Geological disaster post-vegetation recovery monitoring method based on high-resolution remote sensing image

Abstract

The application provides a kind of geological disaster post-vegetation recovery monitoring method based on high-resolution remote sensing image, comprising: constructing fusion type "spectrum-vegetation index-texture" feature set;In sunny weather conditions, more evenly select n typical vegetation samples in the target area;Construct bagging type ensemble learning model;Analysis of the obvious differences of arbor, shrub and herbaceous vegetation in vegetation height, single plant vegetation horizontal projection coverage area, root depth and root extension range;By analyzing the vegetation change rate index and vegetation type change rate index of several continuous time phase remote sensing images according to the time phase change information, the vegetation recovery of the target area is quantitatively analyzed, so as to effectively improve the accuracy of pixel classification method, greatly improve the accuracy of quantitative analysis of vegetation in the target area, and greatly reduce the cost of long-term accurate monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vegetation recovery monitoring, and particularly relates to a method for monitoring vegetation recovery after geological disasters based on high-resolution remote sensing images. BACKGROUND

[0002] Geological disasters occur frequently in China. Geological disasters usually cause serious damage to the ecological environment, and the surface vegetation, as a core component of the ecological environment, is the first to suffer very serious damage. As an important indicator of the ecological environment, monitoring and evaluating the regional vegetation recovery is one of the key tasks of post-disaster reconstruction. Traditional vegetation resource investigation mainly relies on manual ground collection, which has the disadvantages of heavy workload, low efficiency, limited scope, long cycle, and the like, and is difficult to meet the needs of efficient and long-term precision vegetation monitoring. Remote sensing technology is widely used in vegetation resource monitoring due to its macroscopic nature, dynamic nature, repeated access, and the like. With the improvement of the resolution of remote sensing images, high-resolution images are increasingly applied to vegetation resource monitoring due to their rich ground feature information and pure spectral characteristics.

[0003] Traditional pixel-based classification methods are widely used in vegetation resource monitoring, but they are mainly applied to medium and low spatial resolution multispectral images, and have many problems for high spatial resolution multispectral images, such as low classification accuracy and obvious salt and pepper noise. SUMMARY

[0004] The present application provides a method for monitoring vegetation recovery after geological disasters based on high-resolution remote sensing images to solve at least one of the above technical problems.

[0005] To solve the above problems, as one aspect of the present application, a method for monitoring vegetation recovery after geological disasters based on high-resolution remote sensing images is provided, comprising:

[0006] Step 1: Preprocessing the multispectral satellite remote sensing image with a spatial resolution of ≤10 meters in the target area;

[0007] Step 2: Constructing a fusion-type "spectrum-vegetation index-texture" feature set so that each ground feature pixel point corresponds to a feature vector of no less than 50 dimensions, which is used as the input quantity of the model in step 4 below;

[0008] Step 3: Under clear weather conditions, uniformly selecting n typical vegetation sample lands in the target area, and after field investigation, establishing a geographic survey data set containing accurate geographic coordinate values, and extracting f remote sensing images closest in imaging time to the survey time as training and testing data in step 4 below;

[0009] Step 4: Constructing a bagging-type ensemble learning model;

[0010] Step 5, using the learning model to analyze the obvious differences in vegetation height, single-plant horizontal projection coverage area, root depth and root extension range among the three types of vegetation, i.e., trees, shrubs and herbs;

[0011] Step 6, quantitatively analyzing the vegetation restoration in the target region by analyzing the per-time-phase change information of the vegetation change rate index and the vegetation type change rate index of the remote sensing images of several continuous time phases.

[0012] Preferably, in step 1, the multi-spectral satellite remote sensing images include a plurality of multi-spectral images acquired by a plurality of remote sensing satellites, and the preprocessing includes orthorectification, geometric fine correction, radiation calibration and atmospheric correction, etc. The remote sensing images after preprocessing include atmospheric bottom reflectance data, which can meet the demand for accurate analysis of vegetation characteristics.

[0013] Preferably, in the preprocessing process, the remote sensing images can be resampled to uniformly resample the spatial resolutions of the target bands to the same scale, and preferably, the nearest neighbor distance algorithm is used for resampling.

