Classification and recognition method of evergreen broad-leaved forest vegetation subtypes based on hierarchical multi-label network

Abstract

The present invention discloses a method for classifying and identifying evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network, comprising: obtaining multi-source remote sensing data of a sample area and extracting classification features from the multi-source remote sensing data; constructing a sample library of evergreen broad-leaved forests and their subtypes in the sample area based on field survey data and an existing vegetation classification system; constructing a hierarchical multi-label classification network based on the sample library and the classification features, using a multi-level, sequentially arranged classifier to form a hierarchical multi-label classification network; and classifying and identifying the evergreen broad-leaved forest vegetation in the target area using the hierarchical multi-label classification network. The present invention constructs a hierarchical classification system for forests, evergreen forests, evergreen broad-leaved forests, and dry / wet evergreen broad-leaved forests, addressing the shortcomings of traditional methods in distinguishing similar vegetation types. At the same time, through a hierarchical error weighting mechanism, the impact of low-level classification errors on the overall results is reduced, thereby improving the accuracy and reliability of the classification of dry and wet evergreen broad-leaved forests.

Description

Technical Field

[0001] The present invention relates to the fields of remote sensing technology and ecological environment monitoring, and in particular to a method for classifying and identifying evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network. Background Art

[0002] As a globally important gene bank, subtropical evergreen broad-leaved forests play a vital role in maintaining regional soil and water resources, regulating climate, and maintaining ecosystem balance and stability. In-depth research on a quantitative classification system for subtropical evergreen broad-leaved forests and their vegetation subtypes will help better understand regional evergreen broad-leaved forest resources and facilitate their rational planning and utilization.

[0003] The complexity of vegetation information extraction and classification lies in the need to accurately group vegetation types based on their similarities. This process requires consideration not only of the physical characteristics of vegetation but also of the dispersed and discontinuous nature of its geographical distribution, making this task highly generalizable. Remote sensing technology, with its extensive coverage, rich information content, and short return period, has become a powerful tool for vegetation information extraction and classification. By analyzing spectral, textural, and temporal features in remote sensing images and combining them with classifiers, different vegetation types can be effectively distinguished and identified, resulting in accurate vegetation distribution maps. Although research on forest type classification and fine-grained identification based on remote sensing imagery is extensive, more detailed classification of vegetation subtypes within vegetation types remains insufficient.

[0004] There are currently two key issues: (1) The widespread distribution of evergreen broad-leaved forests and the complex and varied terrain pose significant challenges to the acquisition and preprocessing of remote sensing images. (2) Scale and feature selection for evergreen broad-leaved forest vegetation extraction and classification: How to select the most suitable remote sensing image scale, vegetation features, and classification methods for evergreen broad-leaved forest extraction and subtype classification is still underdeveloped. The impact and applicability of different vegetation characteristics (such as spectral characteristics, canopy characteristics, geometric structure, etc.) and classification algorithms (such as supervised classification, unsupervised classification, machine learning, etc.) on classification results need further exploration. Summary of the Invention

[0005] In order to at least overcome the above-mentioned deficiencies in the prior art, the purpose of the present application is to provide a method for classification and identification of evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network.

[0006] In a first aspect, the present application provides a method for classifying and identifying evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network, comprising:

[0007] Acquiring multi-source remote sensing data of a sample area and extracting classification features from the multi-source remote sensing data; the multi-source remote sensing data includes optical image data and radar image data; the classification features include spectral features, texture features, and time series features;

[0008] Construct a sample library of evergreen broad-leaved forests and their subtypes in the sample area based on field survey data and existing vegetation classification systems;

[0009] Constructing a hierarchical multi-label classification network based on the sample library and the classification features; the classification result of each level of classifier is a subdivision of the classification result of the upper level classifier, and the classification error weight of each level of classifier decreases along the hierarchy;

[0010] The evergreen broad-leaved forest vegetation in the target area is classified and identified through the hierarchical multi-label classification network.

[0011] In one possible implementation, building a hierarchical multi-label classification network includes:

[0012] assigning multi-level classification labels to the samples according to the vegetation types of the samples in the sample library;

[0013] According to the hierarchical relationship of the multi-level classification labels, multiple classifiers are trained from the top level to the bottom level to form an initial hierarchical network; each classifier subdivides and classifies one of the classification results of the previous level classifier;

[0014] The classification results of non-bottom-level classifiers are trimmed to eliminate the classification results that do not require the next-level classifier to perform subdivision classification;

[0015] The classification result of the bottom-level classifier is used as the classification result output by the hierarchical multi-label classification network.

