained in the second classification result, and taking the second high-confidence pseudo-label data as the training sample increment samples of the random forest classifier.

[0091] An exemplary manner of screening the first high-confidence pseudo-label data from the pseudo-label data contained in the first classification result can be: screening first pseudo-label data with a classification probability value greater than a first preset probability threshold from the pseudo-label data contained in the first classification result; filtering out error pseudo-label data in the first pseudo-label data by using a confidence learning algorithm to obtain the first high-confidence pseudo-label data.

[0092] The manner of screening the second high-confidence pseudo-label data from the pseudo-label data contained in the second classification result is similar to the above manner, and will not be described here.

[0093] In actual implementation process, the manner of adding high-confidence pseudo-labels can include the following steps:

[0094] First, pseudo-label preliminary screening is used to filter pseudo-labels with low probability to obtain high-confidence pseudo-label data; second, error pseudo-label filtering is used to filter error pseudo-labels in the high-confidence pseudo-label data

[0095] Confidence learning is a new and principled framework to identify label errors, characterize label noise and apply to noisy learning. Confidence learning can directly estimate the joint distribution of noisy labels and true labels, without making the assumption of random and uniform label noise, and without hyperparameters, but only needs to use cross-validation to obtain out-of-sample prediction probabilities. Applying confidence learning to the multi-classifier mutual guidance framework can automatically find out the samples with error labels annotated by base classifiers and filter them out.

[0096] Confidence learning filters out error samples in two steps. The first step is to estimate the joint distribution (count) of noisy labels and true labels. The second step is to find and filter out error samples (clean).

[0097] In the first step, the noisy label is defined as i.e. samples annotated by base classifiers but possibly with errors; and the true label is defined as y *, the true label is estimated by cross-validation. In addition, define the total number of samples as n and the total number of classes as m.

[0098] To estimate the joint distribution, a total of 4 steps are needed:

[0099] a) Cross-validation

[0100] First, the probability P[i][j] of the ith sample in the jth class needs to be calculated by cross-validation on the dataset, then the average probability t[j] of each base classifier in the jth class is calculated as the confidence threshold; Finally, for sample i, the true label y * is calculated as the maximum probability argmaxP[i][j] in the jth class, and P[i][j]>t[j].

[0101]

[0102] where θ is the model.

[0103] b) Calculate the count matrix (similar to the confusion matrix)

[0104]

[0105] where,

[0106]

[0107] where l∈[m];

[0108] c) Calibrate the count matrix

[0109] The purpose is to make the count sum equal to the total number of samples labeled by the base classifier. The formula is as follows:

[0110]

[0111] where, is the pseudo-label The total number of samples.

[0112] d) Estimate the noise label and the joint distribution of the true label y *

[0113]

[0114] The joint distribution can fully reflect the distribution of noise (error) labels and true labels in the real world, and as the data size expands, this estimation method is closer to the true distribution.

[0115] ​After obtaining the joint distribution of noise labels and true labels, erroneous samples can be filtered out.

[0116] An exemplary method for using the first high-confidence pseudo-label data as incremental training samples for a support vector machine classifier is as follows: For each category included in the first classification result, determine the first high-confidence pseudo-label data corresponding to each category; perform inter-class data balancing processing on the first high-confidence pseudo-label data corresponding to each category; and use the first high-confidence pseudo-label data after inter-class balancing processing as incremental training samples for the support vector machine classifier.

[0117] Similarly, the pseudo-label data with the second highest confidence can be used as incremental training samples for the random forest classifier. Class imbalance can eliminate the problem of low generalization performance in classification models caused by class imbalance.

[0118] Sub-step 7: Train the random forest classifier using the incremental training samples corresponding to the random forest classifier, and train the support vector machine classifier using the incremental training samples corresponding to the support vector machine classifier, thereby completing one cross-training of the random forest classifier and the support vector machine classifier.

[0119] In actual implementation, sub-steps 4 to 7 can be repeated to perform multiple iterative cross-training of the random forest classifier and the support vector machine classifier until the two classifiers meet the preset conditions and training stops.

[0120] The following is combined Figure 2 , Figure 3 The method for determining the type of forest disturbance according to an embodiment of this application will be described using a specific example.

[0121] like Figure 2 As shown, the method for determining the type of forest disturbance in this application mainly includes the following parts: data preprocessing, labeling of a small number of samples, construction and selection of attribution features, iterative cross-training of random forest classifier and support vector machine classifier based on a small number of samples, and determination of forest disturbance attribution.

