Urban green space carbon storage assessment method based on high-resolution imaging and machine learning
Through a method based on high-resolution imaging and machine learning, combined with Sentinel-2 remote sensing data, identify urban green space and calculate carbon storage, the problems of low evaluation accuracy and difficulty in dynamic monitoring in the existing technology are solved, and the carbon storage of urban green space is achieved , the monitoring of dynamic timing changes and high-precision evaluation of the carbon sink efficiency of urban ecological space.
Patent Information
- Application Number
- CN202410420296.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-09
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-04-09
AI Technical Summary
The existing technology is difficult to quickly, efficiently and accurately evaluate the carbon storage of urban green spaces and realize large-scale dynamic monitoring, making it difficult to accurately measure the carbon sink efficiency of urban ecological space.
Using a method based on high-resolution imaging and machine learning, a random forest classifier and regression model is combined with Sentinel-2 remote sensing data to identify urban green space and calculate carbon reserves to achieve monitoring and high-precision evaluation of timing dynamic changes.
It has realized the monitoring of time-series dynamic changes in urban green space carbon storage and high-precision assessment, with good timeliness and large-scale monitoring capabilities, and scientifically evaluated the carbon sink efficiency of urban ecological space.
Smart Images

Figure CN118378937B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of urban green space carbon reserve assessment, and in particular to an urban green space carbon reserve assessment method based on high-resolution imaging and machine learning. Background Art
[0002] As the only direct natural carbon sink space within the city, urban green space plays an important role in increasing carbon sinks and reducing emissions. How to accurately assess the carbon storage of urban green space over a large area and clarify its dynamic changes, and effectively enhance the carbon sink function of urban green space has always been a hot topic of research at home and abroad.
[0003] At present, the methods for assessing carbon storage in urban green spaces can be roughly summarized into four categories: plot inventory method, micrometeorological method, remote sensing estimation method and model simulation method. The plot inventory method is based on field surveys to obtain the growth parameters of vegetation in the survey plot (such as diameter at breast height, tree height, etc.), and obtains the carbon storage of urban green spaces by calculating biomass and carbon content coefficient. This method is simple and convenient, but it often requires a lot of manpower, material and financial resources, and is not suitable for large-scale urban green space carbon storage assessment. The micrometeorological method measures the carbon flux in the atmospheric environment through carbon flux observation stations around the world, but it is severely restricted by terrain and meteorological conditions, and is less distributed in urban areas. The remote sensing estimation method refers to the assessment of urban green space carbon storage using medium and low resolution remote sensing images as data sources, but the spatial resolution is generally low, which is difficult to meet the needs of accurate measurement of urban green space carbon storage. The model simulation method is a mathematical method that uses computers to quantitatively describe the dynamics of ecosystem carbon cycle and its relationship with global change to measure urban green space carbon storage. It often involves complex parameters and ecological processes, and it is difficult to achieve large-scale dynamic monitoring.
[0004] In summary, there is still a lack of a method that can quickly, efficiently and accurately assess the carbon storage of urban green spaces and achieve large-scale dynamic monitoring, making it difficult to accurately measure the carbon sink efficiency of my country's urban ecological space. Summary of the invention
[0005] In view of the shortcomings of the existing technology, the present invention provides a method for evaluating carbon reserves in urban green spaces based on high-resolution images and machine learning, which solves the technical problems in the existing technology of lack of timeliness, low evaluation accuracy and difficulty in large-scale dynamic monitoring of carbon reserves in urban green spaces.
[0006] The technical solution adopted by the present invention to solve the above technical problems is: a method for evaluating urban green space carbon reserves based on high-resolution images and machine learning, which is characterized by comprising the following steps:
[0007] Step 1: obtain multiple original high-resolution remote sensing images covering the same field monitoring area; then perform image preprocessing on all the original high-resolution remote sensing images obtained to obtain a remote sensing image of the field monitoring area of interest; wherein, the remote sensing image retains 12 bands, namely, the coast / aerosol band, i.e., the B1 band, the blue light band, i.e., the B2 band, the green light band, i.e., the B3 band, the red light band, i.e., the B4 band, the red edge band 1, i.e., the B5 band, the red edge band 2, i.e., the B6 band, the red edge band 3, i.e., the B7 band, the wide near-infrared band, i.e., the B8 band, the narrow near-infrared band, i.e., the B8A band, the water vapor band, i.e., the B9 band, the short-edge infrared band 1, i.e., the B11 band, and the short-edge infrared band 2, i.e., the B12 band;
[0008] Step 2: In the remote sensing image obtained in step 1, multiple evenly distributed areas of interest are manually marked as samples for various land cover types on the remote sensing image, and the same number of samples are randomly selected from all samples of various land cover types as training samples, and the remaining samples are used as verification samples; then, based on the reflectivity of the narrow near-infrared band and the red light band of the remote sensing image, the normalized difference vegetation index image of the remote sensing image is calculated; then, based on the training samples, the remote sensing image and its normalized difference vegetation index image, a random forest classifier is used to accurately identify the urban green space in the remote sensing image; wherein the land cover types are urban green space, water body, bare land and impervious surface, and the verification samples are used to evaluate the recognition accuracy of the acquired urban green space;