[0016] In one possible implementation, the first level of the multi-level classification label is forest and non-forest; the second level of the multi-level classification label is evergreen forest and non-green forest; the third level of the multi-level classification label is evergreen broad-leaved forest and non-green broad-leaved forest; the fourth level of the multi-level classification label is dry evergreen broad-leaved forest and wet evergreen broad-leaved forest.

[0017] In one possible implementation, training multiple classifiers includes:

[0018] A first-level classifier is trained based on the samples of the first-level label and the spectral features; the input data of the first-level classifier is the spectral features, and the output data of the first-level classifier is classified as forest or non-forest;

[0019] A second-level classifier is trained based on the samples of the second-level label, the spectral features, and the time series features; the input data of the second-level classifier is the spectral features and the time series features, and the output data of the second-level classifier is classified as evergreen forest or non-green forest;

[0020] A third-level classifier is trained based on the samples of the third-level label, the spectral features, and the texture features; the input data of the third-level classifier is the spectral features and the texture features, and the output data of the third-level classifier is classified as evergreen broad-leaved forest or non-evergreen broad-leaved forest;

[0021] A fourth-level classifier is trained based on the samples of the fourth-level label, the spectral features, the texture features, and the time series features; the input data of the fourth-level classifier is the spectral features, the texture features, and the time series features, and the output data of the fourth-level classifier is classified as dry evergreen broad-leaved forest and wet evergreen broad-leaved forest;

[0022] The data classified as forest output by the first-level classifier is used as the input data of the second-level classifier; the data classified as evergreen forest output by the second-level classifier is used as the input data of the third-level classifier; the data classified as evergreen broad-leaved forest output by the third-level classifier is used as the input data of the fourth-level classifier; and the classification result output by the fourth-level classifier is used as the output result of the hierarchical multi-label classification network.

[0023] In one possible implementation, the training of the fourth-level classifier includes:

[0024] The evergreen broad-leaved forest humidity index is constructed using the following formula:

[0025]

[0026] Wherein, EBFHI is the evergreen broad-leaved forest wetness index, NIR is the near-infrared reflectance, and SWIR1 is the short-wave infrared reflectance;

[0027] The evergreen broad-leaved forest wetness index is used as a spectral feature input during the training of the fourth-level classifier.

[0028] In a possible implementation, the first-level classifier, the second-level classifier, the third-level classifier, and the fourth-level classifier adopt at least one of a support vector machine, a random forest model, and a gradient boosting tree.

[0029] In one possible implementation, the classification error weight of each level of classifier is calculated according to the following formula:

[0030]

[0031] Where, w(y i) is the node y in the hierarchical multi-label classification network i The classification error weight, For node y i At the level in the hierarchical multi-label classification network, maxlevel is the length of the longest path in the hierarchical multi-label classification network.

[0032] Secondly, this application also provides a classification and identification system for evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network, including:

[0033] an acquisition unit configured to acquire multi-source remote sensing data of a sample area and extract classification features from the multi-source remote sensing data; the multi-source remote sensing data includes optical image data and radar image data; the classification features include spectral features, texture features, and time series features;

[0034] A sample unit is configured to construct a sample library of evergreen broad-leaved forests and their subtypes in the sample area based on field survey data and an existing vegetation classification system;

[0035] A modeling unit is configured to construct a hierarchical multi-label classification network by sequentially setting multiple levels of classifiers based on the sample library and the classification features; the classification result of each level of classifier is a subdivision of the classification result of the upper level classifier, and the classification error weight of each level of classifier decreases sequentially along the hierarchy;

[0036] The identification unit is configured to classify and identify the evergreen broad-leaved forest vegetation in the target area through the hierarchical multi-label classification network.

[0037] In a possible implementation, the modeling unit is further configured to:

[0038] assigning multi-level classification labels to the samples according to the vegetation types of the samples in the sample library;

[0039] According to the hierarchical relationship of the multi-level classification labels, multiple classifiers are trained from the top level to the bottom level to form an initial hierarchical network; each classifier subdivides and classifies one of the classification results of the previous level classifier;

[0040] The classification results of non-bottom-level classifiers are trimmed to eliminate the classification results that do not require the next-level classifier to perform subdivision classification;

[0041] The classification result of the bottom-level classifier is used as the classification result output by the hierarchical multi-label classification network.