[0014] Preferably, in step 2, the feature set includes a spectral feature subset, a vegetation index feature subset and a texture feature subset, wherein the spectral feature subset is constructed by extracting bands sensitive to ground vegetation in the remote sensing images, the vegetation index feature subset is constructed by selecting a plurality of indices related to vegetation characteristics, and the texture feature subset is constructed by using the statistical quantities-energy, entropy, uniformity, difference entropy, etc. calculated by the gray level co-occurrence matrix of the vegetation reflection significant bands in the spectral feature subset to construct the texture feature subset part at the core pixel point.

[0015] Preferably, the spectral feature subset is preferably constructed by using blue light, green light, red light, vegetation red edge and near-infrared bands to construct the spectral feature subset part in the feature set; and / or, the vegetation index feature subset is preferably constructed by using NDVI, DVI, EVI, PVI, CTVI, TSAVI and RVI to construct the vegetation index feature subset part in the feature set; and / or, the texture feature subset is preferably analyzed in three directions of 0°, 45° and 90°.

[0016] Preferably, in step 3, the number n of sample plots is not less than twice the dimension number of the feature vector in step 2, and the remote sensing images are preferably 5 images before and after the survey time as the midpoint, and n K*K pixel size ROIs are picked up in each remote sensing image according to the accurate geographic coordinate value, a total of f*n ROIs are acquired, and a total of f*n*K*K feature vector sample data of labeled pixels are contained.

[0017] Preferably, in step 4, the learning model comprises several neural network sub-classifiers and several SVM sub-classifiers and an output unit of a neighborhood type combination strategy; wherein,

[0018] The neural network classifiers differ from each other in structure, have different numbers of hidden layers and neural nodes, and are independently trained using different training sets and test sets, so that the overall classification accuracy and Kappa coefficient of each neural network classifier are better than 85% and 0.85 after training;

[0019] The SVM classifiers differ from each other in structure, are SVM classifiers based on different kernel functions or LSSVM classifiers based on different kernel functions, are independently trained using different training sets and test sets, so that the overall classification accuracy and Kappa coefficient of each SVM classifier are better than 80% and 0.85 after training, the sub-classifiers work independently in parallel, and the classification results thereof are input into an output unit of a neighborhood type combination strategy;

[0020] The domain type combination strategy is: if the output of the sub-classifier is not more than two types of ground objects and the proportion of the dominant ground object class is not less than a threshold, the dominant ground object class is the ground object class at the point; otherwise, it is considered that the mixed pixel phenomenon at the pixel point is obvious, a domain of K*K pixels is obtained, and the content of a certain class of ground object in the core pixel point is obtained by multiplying the content of a certain class of ground object in the domain with the content of the certain class of ground object in the output of the multiple classifiers at the core pixel point.

[0021] Preferably, step 5 comprises:

[0022] (1) setting a corresponding vegetation influence factor for each type of vegetation, and determining the vegetation condition index by the following formula:

[0023] VCI = α a *C arbor + α s *C shrub + α h *C herbal

[0024] Wherein: α a , α s , α h are the vegetation influence factors of arbor, shrub and herb types, respectively, and the values of the vegetation influence factors α a , α s , α h are set according to the influence of vegetation on the environment, and generally can be set to 10, 5 and 2, respectively; C arbor , C shrub , C herbalVegetation coverage of tree, shrub and herb types in the study area respectively;

[0025] (2) Quantify vegetation change by vegetation change rate index:

[0026] VCR = VCI new / VCI stadndard

[0027] Wherein: VCI new is the vegetation condition index of a remote sensing image after the earthquake; VCI standard is the vegetation condition index of a remote sensing image before the geological disaster; VCR is the vegetation change rate index, which is less than 1, indicating that the vegetation in the study area is damaged, and which is greater than 1, indicating that the vegetation in the study area is more lush than before;

[0028] (3) Quantify the change of a certain type of vegetation by vegetation type change rate index:

[0029] SCR = C Species_new / C Species_standard

[0030] Wherein: C Species_new is the vegetation coverage of a certain type of vegetation in the study area in a remote sensing image after the geological disaster; C Species_standard is the vegetation coverage of a certain type of vegetation in the study area in a remote sensing image before the geological disaster; SCR is the vegetation type change rate index, which is less than 1, indicating that the vegetation of a certain type is damaged, and which is greater than 1, indicating that the vegetation of a certain type is more lush than before.