[0122] The acquisition and preprocessing of disturbance data are as follows:

[0123] In this example, CCDC is used to obtain forest disturbance pixels for subsequent disturbance attribution. Forest disturbance events are detected in GEE (Google Earth Engine) using CCDC. First, all available Landsat data with cloud cover less than 80% is used, with a time span of 1986-2022. It is idealized that the earliest forest is the most original forest range, so the forest mask is extracted using the CLCD data of 1985, and applied to all Landsat data. Second, in order to detect as much forest disturbance as possible, the input parameters of the algorithm, minObservations and chiSquareProbability, are set to 4 and 0.9 respectively, and multiple disturbance data is detected. Finally, 250 change samples and 250 unchanged samples (sample points conform to normal distribution) are made through visual interpretation for accuracy verification, and the overall accuracy is 90.4%.

[0124] Disturbance attribution is only for forest to other land cover types. Although forest mask is used to ensure forest disturbance, CCDC can detect disturbances of all land cover types, so after the second forest disturbance, there are still other land cover types changing to forest. In order to filter out this type, the time of each disturbance is checked, and it is found that there is almost only one disturbance per year, so annual CLCD data is used to filter the pixels that change from other land cover to forest, and the disturbance pixels of forest to other land cover are obtained.

[0125] Feature extraction, i.e. attribution feature construction and selection:

[0126] The forest disturbance types identified in this study include fire, deforestation, road construction, and reclamation. The features used in existing literature are summarized, and the differences in time, space, spectrum, and terrain of different disturbance types are analyzed to obtain the following table of features.

[0127] Table 1: Features for forest disturbance attribution

[0128]

[0129]

[0130] Based on a small number of samples, the random forest classifier and the support vector machine classifier are iteratively trained:

[0131] The classifier mutual guidance framework based on confident learning is shown in Figure 3 . First, the sample data of each class, i.e. the number of categories is about one hundred, is randomly divided into training samples and test samples according to 7:3.

[0132] Secondly, a small amount of sample data is used to train the RF and SVM models, and then the trained models are used to preliminarily classify and predict a large amount of unlabeled data, respectively, to obtain two classification results with differences, i.e. pseudo labels; meanwhile, the test samples are used to verify the accuracy of the classification results. Next, the high-confidence pseudo labels obtained from the RF are selected and added to the small amount of sample data of the SVM to expand the sample set of the SVM, and the high-confidence pseudo labels obtained from the SVM are selected and added to the small amount of sample data of the RF to expand the sample set of the RF. Then, the expanded sample sets are used to train the RF and SVM, and the trained models are used to classify and predict the updated unlabeled data, to realize mutual guidance of the RF and SVM models; meanwhile, the test samples are used to verify the accuracy of the classification results. In this way, iteration is performed, and the generalization performance and stability of the model are judged by the size of the verification accuracy and the difference between the verification accuracies before and after.

[0133] The selection of high-confidence pseudo labels includes:

[0134] Firstly, preliminary screening of pseudo labels is performed to obtain the probabilities Prob_SVM and Prob_RF in the classification of the SVM and the RF, and an empirical threshold is set to filter the pseudo labels with low probability values. The probabilities are respectively:

[0135] Pseudo_label SVM =Prob_SVM>=0.9;

[0136] Pseudo_label RF =Prob_RF>=0.7;

[0137] Filtering of error pseudo labels:

[0138] Secondly, the confidence learning is used to filter out the error pseudo labels, and the filtering rule is as follows:

[0139] The samples with are filtered, i.e. the samples with the index j corresponding to the maximum probability of P[i][j] and the pseudo labels are inconsistent.

[0140] Balance of the number of classes:

[0141] The number of pseudo labels of different perturbation types in the pseudo label results obtained after filtering of the error pseudo labels is unbalanced, which will make the classifier learn towards the class with more quantity when input into the classifier, resulting in low classification accuracy. Therefore, the number of classes needs to be balanced, and the balance rule of the number of classes is set as:

[0142] Label_numMin=min(label_numi)

[0143] wherein i∈[1, 2, 3, 4] represents different disturbance types;

[0144] If Label_numMin is less than 50, then

[0145] labeli_num = 50;

[0146] If Label_numMin is greater than or equal to 50 and less than 100, then

[0147] labeli_num = Label_numMin / / 2;

[0148] If Label_numMin is greater than or equal to 100 and less than 150, then

[0149] labeli_num = Label_numMin / / 3;