[0042] In one possible implementation, the classification error weight of each level of classifier is calculated according to the following formula:

[0043]

[0044] Where, w(y i ) is the node y in the hierarchical multi-label classification network i The classification error weight, For node y i At the level in the hierarchical multi-label classification network, maxlevel is the length of the longest path in the hierarchical multi-label classification network.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] The present invention is based on a hierarchical multi-label network-based evergreen broad-leaved forest vegetation subtype classification and identification method, which constructs a hierarchical classification system for forests, evergreen forests, evergreen broad-leaved forests, and dry / wet evergreen broad-leaved forests, solving the shortcomings of traditional methods in distinguishing similar vegetation types; at the same time, through a hierarchical error weighting mechanism, it reduces the impact of low-level classification errors on the overall results, thereby improving the accuracy and reliability of the classification of dry and wet evergreen broad-leaved forests. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0048] Figure 1 This is a schematic diagram of the steps of the method according to the embodiment of the present application;

[0049] Figure 2 This is a schematic diagram of the hierarchical architecture of an embodiment of the present application;

[0050] Figure 3 This is a schematic diagram of the spectrum curves of the dry evergreen broad-leaved forest and the moist evergreen broad-leaved forest in the embodiment of the present application;

[0051] Figure 4 This is a schematic diagram of the classification results of the first-level classifier in the embodiment of the present application;

[0052] Figure 5 This is a schematic diagram of the classification results of the second-level classifier in the embodiment of the present application;

[0053] Figure 6 This is a schematic diagram of the classification results of the third-level classifier in the embodiment of the present application;

[0054] Figure 7 This is a schematic diagram of the classification results of the fourth-level classifier in an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of illustration and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate the operations implemented according to some embodiments of the embodiments of the present application. It should be understood that the operations of the flowcharts can be implemented out of sequence, and steps without logical context can be reversed or implemented simultaneously. In addition, those skilled in the art, under the guidance of the contents of this application, can add one or more other operations to the flowchart, or remove one or more operations from the flowchart.

[0056] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0057] Please refer to Figure 1 , which is a flow chart of the method for classifying and identifying evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network provided in an embodiment of the present invention. Furthermore, the method for classifying and identifying evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network can specifically include the contents described in the following steps S1 to S4.

[0058] S1: Acquire multi-source remote sensing data of a sample area and extract classification features from the multi-source remote sensing data; the multi-source remote sensing data includes optical image data and radar image data; the classification features include spectral features, texture features, and time series features;

[0059] S2: Construct a sample library of evergreen broad-leaved forests and their subtypes in the sample area based on field survey data and existing vegetation classification systems;

[0060] S3: constructing a hierarchical multi-label classification network based on the sample library and the classification features, wherein the classification result of each level of the classifier is a subdivision of the classification result of the upper level classifier, and the classification error weight of each level of the classifier decreases along the hierarchy;

[0061] S4: Classify and identify the evergreen broad-leaved forest vegetation in the target area through the hierarchical multi-label classification network.

[0062] When implementing the embodiment of the present application, the optical image data also needs to be preprocessed; specifically, the preprocessing of the optical image data includes: obtaining the dry season image data and wet season image data of the target year from the optical image data and removing clouds and cloud shadows to form first preprocessed image data; obtaining the impact data of the corresponding seasons of the adjacent years of the target year from the optical image data and synthesizing them to form second preprocessed image data; extracting the images of the vacant positions in the first preprocessed image data from the second preprocessed image data, and performing interpolation processing at the vacant positions of the first preprocessed image data to fill the vacant positions to form third preprocessed image data to complete the preprocessing.

[0063] For this example, Landsat 8 optical imagery was used. Clouds and cloud shadows were effectively removed from 2020 using the 'QA_PIXEL' quality assessment band of the Landsat 8 imagery. The 'QA_PIXEL' layer contains information about clouds, cloud shadows, and other factors that may affect image quality. By creating a mask, clouds and cloud shadows can be effectively identified and removed. Subsequently, remote sensing imagery from the dry and wet seasons of 2019 (dry season: November 1, 2018, to March 31, 2019; wet season: May 1, 2019, to September 30, 2021) and 2021 (dry season: November 1, 2020, to March 31, 2021; wet season: May 1, 2021, to September 30, 2021) was synthesized. Finally, the synthesized dry and wet season remote sensing images of 2019 and 2021 were used to interpolate the null values ​​after removing clouds and cloud shadows in 2020 to obtain clearer and more coherent remote sensing images.

[0064] In the embodiments of the present application, radar image data is insensitive to clouds and can obtain surface information under any weather conditions, which can effectively supplement optical image data. For example, radar image data uses Sentinel-1, which also requires corresponding preprocessing, such as boundary noise correction and speckle filtering. Different preprocessing methods can be selected according to research needs.