[0031] Due to the adoption of the above technical scheme, the present application has the following innovations:

[0032] (1) The present application organically integrates the spectral information of the core pixel, the nonlinear correlation information between multiple bands and the texture information within a certain field range of the core pixel by constructing a "spectrum-vegetation index-texture" feature set, and can effectively improve the accuracy of pixel classification method combined with the subsequent bagging type integrated learning model.

[0033] (2) The present application designs a model containing a plurality of neural network sub-classifiers, a plurality of SVM sub-classifiers and an output unit of a neighborhood type combination strategy, which realizes accurate single-pixel proportional feature classification, greatly improves the accuracy of vegetation quantification analysis in the target area.

[0034] (3) The present application fully utilizes the macroscopic, dynamic and repeated access characteristics of remote sensing technology to reduce the cost of environmental investigation, and is particularly suitable for free and open source high spatial resolution remote sensing image application scenarios, which can greatly reduce the cost of long-term accurate monitoring. DETAILED DESCRIPTION

[0035] The following detailed description of the embodiments of the present application is provided, but the present application can be implemented in various different ways limited and covered by the claims.

[0036] The present application provides a geological disaster post-vegetation recovery monitoring method based on high-resolution remote sensing images, which introduces machine learning technology into fine analysis of remote sensing images, systematically establishes a vegetation recovery monitoring model according to the principle of layer-by-layer interpretation, and applies an improved integrated learning method based on a fusion feature set to qualitatively and quantitatively analyze the vegetation recovery condition, thereby providing technical support for planning and management of post-disaster reconstruction.

[0037] The following detailed description of the implementation process of the present application is provided.

[0038] (1) The multispectral satellite remote sensing image of the target area with a spatial resolution of less than or equal to 10 meters is preprocessed. The multispectral satellite remote sensing image contains multispectral images obtained by several remote sensing satellites. The preprocessing mainly involves orthorectification, geometric precise correction, radiation calibration, and atmospheric correction, etc. The remote sensing image after preprocessing contains atmospheric bottom reflectance data, which can meet the demand of accurate analysis of vegetation characteristics. Optionally, in order to improve the accuracy of subsequent analysis, the remote sensing image can be resampled during the preprocessing process, and the spatial resolution of each target band is uniformly resampled to the same scale, and preferably the nearest neighbor distance algorithm is used for resampling.

[0039] (2) Construct a fusion feature set, which contains three main feature subsets: spectral feature subset, vegetation index feature subset and texture feature subset. Construction of spectral feature subset: extract the bands sensitive to the ground vegetation in the remote sensing image, preferably the blue light, green light, red light, vegetation red edge and near-infrared band to construct the spectral feature subset part in the feature set. Construction of vegetation index feature subset: select several vegetation feature related indexes, preferably use NDVI (Normalized Difference Vegetation Index, Normalized Difference Vegetation Index), DVI (Difference Vegetation Index, Difference Vegetation Index), EVI (enhanced vegetation index), PVI (Perpendicular Vegetation Index), CTVI (Corrected Transformed Vegetation Index), TSAVI (Transformed Soil Adjusted Vegetation Index) and RVI (Ratio vegetation index) to construct the vegetation index feature subset part in the feature set. Construction of texture feature subset: the statistical quantities-energy, entropy, uniformity, difference entropy, etc. calculated by the gray level co-occurrence matrix of the vegetation reflection significant bands in the spectral feature subset constitute the texture feature subset part at the core pixel point, preferably, the analysis direction is 0°, 45°, 90°. Through the above method, the "spectrum-vegetation index-texture" feature set is constructed, so that each ground feature pixel point corresponds to a feature vector not less than 50 dimensions, which is used as the input quantity of the model in step (4).