[0150] If Label_numMin is greater than or equal to 150 and less than 200, then

[0151] labeli_num = Label_numMin / / 4;

[0152] If Label_numMin is greater than or equal to 200 and less than 250, then

[0153] labeli_num = Label_numMin / / 5;

[0154] If Label_numMin is greater than or equal to 250 and less than 300, then

[0155] labeli_num = Label_numMin / / 6;

[0156] If Label_numMin is greater than or equal to 300, then

[0157] labeli_num = 50;

[0158] The method for determining the forest disturbance type provided in the specific example explores the application of small sample in the attribution of subtropical evergreen forest disturbance. Small sample learning is often used in classification of high-resolution hyperspectral data, and has not been applied to the attribution of evergreen forest disturbance based on medium-resolution and long-time series. Combined with confidence learning, it solves the noise problem and sample quantity balance problem in small sample learning. In addition, the mutual guidance of the two base classifiers, random forest and support vector machine, in the example can improve the classification accuracy of the classifier.

[0159] Figure 4A structural block diagram of a device for determining a forest disturbance type according to an embodiment of the present application.

[0160] The device for determining a forest disturbance type provided by the embodiments of the present application comprises the following functional modules:

[0161] The acquisition module 401 is configured to acquire forest disturbance data from a land cover data set; wherein the disturbance data comprises a disturbance time and a disturbance occurrence position;

[0162] The classification module 402 is configured to perform feature analysis on the forest disturbance data to obtain forest disturbance data under each classification;

[0163] The mutual guidance training module 403 is configured to take the forest disturbance data under each classification as basic data, and perform sample data volume expansion and classifier training through multiple iterations in a mutual guidance manner of a random forest classifier and a support vector machine classifier to obtain a random forest classifier and a support vector machine classifier that meet preset conditions;

[0164] The prediction module 404 is configured to predict forest disturbance data to be predicted based on the random forest classifier and the support vector machine classifier that meet the preset conditions to obtain a forest disturbance type.

[0165] Optionally, the acquisition module comprises:

[0166] The first sub-module is configured to perform forest mask on an initial land cover data set to obtain forest-related data;

[0167] The second sub-module is configured to detect the forest-related data by using a continuous change detection and classification algorithm (CCDC) to obtain a detection result, wherein the detection result comprises disturbances in which forest changes into other land types and disturbances in which other land types change into forest;

[0168] The third sub-module is configured to perform disturbance pixel filtering on the disturbances in which forest changes into other land types based on the land cover data set to obtain forest disturbance data.

[0169] Optionally, the classification module comprises:

[0170] The fourth sub-module is configured to analyze the forest disturbance data based on time features, spatial features, spectral features and terrain features of subtropical evergreen forest disturbance classification to obtain a classification corresponding to each forest disturbance;

[0171] The fifth sub-module is configured to divide forest disturbance data of the same classification into the same classification to obtain forest disturbance data under each classification.

[0172] Optionally, the mutual guidance training module comprises:

[0173] The sixth sub-module is configured to take a preset number of forest disturbance data as sample data for each classification of forest disturbance data;

[0174] The seventh sub-module is configured to divide the sample data of each classification into training samples and test samples according to a preset ratio;

[0175] The eighth sub-module is configured to train the random forest classifier and the support vector machine classifier respectively using the training samples;

[0176] The ninth sub-module is configured to respectively predict the test samples corresponding to each classification according to the trained random forest classifier and the support vector machine classifier, and determine the test accuracy of the random forest classifier and the support vector machine classifier;

[0177] The tenth sub-module is configured to, in a case where the test accuracy does not meet a preset condition, respectively classify unlabelled forest disturbance data according to the trained random forest classifier and the support vector machine classifier, and obtain a first classification result and a second classification result;

[0178] The eleventh sub-module is configured to determine training sample increment samples of the random forest classifier and the support vector machine classifier according to the first classification result and the second classification result;

[0179] The twelfth sub-module is configured to train the random forest classifier using the increment training samples corresponding to the random forest classifier, and train the support vector machine classifier using the increment training samples corresponding to the support vector machine classifier, so as to complete one mutual training of the random forest classifier and the support vector machine classifier.