[0065] When implementing the embodiment of this application, it is necessary to build a sample library based on field survey data and the existing vegetation classification system;

[0066] For example, the existing vegetation classification system uses the classification systems of the Flora of China (FOC) and the Sichuan Vegetation (1980). When constructing a sample library, samples can be divided into four levels: forest and non-forest, evergreen forest and non-green forest, evergreen broad-leaved forest and non-green broad-leaved forest, and dry evergreen broad-leaved forest and moist evergreen broad-leaved forest. For example, when classifying 6,866 sample plots obtained through field surveys, taking the founding species of the sample plot as Phoebe zhennan, if the FOC classification system records Phoebe zhennan as an evergreen tree, the sample plot would be classified as evergreen forest. However, as a broad-leaved tree, Phoebe zhennan would be classified as evergreen broad-leaved forest. However, according to the Sichuan Vegetation (1980) classification system, the corresponding sample plot data would be classified as dry evergreen broad-leaved forest or moist evergreen broad-leaved forest.

[0067] In the embodiment of the present application, a hierarchical multi-label classification network can be constructed based on the classification content in the above sample library. Figure 2 , shows a schematic diagram of the architecture of the hierarchical multi-label classification network of an embodiment of the present application, which satisfies the hierarchical constraint, that is, any node in the label hierarchy can be labeled as positive only if the node is the root node or the parent node of the node is labeled as positive. From the top-level categories (such as "forest" and "non-forest") to more detailed categories (such as "dry evergreen broad-leaved forest" and "humid evergreen broad-leaved forest"), the output of each layer provides a possible classification path for the next layer. In this hierarchical classification system, an instance may be assigned multiple labels, each of which corresponds to a node in the hierarchy. The label of a parent node is directly affected by the labels of all its child nodes: if an instance belongs to a child node, it should also belong to all parent nodes related to the child node. This hierarchical structure allows the model to recognize not only more general categories, but also more specific subcategories. It reflects the natural hierarchy of land feature classification, where higher-level categories are more abstract and lower-level categories are more specific. When training such models, one needs to ensure that the correct hierarchical relationships are learned and predicted, which is usually achieved through hierarchical errors and embedding hierarchical logic in the model structure.

[0068] In an embodiment of the present application, for the label nodes in the hierarchical structure, their hierarchical positions are taken into consideration to determine their weights in all scores. Normally, the classification error cost occurring at the top of the hierarchical label tree is higher than the error cost at the bottom. This is because the top of the hierarchical tree has more training samples and the differences between categories are more significant, so it contributes more to the study of classification problems. Therefore, the inventors assign higher weights to the nodes at the top of the hierarchical tree, while the weights of the bottom nodes are relatively low. This weight distribution method helps to more accurately weigh the importance of nodes in different positions during the classification process, and can effectively improve the accuracy and efficiency of classification. In an embodiment of the present application, the classifier of each level needs to be trained with data from the corresponding sample library. After completing the training, the complete hierarchical multi-label classification network can be directly applied to the classification of evergreen broad-leaved forest vegetation in the target area to identify the dry evergreen broad-leaved forest and the wet evergreen broad-leaved forest in the target area.

[0069] In one possible implementation, building a hierarchical multi-label classification network includes:

[0070] assigning multi-level classification labels to the samples according to the vegetation types of the samples in the sample library;

[0071] According to the hierarchical relationship of the multi-level classification labels, multiple classifiers are trained from the top level to the bottom level to form an initial hierarchical network; each classifier subdivides and classifies one of the classification results of the previous level classifier;

[0072] The classification results of non-bottom-level classifiers are trimmed to eliminate the classification results that do not require the next-level classifier to perform subdivision classification;

[0073] The classification result of the bottom-level classifier is used as the classification result output by the hierarchical multi-label classification network.

[0074] In one possible implementation, the first level of the multi-level classification label is forest and non-forest; the second level of the multi-level classification label is evergreen forest and non-green forest; the third level of the multi-level classification label is evergreen broad-leaved forest and non-green broad-leaved forest; the fourth level of the multi-level classification label is dry evergreen broad-leaved forest and wet evergreen broad-leaved forest.