[0040] (3) Under sunny weather conditions, n typical vegetation sample plots are selected in the target area, generally, the number of sample plots n is not less than twice the dimension number of the feature vector in step 2. After field investigation, a geographic survey data set containing accurate geographic coordinate values is established. Extract f remote sensing images closest to the survey time, preferably 5 images before and after the survey time as the center; according to the accurate geographic coordinate values, n K*K pixel size ROIs (regions of interest) are picked up in each remote sensing image, a total of f*n ROIs are obtained, which contain f*n*K*K labeled pixel feature vector sample data, which are used as training and test data in step (4).

[0041] (4) Construct a bagging type ensemble learning model, which contains several neural network sub-classifiers and several SVM sub-classifiers and an output unit of neighborhood type combination strategy. The neural network classifiers differ from each other in structure, and the number of hidden layers and the number of neurons are different, and each independently uses different training sets and test sets for model training, so that the overall classification accuracy (Overall Accuracy) and Kappa coefficient of each neural network classifier are better than 85% and 0.85. The SVM classifiers differ from each other in structure, which are different kernel function based SVM classifiers or different kernel function based LSSVM classifiers, each independently uses different training sets and test sets for model training, and after training, the overall classification accuracy and Kappa coefficient of each SVM classifier are better than 80% and 0.85. The sub-classifiers work independently in parallel, and input their classification results into an output unit of neighborhood type combination strategy. The field type combination strategy is: if the output of the sub-classifier is not more than two types of ground objects and the proportion of the dominant ground object class is not less than the threshold value (generally 85%), the dominant ground object class is the ground object class at the point (considering that the mixed pixel phenomenon at the pixel point is not obvious); Otherwise, it is considered that the mixed pixel phenomenon at the pixel point is obvious, and the K*K pixel field is taken (generally, k=3), and the content of a certain class of ground object in the core pixel point is obtained by multiplying the content of a certain class of ground object in the core pixel point. The content of a certain class in the core pixel point.

[0042] (5) The three types of vegetation, trees, shrubs and herbs, have obvious differences in vegetation height, single plant vegetation horizontal projection coverage area, root depth and root extension range. The significant degree of their influence on the environment is different, and it is necessary to set corresponding vegetation influence factors for each type of vegetation. The present application designs the following formula to determine the vegetation condition index (VCI, Vegetaion Condition Index):

[0043] VCI = α a *C arbor + α s *C shrub + α h *C herbal

[0044] Wherein: α a , α s , α h are the vegetation influence factors of tree, shrub and herb types respectively, and the vegetation influence factors α a , α s , α hThe values of C, C and C can be set to 10, 5 and 2 respectively arbor , C shrub , C herbal are the vegetation cover degrees (calculated according to the pixel ratio of remote sensing images) of the tree, shrub and herb types of vegetation in the study area. In order to quantify the vegetation change, the vegetation change rate (VCR, Vegetaion Change Rate) index is introduced:

[0045] VCR = VCI new / VCI standard

[0046] VCI new is the vegetation condition index of a remote sensing image after the occurrence of the geological disaster; VCI standard is the vegetation condition index of a remote sensing image before the occurrence of the geological disaster; VCR is the vegetation change rate index, and when the value is less than 1, it indicates that the vegetation in the study area is damaged, and when the value is greater than 1, it indicates that the vegetation in the study area is more flourishing than before. The vegetation change in the study area can be analyzed through the vegetation change rate index. In order to further quantify the change of a certain type of vegetation, the vegetation type change rate (SCR, Species Change Rate) index is designed:

[0047] SCR = C Species_new / C Species_andard

[0048] C Species_new is the vegetation cover degree of a certain type of vegetation in the study area in a remote sensing image after the occurrence of the geological disaster; C Species_standard is the vegetation cover degree of a certain type of vegetation in the study area in a remote sensing image before the occurrence of the geological disaster; SCR is the vegetation type change rate index, and when the value is less than 1, it indicates that the vegetation of the certain type is damaged, and when the value is greater than 1, it indicates that the vegetation of the certain type is more flourishing than before. The change of the certain type of vegetation can be analyzed through the vegetation type change rate index.