[0180] Optionally, the eleventh sub-module is specifically configured to:

[0181] Compare the first classification result and the second classification result, and determine the classification result with a difference as pseudo-label data;

[0182] Filter first high-confidence pseudo-label data from the pseudo-label data contained in the first classification result, and take the first high-confidence pseudo-label data as training sample increment samples of the support vector machine classifier;

[0183] Filter second high-confidence pseudo-label data from the pseudo-label data contained in the second classification result, and take the second high-confidence pseudo-label data as training sample increment samples of the random forest classifier.

[0184] Optionally, when the eleventh sub-module filters first high-confidence pseudo-label data from the pseudo-label data contained in the first classification result, the eleventh sub-module is specifically configured to:

[0185] From the pseudo-label data included in the first classification result, filter out the first pseudo-label data whose classification probability value is greater than the first preset probability threshold;

[0186] The erroneous pseudo-label data in the first pseudo-label data is filtered out using a confidence learning algorithm to obtain the first high-confidence pseudo-label data.

[0187] Optionally, when the eleventh submodule uses the first high-confidence pseudo-label data as incremental training samples for the support vector machine classifier, it is specifically used for:

[0188] For each category included in the first classification result, determine the first high-confidence pseudo-label data corresponding to each category;

[0189] Perform inter-class data balancing on the first high-confidence pseudo-label data corresponding to each category;

[0190] The first high-confidence pseudo-label data after inter-class balancing is used as the incremental training sample for the support vector machine classifier.

[0191] The device for determining the type of forest disturbance provided in this application combines few-shot learning with the mutual learning method of two base classifiers, random forest and support vector machine, to train the model. Users only need to manually label a small number of samples. Subsequently, the amount of sample data can be increased by iterative mutual learning, which can save the time and labor cost of sample labeling.

[0192] The embodiments provided in this application Figure 4 The device shown for determining the type of forest disturbance can achieve Figure 1 The various processes implemented in the method implementation examples will not be described again here to avoid repetition.

[0193] This invention also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0194] Memory, used to store computer programs;

[0195] When the processor executes a program stored in memory, it implements any of the methods for determining the type of forest disturbance described in the above embodiments.

[0196] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0197] The communication interface is used for communication between the terminal and other devices.

[0198] The memory can include a Random Access Memory (RAM) and can also include a non-volatile memory, such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the processor.

[0199] The processor mentioned above can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0200] In another embodiment provided by the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores instructions, when the instructions are run on a computer, the computer is caused to implement the method for determining the forest disturbance type in any of the above embodiments.

[0201] In another embodiment provided by the present application, a computer program product containing instructions is also provided, when the instructions are run on a computer, the computer is caused to implement the method for determining the forest disturbance type in any of the above embodiments.

[0202] In the embodiments described above, all or some of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or some of the embodiments can be implemented in the form of a computer program product storing computer program instructions. The computer program instructions are loaded in a computer and executed, so as to produce all or some of the processes or functions described by the embodiments of the present application. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer program instructions can be stored in a computer readable storage medium, or transferred from one computer readable storage medium to another computer readable storage medium, for example, the computer program instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (for example, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available media. The available medium can be a magnetic medium (for example, floppy disk, hard disk, magnetic tape), an optical medium (for example, DVD), or a semiconductor medium (for example, solid state disk (SSD)), etc.

[0203] It should be noted that, in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0204] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered as the protection scope of the present application.

Claims

1. A method of determining a type of forest disturbance, characterized by, The method comprises: obtaining forest disturbance data from a land cover dataset CLCD; wherein the disturbance data comprises: disturbance time and disturbance occurrence position; performing feature analysis on the forest disturbance data to obtain forest disturbance data under each classification; based on the forest disturbance data under each classification as basic data, using the mutual induction mode of random forest classifier and support vector machine classifier to perform sample data volume expansion and classifier training multiple times to obtain a random forest classifier and a support vector machine classifier satisfying a preset condition; based on the random forest classifier and the support vector machine classifier satisfying the preset condition, predicting the forest disturbance data to be predicted to obtain a forest disturbance type.

2. The method of claim 1, wherein, The step of obtaining forest disturbance data from a land cover dataset comprises: performing forest mask on the initial land cover dataset to obtain forest-related data; using a continuous change detection and classification algorithm CCDC to detect the forest-related data to obtain a detection result, wherein the detection result includes: forest change to other land disturbance and other land change to forest disturbance; based on the land cover dataset, filtering disturbance pixels of the forest change to other land disturbance to obtain forest disturbance data.