[0075] When the embodiment of the present application is implemented, it is necessary to label the samples in the sample library with multi-level classification labels. It should be understood that in order for the samples to be applied to the training of multi-level classifiers, if the same sample can be further subdivided, it will be given multiple labels. For example, if the founding species of a sample is Cyclobalanopsis gracilis, it will be given four labels: forest, evergreen forest, evergreen broad-leaved forest, and dry evergreen broad-leaved forest. When training the model, when training the first-level classifier, the sample is trained with the label forest; when training the second-level classifier, the sample is trained with the label evergreen forest; when training the third-level classifier, the sample is trained with the label evergreen broad-leaved forest; and when training the fourth-level classifier, the sample is trained with the label dry evergreen broad-leaved forest. In the embodiment of the present application, by clipping the classification results of the classifiers other than the bottom one, the data that does not need to be processed subsequently can be eliminated, thereby improving the computational efficiency of the model.

[0076] In one possible implementation, training multiple classifiers includes:

[0077] A first-level classifier is trained based on the samples of the first-level label and the spectral features; the input data of the first-level classifier is the spectral features, and the output data of the first-level classifier is classified as forest or non-forest;

[0078] A second-level classifier is trained based on the samples of the second-level label, the spectral features, and the time series features; the input data of the second-level classifier is the spectral features and the time series features, and the output data of the second-level classifier is classified as evergreen forest or non-green forest;

[0079] A third-level classifier is trained based on the samples of the third-level label, the spectral features, and the texture features; the input data of the third-level classifier is the spectral features and the texture features, and the output data of the third-level classifier is classified as evergreen broad-leaved forest or non-evergreen broad-leaved forest;

[0080] A fourth-level classifier is trained based on the samples of the fourth-level label, the spectral features, the texture features, and the time series features; the input data of the fourth-level classifier is the spectral features, the texture features, and the time series features, and the output data of the fourth-level classifier is classified as dry evergreen broad-leaved forest and wet evergreen broad-leaved forest;

[0081] The data classified as forest output by the first-level classifier is used as the input data of the second-level classifier; the data classified as evergreen forest output by the second-level classifier is used as the input data of the third-level classifier; the data classified as evergreen broad-leaved forest output by the third-level classifier is used as the input data of the fourth-level classifier; and the classification result output by the fourth-level classifier is used as the output result of the hierarchical multi-label classification network.

[0082] When implementing the embodiment of the present application, the features used in training each layer are shown in the following table:

[0083] Table 1 Hierarchical feature table

[0084]

[0085] Among them, blue, red, green, nir, swir1, and swir2 are the blue band, red band, green band, near-infrared band, shortwave infrared band 1, and shortwave infrared band 2 of Landsat 8 OLI, respectively. NDVI is the Normalized Difference Vegetation Index; EVI is the Enhanced Vegetation Index; and FDI is the Forest Differentiation Index. FDI is calculated according to the following formula:

[0086]

[0087] Due to the introduction of FDI, the sum of the red and green band reflectances can more clearly distinguish forest vegetation compared with the reflectance of the near-infrared band.

[0088] For the second-level classification, compared with the first-level features, temporal features are introduced, that is, the features brought about by changes in different seasons. The reason is that the key to distinguishing evergreen forests from non-green forests is to identify the phenological characteristics of vegetation. There are significant differences in phenology between evergreen forests and non-green forests. Evergreen forests show smaller NDVI changes due to green leaves throughout the year, while non-green forests, especially deciduous forests, have significant fluctuations in NDVI due to seasonal leaf growth and shedding. Remote sensing images acquired in the dry season are particularly critical for identifying evergreen forests, as the NDVI values ​​in the dry season will drop significantly. By comparing the differences in NDVI and EVI between the dry and wet seasons, evergreen forests and non-green forests can be effectively distinguished. This method improves the accuracy of vegetation type identification by analyzing the changes in vegetation greenness index in different seasons. Time series remote sensing data analysis can not only capture the vegetation status in a specific season, but also track the dynamic changes of vegetation over time, which is crucial for accurately identifying vegetation types and conducting long-term ecological monitoring. Among them, NDVI wet NDVI value calculated from remote sensing images in the wet season, EVI wet EVI value calculated from remote sensing images in the wet season, NDVI dry NDVI value calculated from remote sensing images in the dry season, EVI dry EVI value calculated from remote sensing images in the dry season; NDVI diff and EVI diff Calculated according to the following formula:

[0089]

[0090]