[0049] (6) By analyzing the vegetation change rate index and the vegetation type change rate index of a plurality of continuous time phase remote sensing images according to the time phase change information, the vegetation recovery of the target area can be quantitatively analyzed.

[0050] In the following, a preferred embodiment of the present application is further described.

[0051] European Space Agency only released Sentinel-2 L1C level multi-spectral data, Sentinel-2 L1C level product is the atmospheric apparent reflectance product after orthorectification and geometric correction, the spatial resolution of this type of product can meet the requirements of accurate analysis of spatial distribution of ground objects; but this type of product does not carry out radiation calibration and atmospheric correction, which cannot meet the demand of accurate analysis of vegetation, so it is necessary to carry out accurate radiation calibration and atmospheric correction on this product. We use Sen2cor software to preprocess L1C level data to obtain L2A level atmospheric bottom reflectance data (Bottom-of-Atmosphere corrected reflectance) after accurate radiation calibration and atmospheric correction, which can meet the demand of accurate analysis of vegetation characteristics. In order to improve the accuracy of subsequent analysis, the obtained L2A level remote sensing data is resampled according to the nearest neighbor distance algorithm, and the spatial resolution of each interested band is resampled to 10m to obtain 10m resolution remote sensing image. Then, using the method commonly used in ENVI software, the remote sensing image is cut according to the administrative boundary of Jiuzhaigou scenic area, and the data outside the target research area is removed to improve the efficiency of data processing and analysis, and the final research area remote sensing image is obtained.

[0052] According to the principle of step-by-step interpretation, first of all, the feature set for ground object classification is constructed, according to the NDVI characteristics of different target ground objects, the ground objects are initially classified into three categories: vegetation, water and bare soil, and then the vegetation is accurately subdivided into three subcategories: trees, shrubs and herbs using machine learning algorithm, and then subsequent research is carried out.

[0053] The feature set includes three main feature subsets: spectral feature subset, vegetation index feature subset and texture feature subset. The construction of the feature set: (1) extract B2 (blue light, 490 nm), B3 (green light, 560 nm), B4 (red light, 665 nm), B5 (vegetation red edge 1, 705 nm), B6 (vegetation red edge 2, 740 nm), B7 (vegetation red edge 3, 705 nm), B8 (near infrared, 842 nm), B8A (vegetation red edge 4, 842 nm), B11 (SWIR1, 705 nm) and B12 (SWIR2, 842 nm) in the final study area remote sensing image to construct the spectral feature subset part in the feature set; (2) use NDVI (Normalized Difference Vegetation Index, normalized vegetation index), DVI (Difference Vegetation Index, difference vegetation index), EVI (enhanced vegetation index), PVI (Perpendicular Vegetation Index, perpendicular vegetation index), CTVI (Corrected Transformed Vegetation Index, corrected vegetation index), TSAVI (Transformed Soil Adjusted Vegetation Index, transformed soil adjusted vegetation index) and RVI (Ratio vegetation index) to construct the vegetation index feature subset part in the feature set; (3) use the statistical quantities calculated by the gray level co-occurrence matrix (the window size is 5X5; the direction is 0°, 45° and 90° respectively) in the B2, B3, B4 and B6 bands in the spectral feature subset to calculate the parameters such as energy, entropy, uniformity and difference entropy to construct the texture feature subset part at the core pixel point. By constructing the "spectrum-vegetation index-texture" feature set, each ground feature pixel point corresponds to a 54-dimensional feature vector.

[0054] The vegetation mask of the remote sensing image is made by using the NDVI coefficient value as a rule (NDVI>0.035 is regarded as a vegetation type ground feature), the vegetation area of the remote sensing image is extracted, and the vegetation classification is focused.