3. The method of claim 1, wherein, The step of performing feature analysis on the forest disturbance data to obtain forest disturbance data under each classification comprises: based on the time feature, spatial feature, spectral feature and terrain feature of subtropical evergreen forest disturbance classification, analyzing the forest disturbance data to obtain the classification corresponding to each forest disturbance; forest disturbance data of the same classification is divided into the same classification to obtain forest disturbance data under each classification.

4. The method of claim 1, wherein, The step of using the forest disturbance data under each classification as basic data, using the mutual induction mode of random forest classifier and support vector machine classifier to perform sample data volume expansion and classifier training multiple times to obtain a random forest classifier and a support vector machine classifier satisfying a preset condition comprises: for each classification of forest disturbance data, taking a preset number of forest disturbance data as sample data; dividing the sample data of each classification into training samples and test samples according to a preset proportion; training the random forest classifier and the support vector machine classifier using the training samples respectively; predicting the test samples corresponding to each classification according to the trained random forest classifier and support vector machine classifier respectively to determine the test accuracy of the random forest classifier and the support vector machine classifier; in the case where the test accuracy does not satisfy the preset condition, classifying unmarked forest disturbance data according to the trained random forest classifier and support vector machine classifier respectively to obtain a first classification result and a second classification result; determining the training sample increment sample of the random forest classifier and the support vector machine classifier according to the first classification result and the second classification result. The random forest classifier is trained by using the incremental training sample corresponding to the random forest classifier, and the support vector machine classifier is trained by using the incremental training sample corresponding to the support vector machine classifier, so as to complete one iteration of mutual training of the random forest classifier and the support vector machine classifier.

5. The method of claim 4, wherein, The step of determining the training sample increment sample of the random forest classifier and the support vector machine classifier according to the first classification result and the second classification result comprises: By comparing the first classification result and the second classification result, the classification result with differences is determined as pseudo-label data; First high-confidence pseudo-label data is screened from the pseudo-label data contained in the first classification result, and the first high-confidence pseudo-label data is used as the training sample increment sample of the support vector machine classifier; Second high-confidence pseudo-label data is screened from the pseudo-label data contained in the second classification result, and the second high-confidence pseudo-label data is used as the training sample increment sample of the random forest classifier.

6. The method of claim 5, wherein, The step of screening first high-confidence pseudo-label data from the pseudo-label data contained in the first classification result comprises: First pseudo-label data with a classification probability value greater than a first preset probability threshold is screened from the pseudo-label data contained in the first classification result; Error pseudo-label data in the first pseudo-label data is filtered out by using a confidence learning algorithm to obtain first high-confidence pseudo-label data.

7. The method of claim 5, wherein, The step of using the first high-confidence pseudo-label data as the training sample increment sample of the support vector machine classifier comprises: First high-confidence pseudo-label data corresponding to each classification in the first classification result is determined; The first high-confidence pseudo-label data corresponding to each classification is subjected to inter-class data balancing processing; The first high-confidence pseudo-label data subjected to inter-class balancing processing is used as the training sample increment sample of the support vector machine classifier.

8. An apparatus for determining the type of forest disturbance, characterized in that, The device comprises: An acquisition module is configured to acquire forest disturbance data from a land cover dataset (CLCD). The disturbance data comprises a disturbance time and a disturbance occurrence position. A classification module is configured to perform feature analysis on the forest disturbance data to obtain forest disturbance data under each classification. A mutual training module is configured to use the forest disturbance data under each classification as basic data, and iteratively perform sample data volume expansion and classifier training in a manner of mutual training of a random forest classifier and a support vector machine classifier, to obtain a random forest classifier and a support vector machine classifier satisfying a preset condition. A prediction module is configured to predict forest disturbance data to be predicted based on the random forest classifier and the support vector machine classifier satisfying the preset condition, to obtain a forest disturbance type.

9. The apparatus of claim 8, wherein, The acquisition module comprises: A first sub-module is configured to perform forest mask on an initial land cover dataset to obtain forest-related data. The second sub-module is configured to detect the forest-related data by using a continuous change detection and classification algorithm (CCDC) to obtain a detection result, wherein the detection result includes disturbances in which forest changes into other land types and disturbances in which other land types change into forest. The third sub-module is configured to perform disturbance pixel filtering on the disturbances in which forest changes into other land types based on a land cover data set to obtain forest disturbance data.

10. An electronic device, comprising: The device comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus. The memory is configured to store a computer program. The processor is configured to execute the program stored on the memory to implement the method for determining a forest disturbance type according to any one of claims 1-7.

Images