[0091] For the third-level classification, texture features are introduced, compared to the first-level. Eleven texture features—Angular Second Moment (ASM), Contrast, Correlation, Variance, Homogeneity, Sum Entropy, Entropy, Dissimilarity, Mean, Cluster Prominence (CP), and Cluster Shade (CS)—are commonly used for vegetation type extraction and classification. Texture features are introduced because the spectral information of ground objects, i.e., their ability to reflect electromagnetic energy, directly reflects their characteristics and is widely used in remote sensing data analysis and interpretation. However, the application of spectral information is subject to interference and influence from various factors, including atmospheric conditions, intrinsic variability of the ground objects, and varying observation conditions. These variations can result in different ground objects exhibiting similar spectral characteristics, or the same ground object exhibiting different spectral characteristics under different conditions, leading to the phenomenon of "different objects exhibiting the same spectrum" or "same object exhibiting different spectra." This phenomenon is particularly pronounced when distinguishing between landforms with similar spectral characteristics. Studies have shown that while broadleaf forests and non-broadleaf forests are difficult to distinguish spectrally, they exhibit distinct texture differences in remote sensing images.

[0092] In one possible implementation, the training of the fourth-level classifier includes:

[0093] The evergreen broad-leaved forest humidity index is constructed using the following formula:

[0094]

[0095] Wherein, EBFHI is the evergreen broad-leaved forest wetness index, NIR is the near-infrared reflectance, and SWIR1 is the short-wave infrared reflectance;

[0096] The evergreen broad-leaved forest wetness index is used as a spectral feature input during the training of the fourth-level classifier.

[0097] For the fourth level classification, EBFHI, RVI, NDII, MSI, TC1, TC2, TC3 and Sentinel-1VV / VH are newly introduced in the spectral characteristics. Among them, EBFHI evergreen broad-leaved forest humidity index, please refer to Figure 3 , Figure 3The spectral curves of dry evergreen broad-leaved forests and moist evergreen broad-leaved forests are shown; the vertical axis in the figure is reflectance, and the horizontal axis is spectral band. The colored areas indicate the spectral distribution of different evergreen broad-leaved forests. The blue line is the boundary curve of dry evergreen broad-leaved forests, and the red line is the boundary curve of moist evergreen broad-leaved forests. The reason for introducing EBFHI is that the sampling time of field survey points is mostly concentrated in the wet season. Therefore, the spectral curves of dry and moist evergreen broad-leaved forests in the sample library in the wet season remote sensing images were analyzed. It was found that the reflectance of moist evergreen broad-leaved forests in the near-infrared (Nir) and short-wave infrared (Swir1) bands is higher than that of dry evergreen broad-leaved forests in these two bands. This suggests that during the wet season, vegetation moisture in moist evergreen broad-leaved forests may be higher than in dry evergreen broad-leaved forests, or that moist evergreen broad-leaved forests have denser vegetation structures, resulting in higher reflectance. Introducing this feature enhances the differences between dry and moist evergreen broad-leaved forests in the Nir and Swir1 bands. The Ratio Vegetation Index (RVI), Normalized Difference Infrared Index (NDII), and Water Stress Index (MSI) reflect canopy moisture content by combining the absorption characteristics of water in the near-infrared and short-wave infrared ranges with the penetration of light in the near-infrared range. TC1, TC2, and TC3 represent the brightness, greenness, and yellowness of vegetation, respectively, expressed through the Tasseled Cap Transform. Sentinel-1VV / VH, which uses VV (vertical transmission and vertical reception) and VH (vertical transmission and horizontal reception) polarization modes of Sentinel-1, is more sensitive to the geometric structure and orientation of vegetation, making it superior in assessing vegetation moisture. It generally provides more information in detecting vegetation moisture content and can serve as a key indicator of characteristic information for classifying dry evergreen broad-leaved forests and moist evergreen broad-leaved forests.

[0098] For the fourth level classification, the temporal characteristics are the data of RVI, NDII and MSI in the dry and wet seasons, as well as the difference between the dry and wet season data. Specifically, RVI wet , RVI dry , RVI diff NDII wet NDII diff NDII wet 、MSI wet 、MSI dry 、MSI diff As the layers deepen, the set of selected features gradually increases. Each level of features not only increases the detail and depth of extraction and classification, but also reflects the natural classification logic from abstract to concrete. This hierarchical feature selection improves the accuracy and efficiency of extraction and classification.

[0099] For example, the results of classification by the trained first-level classifier are Figure 4, the results of classification by the trained second-level classifier are Figure 5 , the results of classification by the trained third-level classifier are Figure 6 , the results of classification by the trained fourth-level classifier are included in Figure 7 .

[0100] In a possible implementation, the first-level classifier, the second-level classifier, the third-level classifier, and the fourth-level classifier adopt at least one of a support vector machine, a random forest model, and a gradient boosting tree.