[0055] In mid-July (the vegetation in the study area is in full growth, and the meteorological conditions are good at this time, which is beneficial to obtain effective remote sensing data without cloud cover), typical vegetation sample sites (50 for trees, 50 for shrubs, and 50 for herbs) were selected uniformly in the core study area to establish a geographic survey dataset containing precise geographic coordinate values. Ten remote sensing images closest in time to the survey time (with the survey time as the midpoint, 5 images before and after) were extracted. According to the precise geographic coordinate values, 150 5x5 pixel size ROIs (regions of interest) were picked up in each remote sensing image, a total of 1500 ROIs containing 27500 labeled pixel point sample data were obtained. After obtaining the labeled pixel point sample data, each time the training set and test set were made: 70% of the labeled sample data was extracted as the training set according to the random sampling method, and the remaining 30% of the labeled sample data was used as the test set.

[0056] A bagging type ensemble learning model was used for fine classification of vegetation, which included three neural network classifiers and two SVM classifiers, and an output unit with a domain type combination strategy. The three neural network classifiers differ in structure, with different numbers of hidden layers and neurons, and each uses different training and test sets for model training independently; after training, the overall classification accuracy (Overall Accuracy) and Kappa coefficient of each neural network classifier are better than 98% and 0.92. The two SVM classifiers are an RFB-based SVM classifier and an LSSVM classifier based on a Laplace kernel function, each using different training and test sets for model training independently; after training, the overall classification accuracy (Overall Accuracy) and Kappa coefficient of each SVM classifier are better than 97% and 0.92. The five classifiers work independently in parallel: if their output results are not more than two classes and the proportion of the dominant subclass is not less than 4 / 5, the dominant subclass is the vegetation subtype at that point (considering that the mixed pixel phenomenon at that pixel point is not obvious); if it is not the above case, it is considered that the mixed pixel phenomenon at that pixel point is obvious, and a 3x3 pixel domain is taken, and the content of a certain class in the pixel is obtained by multiplying the content of a certain class in the domain by the content of a certain class in the output of multiple classifiers. Through this type of ensemble learning model, fine classification of vegetation in each study area remote sensing image can be achieved.

[0057] The three types of vegetation, i.e. tree, shrub and herb, have obvious differences in vegetation height, horizontal projection coverage of single plant, root depth and root extension range. The analysis of the significant degree of their influence on the environment shows that the corresponding vegetation influence factor should be set for each type of vegetation, and the vegetation condition index (VCI) is determined according to the following formula:

[0058] VCI = a a * C arbor + a s * C shrub + a h * C herbal

[0059] wherein a a , a s , a h are the vegetation influence factors of tree, shrub and herb types respectively, and the values of the vegetation influence factors a a , a s , a h are 10, 5 and 2 respectively according to the relevant literature on the influence of vegetation on the environment and the dominant vegetation types in Jiuzhai Scenic Area; C arbor , C shrub , C herbal are the vegetation coverages of tree, shrub and herb types in the study area (calculated according to the pixel ratio of remote sensing image). In order to quantify the vegetation change, the vegetation change rate (VCR) index is introduced:

[0060] VCR = VCI new / VCI standard

[0061] wherein VCI new is the vegetation condition index of a remote sensing image after the earthquake; VCI standard is the vegetation condition index of a remote sensing image before the earthquake (the day of the earthquake: 11:53, the Sentinel-2 satellite passed over the Jiuzhai Valley Scenic Area and obtained a cloud-free remote sensing image with good imaging conditions; 21:19, the earthquake occurred); VCR is the vegetation change rate index, and its value less than 1 indicates that the vegetation in the study area is damaged, and its value greater than 1 indicates that the vegetation in the study area is more lush than before. The vegetation change rate index can be used to analyze the vegetation change in the study area. In order to further quantify the change of a certain type of vegetation, the species change rate (SCR) index is introduced:

[0062] SCR = C Species_new / CSpecies_standard

[0063] wherein: C Species_new is the vegetation coverage of a certain type of vegetation in the study area in a remote sensing image of a certain scene after the earthquake; C SPecies_standard is the vegetation coverage of a certain type of vegetation in the study area in a remote sensing image of the day before the earthquake; SCR is the vegetation type change rate index, and when the value is less than 1, it indicates that the vegetation of a certain type is damaged, and when the value is greater than 1, it indicates that the vegetation of a certain type is more lush than before. The change of a certain type of vegetation can be analyzed through the vegetation change rate index.