[0009] Step 3: After identifying the urban green space in the remote sensing image, determine the actual urban green space; then use the random sampling method to randomly select multiple urban green space plots to be investigated from the actual urban green space; then, based on expert knowledge and visual interpretation, delete all the randomly selected urban green space plots with unstable growth status; then conduct field investigations and on-site surveys on all the retained urban green space plots, record the latitude and longitude of each retained urban green space plot, and clarify the actual geographical location of each urban green space plot;
[0010] Step 4: During the field survey and on-site investigation of each retained urban green space plot, conduct a tree-by-tree survey of all vegetation types of interest in each urban green space plot, and record the type and growth parameters of each plant; the vegetation types of interest include trees and shrubs, the growth parameters of trees include breast diameter, crown width and plant height, and the growth parameters of shrubs include base diameter, crown width and plant height;
[0011] Step 5: Calculate the total biomass of each plant in each urban green space sample plot based on the growth parameters of each plant in each urban green space sample plot; then multiply the total biomass of each plant in each urban green space sample plot by the corresponding carbon content coefficient to obtain the carbon storage of each plant in each urban green space sample plot; and then calculate the total carbon storage of all vegetation in each urban green space sample plot;
[0012] Step 6: According to the actual geographical location recorded in each urban green space sample plot, the corresponding coordinate position is matched in the remote sensing image; then, with the coordinate position of each urban green space sample plot matched in the remote sensing image as the center, the spatial offset method is used to perform spatial offset and expansion of the urban green space sample plot in the surrounding areas in the remote sensing image to obtain 9 times the new urban green space sample plots. The spatial resolution of each new urban green space sample plot is consistent with the spatial resolution of the remote sensing image; finally, the total carbon storage of all vegetation in the urban green space sample plot is evenly distributed according to the number of new urban green space sample plots corresponding to each urban green space sample plot, and the carbon storage of each new urban green space sample plot is obtained;
[0013] Step 7: extract 20 remote sensing parameters corresponding to each new urban green space plot in the remote sensing image, including 12 spectral information and 8 vegetation indices corresponding to the new urban green space plot; wherein, the 12 spectral information corresponding to the new urban green space plot is the spectral reflectance of the corresponding position in the 12 bands of the remote sensing image, and the 8 vegetation indices corresponding to the new urban green space plot are the values of the corresponding positions in 8 vegetation index images calculated based on the remote sensing image, and the 8 vegetation index images are normalized difference vegetation index image, normalized green difference vegetation index image, difference vegetation index image, enhanced vegetation index image, vegetation moisture index image, ratio vegetation index image, soil adjustment index image, and modified soil adjustment index image;
[0014] Step 8: Based on the carbon storage of each new urban green space sample in the remote sensing image and its corresponding 20 remote sensing parameters, the random forest regression model is used to preliminarily fit the nonlinear relationship between the carbon storage of the new urban green space sample and the 20 remote sensing parameters, and the preliminary urban green space carbon storage assessment model and the MDA parameters corresponding to each remote sensing parameter are obtained; then, the top five remote sensing parameters are selected as the optimal remote sensing parameters according to the descending order of the MDA parameters; then, based on the carbon storage of each new urban green space sample in the remote sensing image and its corresponding five optimal remote sensing parameters, the random forest regression model is used again to obtain the optimized urban green space carbon storage assessment model; finally, after the model robustness test, the final urban green space carbon storage assessment model is determined;
[0015] Step 9: The five images corresponding to the five optimal remote sensing parameters are synthesized into a remote sensing image of urban green space carbon storage to be predicted; the remote sensing image to be predicted is then input into the final urban green space carbon storage assessment model for simulation and prediction to obtain the carbon storage prediction result image of the remote sensing image to be predicted; based on the urban green space space in the remote sensing image identified in step 2, the carbon storage prediction result image of the remote sensing image to be predicted is masked to obtain the predicted carbon storage of each pixel in the urban green space space in the remote sensing image to be predicted.
[0016] In the step 1, the original high-resolution remote sensing image is obtained based on Sentinel-2 remote sensing data, and the level of the original high-resolution remote sensing image is L2A or L1C.
[0017] In the step 1, the process of image preprocessing for all the original high-resolution remote sensing images obtained is as follows: if the level of all the original high-resolution remote sensing images obtained is L2A, then all the bands of all the original high-resolution remote sensing images are resampled to 10 meters, and then all the 10-meter bands are mosaicked, and the mosaicked results are cropped according to the scope of the field monitoring area to obtain a remote sensing image containing the field monitoring area of interest; if there are images with levels of L2A and L1C among all the original high-resolution remote sensing images obtained, then the image with level L1C is firstly subjected to radiation calibration and atmospheric correction, and the level of the image obtained after the atmospheric correction is L2A, and then all the bands of all the images with level L2A are resampled to 10 meters, and then all the 10-meter bands are mosaicked, and the mosaicked results are cropped according to the scope of the field monitoring area to obtain a remote sensing image containing the field monitoring area of interest.