[0101] In one possible implementation, the classification error weight of each level of classifier is calculated according to the following formula:

[0102]

[0103] Where, w(y i ) is the node y in the hierarchical multi-label classification network i The classification error weight, For node y i At the level in the hierarchical multi-label classification network, maxlevel is the length of the longest path in the hierarchical multi-label classification network.

[0104] When the embodiment of the present application is implemented, the node weight calculated by the above formula can decay linearly as the node level decreases, thereby ensuring that the weight is evenly distributed along the level. The weights of the lower-level nodes will neither drop to 0 quickly nor decay too slowly. For example, the weights of the hierarchical architecture are 1 for the first layer, 0.75 for the second layer, 0.5 for the third layer, and 0.25 for the fourth layer. After normalization, the weight of the first layer is 0.4, the weight of the second layer is 0.3, the weight of the third layer is 0.2, and the weight of the fourth layer is 0.1. This formula not only reflects the transmission and accumulation of errors between levels, but also reflects the different influences of each level in the overall task through weights, thereby giving a quantitative evaluation of the overall performance of the model.

[0105] Based on the same inventive concept, the present application also provides an evergreen broad-leaved forest vegetation subtype classification and identification system based on a hierarchical multi-label network, including:

[0106] an acquisition unit configured to acquire multi-source remote sensing data of a sample area and extract classification features from the multi-source remote sensing data; the multi-source remote sensing data includes optical image data and radar image data; the classification features include spectral features, texture features, and time series features;

[0107] A sample unit is configured to construct a sample library of evergreen broad-leaved forests and their subtypes in the sample area based on field survey data and an existing vegetation classification system;

[0108] A modeling unit is configured to construct a hierarchical multi-label classification network by sequentially setting multiple levels of classifiers based on the sample library and the classification features; the classification result of each level of classifier is a subdivision of the classification result of the upper level classifier, and the classification error weight of each level of classifier decreases sequentially along the hierarchy;

[0109] The identification unit is configured to classify and identify the evergreen broad-leaved forest vegetation in the target area through the hierarchical multi-label classification network.

[0110] In a possible implementation, the modeling unit is further configured to:

[0111] assigning multi-level classification labels to the samples according to the vegetation types of the samples in the sample library;

[0112] According to the hierarchical relationship of the multi-level classification labels, multiple classifiers are trained from the top level to the bottom level to form an initial hierarchical network; each classifier subdivides and classifies one of the classification results of the previous level classifier;

[0113] The classification results of non-bottom-level classifiers are trimmed to eliminate the classification results that do not require the next-level classifier to perform subdivision classification;

[0114] The classification result of the bottom-level classifier is used as the classification result output by the hierarchical multi-label classification network.

[0115] In one possible implementation, the classification error weight of each level of classifier is calculated according to the following formula:

[0116]

[0117] Where, w(y i ) is the node y in the hierarchical multi-label classification network i The classification error weight, For node y i At the level in the hierarchical multi-label classification network, maxlevel is the length of the longest path in the hierarchical multi-label classification network.

[0118] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0119] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or can be electrical, mechanical or other forms of connection.