[0064] Due to the adoption of the above technical scheme, the present application has the following innovations:

[0065] (1) Overcome the low accuracy of the traditional pixel classification method.

[0066] In high spatial resolution remote sensing images, the same object with different spectra and different objects with the same spectrum are particularly prominent, which leads to the low classification accuracy of the traditional pixel spectrum information-based ground object classification method in this scene, and cannot meet the requirements of accurate quantitative analysis. The present application organically integrates the spectral information of the core pixel, the nonlinear correlation information between multiple bands and the texture information within a certain field range of the core pixel by constructing a "spectrum-vegetation index-texture" feature set, and combines the subsequent bagging-type integrated learning model to effectively improve the accuracy of the pixel classification method.

[0067] (2) Overcome the difficulty of quantitative classification of the traditional pixel classification method

[0068] The traditional remote sensing classifier can only give an exclusive classification result, i.e. a certain pixel can only belong to a certain type of ground object. However, the actual ground object distribution leads to the widespread existence of mixed pixels. According to the exclusive classification result, the vegetation analysis of the target area often leads to large errors. The present application designs a model containing several neural network sub-classifiers, several SVM sub-classifiers and an output unit of a neighborhood type combination strategy, which realizes accurate single-pixel proportional ground object classification. The accuracy of the quantitative analysis of the vegetation of the target area is greatly improved.

[0069] (3) Greatly reduce the cost of monitoring and evaluating the vegetation restoration of the target area

[0070] Traditional vegetation resource investigation mainly adopts manual ground collection, which has the disadvantages of heavy workload, low efficiency, limited range, long cycle and the like, and is difficult to meet the needs of efficient long-time-span accurate vegetation monitoring. The present application makes full use of the macroscopic, dynamic and repeated access characteristics of remote sensing technology to reduce the cost of environmental investigation, and is particularly suitable for free open source high spatial resolution remote sensing image application scenarios, which can greatly reduce the cost of long-term accurate monitoring.

[0071] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Various modifications and changes can be made by those skilled in the art based on the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for monitoring vegetation restoration after geological disasters based on high-resolution remote sensing images, characterized in that, include: Step 1: Preprocess the multispectral satellite remote sensing images of the target area with a spatial resolution ≤10 meters; Step 2: Construct a fusion-type "spectral-vegetation index-texture" feature set so that each ground feature pixel corresponds to a feature vector of no less than 50 dimensions, which will be used as the input to the model in Step 4 below. Step 3: Under clear weather conditions, select n typical vegetation sample plots in the target area relatively evenly. After conducting on-site surveys, establish a geographic survey dataset containing accurate geographic coordinates. Extract f remote sensing images whose imaging time is closest to the survey time as training and testing data in Step 4 below. Step 4: Construct a bagging-like ensemble learning model; The learning model includes several neural network sub-classifiers, several SVM sub-classifiers, and an output unit with a neighborhood-based combination strategy; among them, Neural network classifiers differ in structure, with different numbers of hidden layers and neurons, and each is trained independently using different training and test sets. After training, the overall classification accuracy and Kappa coefficient of each neural network classifier are better than 85% and 0.85, respectively. SVM classifiers differ structurally from one another, being either SVM classifiers based on different kernel functions or LSSVM classifiers based on different kernel functions. Each classifier is trained independently using different training and test sets. After training, each SVM classifier achieves an overall classification accuracy and Kappa coefficient better than 80% and 0.85, respectively. The subclassifiers work in parallel and independently, and their classification results are input into the output unit of a neighborhood-based fusion strategy. The domain-based combination strategy is as follows: if the output of the subclassifier has no more than two types of land cover and the proportion of the dominant land cover type is not less than the threshold, the dominant land cover type is the land cover type at that point; otherwise, it is considered that the mixed pixel phenomenon at that pixel point is obvious, and a domain of K*K pixels is obtained. The content of a certain type of land cover in the pixel point is obtained by multiplying the content of a certain type of land cover in the de-centered domain by the content of a certain type of land cover in the output of multiple classifiers at that pixel point. Step 5: Use the learning model to analyze the significant differences among the three types of vegetation—trees, shrubs, and herbs—in terms of vegetation height, horizontal projection coverage of a single plant, root depth, and root extension range. Step 6: Quantitatively analyze the vegetation restoration status of the target area by analyzing the temporal changes of the vegetation change rate index and the vegetation type change rate index of several consecutive time-phase remote sensing images.

2. The method for monitoring vegetation restoration after geological disasters based on high-resolution remote sensing images according to claim 1, characterized in that, In step 1, the multispectral satellite remote sensing image contains multispectral images acquired by several remote sensing satellites. The preprocessing includes orthorectification, geometric fine correction, radiometric calibration, and atmospheric correction. The preprocessed remote sensing image contains atmospheric bottom reflectance data, which can meet the needs of accurate analysis of vegetation characteristics.

3. The method for monitoring vegetation restoration after geological disasters based on high-resolution remote sensing images according to claim 2, characterized in that, During the preprocessing process, the remote sensing images are resampled to unify the spatial resolution of each target band to the same scale, and the nearest neighbor distance algorithm is used for resampling.

4. The method for monitoring vegetation restoration after geological disasters based on high-resolution remote sensing images according to claim 1, characterized in that, In step 2, the feature set includes a spectral feature subset, a vegetation index feature subset, and a texture feature subset. The spectral feature subset is constructed by extracting bands in the remote sensing image that are sensitive to surface vegetation. The vegetation index feature subset is constructed by selecting several indices related to vegetation features. The texture feature subset is constructed by using the statistics—energy, entropy, uniformity, and differential entropy parameters—calculated from the gray-level co-occurrence matrix for the vegetation reflectance bands in the spectral feature subset to form the texture feature subset part at the core pixel.

5. The method for monitoring vegetation restoration after geological disasters based on high-resolution remote sensing images according to claim 4, characterized in that, The spectral feature subset is constructed from blue light, green light, red light, vegetation red edge, and near-infrared bands; and / or, the vegetation index feature subset is constructed from NDVI, DVI, EVI, PVI, CTVI, TSAVI, and RVI; and / or, the texture feature subset is preferentially analyzed in three directions: 0°, 45°, and 90°.

6. The method for monitoring vegetation restoration after geological disasters based on high-resolution remote sensing images according to claim 4, characterized in that, In step 3, the number of sample plots n is not less than twice the number of feature vector dimensions in step 2. Five remote sensing images are taken before and after the survey time. Based on the accurate geographic coordinates, n K*K pixel-sized ROIs are picked in each remote sensing image, resulting in a total of f*n ROIs containing feature vector sample data of f*n*K*K labeled pixels.

7. The method for monitoring vegetation restoration after geological disasters based on high-resolution remote sensing images according to claim 1, characterized in that, Step 5 includes: (1) Set corresponding vegetation impact factors for each type of vegetation, and determine the vegetation status index using the following formula: , where: α a α s α h These are vegetation impact factors for tree, shrub, and herbaceous types, respectively. The vegetation impact factor α is then rationally set based on the environmental impact of vegetation. a α s α h The values ​​for C are set to 10, 5, and 2 respectively; arbor C shrub C herbal The vegetation cover of tree, shrub, and herbaceous vegetation types in the study area are respectively; (2) Quantify vegetation change using the vegetation change rate index: ,in: VCI new The vegetation condition index of a remote sensing image of a scene after an earthquake; VCI standard The vegetation status index is the vegetation condition index of the remote sensing image before the geological disaster occurs; VCR is the vegetation change rate index, which indicates that the vegetation in the study area is damaged when the value is less than 1, and that the vegetation in the study area is more lush than before when the value is greater than 1. (3) Quantify the change of a certain type of vegetation through the vegetation type change rate index: ,in: C Species_new The vegetation coverage of a certain type of vegetation in the study area is shown in a remote sensing image of a scene after a geological disaster. C Species_standard The vegetation cover of a certain type of vegetation in the study area is represented by a remote sensing image before the geological disaster. The SCR is the vegetation type change rate index, which indicates that a certain type of vegetation has been damaged when its value is less than 1, and that a certain type of vegetation is more lush than before when its value is greater than 1.