[0120] The units described as separate components may or may not be physically separated. As units, it is obvious that a person of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0121] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0122] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or grid device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0123] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A classification and identification method for evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network, characterized by: include: Acquiring multi-source remote sensing data of a sample area and extracting classification features from the multi-source remote sensing data; The multi-source remote sensing data includes optical image data and radar image data; the classification features include spectral features, texture features and time series features; Construct a sample library of evergreen broad-leaved forests and their subtypes in the sample area based on field survey data and existing vegetation classification systems; Constructing a hierarchical multi-label classification network based on the sample library and the classification features; the classification result of each level of classifier is a subdivision of the classification result of the upper level classifier, and the classification error weight of each level of classifier decreases along the hierarchy; Classify and identify the evergreen broad-leaved forest vegetation in the target area through the hierarchical multi-label classification network; Building a hierarchical multi-label classification network involves: assigning multi-level classification labels to the samples according to the vegetation types of the samples in the sample library; According to the hierarchical relationship of the multi-level classification labels, multiple classifiers are trained from the top level to the bottom level to form an initial hierarchical network; each classifier subdivides and classifies one of the classification results of the previous level classifier; The classification results of non-bottom-level classifiers are trimmed to eliminate the classification results that do not require the next-level classifier to perform subdivision classification; The classification result of the bottom-level classifier is used as the classification result output by the hierarchical multi-label classification network; The first level of the multi-level classification label is forest and non-forest; the second level of the multi-level classification label is evergreen forest and non-green forest; the third level of the multi-level classification label is evergreen broad-leaved forest and non-green broad-leaved forest; the fourth level of the multi-level classification label is dry evergreen broad-leaved forest and wet evergreen broad-leaved forest; Training multiple classifiers involves: A first-level classifier is trained based on the samples of the first-level label and the spectral features; the input data of the first-level classifier is the spectral features, and the output data of the first-level classifier is classified as forest or non-forest; A second-level classifier is trained based on the samples of the second-level label, the spectral features, and the time series features; the input data of the second-level classifier is the spectral features and the time series features, and the output data of the second-level classifier is classified as evergreen forest or non-green forest; A third-level classifier is trained based on the samples of the third-level label, the spectral features, and the texture features; the input data of the third-level classifier is the spectral features and the texture features, and the output data of the third-level classifier is classified as evergreen broad-leaved forest or non-evergreen broad-leaved forest; A fourth-level classifier is trained based on the samples of the fourth-level label, the spectral features, the texture features, and the time series features; the input data of the fourth-level classifier is the spectral features, the texture features, and the time series features, and the output data of the fourth-level classifier is classified as dry evergreen broad-leaved forest and wet evergreen broad-leaved forest; The data classified as forest by the first-level classifier is used as the input data of the second-level classifier; the data classified as evergreen forest by the second-level classifier is used as the input data of the third-level classifier; the data classified as evergreen broad-leaved forest by the third-level classifier is used as the input data of the fourth-level classifier; and the classification result output by the fourth-level classifier is used as the output result of the hierarchical multi-label classification network; The training of the fourth-level classifier includes: The evergreen broad-leaved forest humidity index is constructed using the following formula: Wherein, EBFHI is the evergreen broad-leaved forest wetness index, NIR is the near-infrared reflectance, and SWIR1 is the short-wave infrared reflectance; The evergreen broad-leaved forest wetness index is used as a spectral feature input during the training of the fourth-level classifier.

2. The method for classification and identification of evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network according to claim 1, characterized in that: The first-level classifier, the second-level classifier, the third-level classifier and the fourth-level classifier adopt at least one of a support vector machine, a random forest model and a gradient boosting tree.

3. The method for classification and identification of evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network according to claim 1, characterized in that: The classification error weight of each level classifier is calculated according to the following formula: Where, w(y i ) is the node y in the hierarchical multi-label classification network i The classification error weight, For node y i At the level in the hierarchical multi-label classification network, maxlevel is the length of the longest path in the hierarchical multi-label classification network.

4. A system for classifying and identifying evergreen broad-leaved forest vegetation subtypes based on a hierarchical multi-label network using the method according to any one of claims 1 to 3, characterized in that: include: an acquisition unit configured to acquire multi-source remote sensing data of a sample area and extract classification features from the multi-source remote sensing data; The multi-source remote sensing data includes optical image data and radar image data; the classification features include spectral features, texture features and time series features; A sample unit is configured to construct a sample library of evergreen broad-leaved forests and their subtypes in the sample area based on field survey data and an existing vegetation classification system; A modeling unit is configured to construct a hierarchical multi-label classification network by sequentially setting multiple levels of classifiers based on the sample library and the classification features; the classification result of each level of classifier is a subdivision of the classification result of the upper level classifier, and the classification error weight of each level of classifier decreases sequentially along the hierarchy; The identification unit is configured to classify and identify the evergreen broad-leaved forest vegetation in the target area through the hierarchical multi-label classification network.

5. The evergreen broad-leaved forest vegetation subtype classification and identification system based on hierarchical multi-label network according to claim 4 is characterized in that: The modeling unit is further configured to: assigning multi-level classification labels to the samples according to the vegetation types of the samples in the sample library; According to the hierarchical relationship of the multi-level classification labels, multiple classifiers are trained from the top level to the bottom level to form an initial hierarchical network; each classifier subdivides and classifies one of the classification results of the previous level classifier; The classification results of non-bottom-level classifiers are trimmed to eliminate the classification results that do not require the next-level classifier to perform subdivision classification; The classification result of the bottom-level classifier is used as the classification result output by the hierarchical multi-label classification network.

6. The evergreen broad-leaved forest vegetation subtype classification and identification system based on hierarchical multi-label network according to claim 4 is characterized in that: The classification error weight of each level classifier is calculated according to the following formula: Where, w(y i ) is the node y in the hierarchical multi-label classification network i The classification error weight, For node y i At the level in the hierarchical multi-label classification network, maxlevel is the length of the longest path in the hierarchical multi-label classification network.

Images