[0018] The specific process of step 2 is:
[0019] Step 2.1: Divide the field monitoring area in the remote sensing image into four different land cover types, namely urban green space, water body, bare land and impervious surface; then, based on the remote sensing image, use the manual annotation method to evenly select multiple areas of interest as samples for each land cover type on the remote sensing image, and the number of samples of the four land cover types is the same; finally, use the stratified sampling method to randomly select the same number of samples from all samples of each of the four land cover types as training samples, and use the remaining samples of each land cover type as verification samples;
[0020] Step 2.2: Calculate the normalized difference vegetation index image of the remote sensing image, and record the value of any pixel in the normalized difference vegetation index image of the remote sensing image as NDVI, NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red), where ρ NIR Represents the spectral reflectance of the corresponding position in the narrow near-infrared band of the remote sensing image, ρ Red Indicates the reflectance of the corresponding position in the red light band of the remote sensing image;
[0021] Step 2.3: Use the training samples, remote sensing images and their normalized difference vegetation index images to train a random forest classifier to obtain a trained random forest classifier; then input the remote sensing image into the trained random forest classifier to classify each land cover type in the field monitoring area in the remote sensing image to obtain a classified result image, in which the spatial distribution of the land cover types of all pixels in the field monitoring area in the remote sensing image is clearly defined, and the spatial distribution of urban green space is extracted, that is, the urban green space in the remote sensing image is identified;
[0022] Step 2.4: Based on the verification samples and the classified result images, the confusion matrix method is used to select four indicators to evaluate the recognition accuracy of the acquired urban green space. The four indicators are overall accuracy, kappa coefficient, user accuracy and mapping accuracy. If all four indicators meet the accuracy requirements, the identified urban green space will be used as the final recognition result. If some of the four indicators do not meet the accuracy requirements, adjust the ntree parameters in the random forest classifier and return to step 2.3 to continue training.
[0023] In the step 2.1, the number of samples of each land cover type is Y, of which 2 / 3 of the samples are used as training samples and the remaining 1 / 3 of the samples are used as verification samples, where Y>200; in the step 2.3, when the random forest classifier is used for training, the initial value of the random forest number is 500.
[0024] In step 3, the number of randomly selected urban green space plots is 200, and the size of the urban green space plot is set to 30 meters × 30 meters; the urban green space plot with unstable growth status refers to the urban green space plot that has had large-scale construction activities or obvious human interference in the past two to three years; the actual geographical location of the urban green space plot refers to the actual latitude and longitude coordinates marked during the field investigation and on-site survey.
[0025] In step 5, the process of calculating the total biomass of each plant in each urban green space plot is as follows:
[0026] Step 5.1: Based on the allometric growth equation of each plant, calculate the aboveground biomass of each plant in each urban green space plot, and record the aboveground biomass of any plant in any urban green space plot as W above , if the vegetation is a tree and its height is difficult to measure accurately, then W above =aD b, if the vegetation is a tree and its height is accurately measured or if the vegetation is a shrub, then W above =a(D 2 H) b ; Among them, W above The unit is kg. When the vegetation is a tree, D represents the diameter at breast height of the vegetation. When the vegetation is a shrub, D represents the base diameter of the vegetation. The unit of D is cm. H represents the plant height. The unit of H is m. a and b are both regression constants.
[0027] Step 5.2: Calculate the underground biomass of each plant in each urban green space sample plot, and record the underground biomass of any plant in any urban green space sample plot as W under , W above With W under The ratio is 4.28;
[0028] Step 5.3: Calculate the total biomass of each plant in each urban green space plot, and record the total biomass of any plant in any urban green space plot as W tottal If the vegetation is trees or unpruned shrubs, then W total =W above +W under , if the vegetation is shrubs that are often pruned, then W total =(W above +W under )×n%; where n% is the reduction ratio of the total biomass of the vegetation;
[0029] In step 5, the process of calculating the total carbon storage of all vegetation in each urban green space sample plot is as follows: the total carbon storage of all vegetation in any urban green space sample plot is recorded as C total , Where N represents the number of vegetation in any urban green space sample, C i Represents the carbon storage of the i-th plant in any urban green space plot.
[0030] In step 6, the size of the new urban green space plot is set to 10 meters × 10 meters. For any urban green space plot surveyed on the spot, 9 new urban green space plots are obtained after spatial offset and expansion in the remote sensing image.
[0031] In step 7, the process of extracting 20 remote sensing parameters corresponding to a new urban green space plot is as follows:
[0032] Step 7.1: Extract 12 types of spectral information corresponding to the new urban green space sample: Use the point value extraction method to obtain each type of spectral information corresponding to the new urban green space sample. Any type of spectral information is the spectral reflectance of the position corresponding to the new urban green space sample in the band corresponding to the remote sensing image;
[0033] Step 7.2: Extract the eight vegetation indices corresponding to the new sample plot of urban green space: record the value of any pixel in the normalized difference vegetation index image as NDVI, NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ), the value of any pixel in the normalized green difference vegetation index image is recorded as GNDVI, GNDVI = (ρ NIR -ρ Green ) / (ρ NIR +ρ Green ), the value of any pixel in the difference vegetation index image is recorded as DVI, DVI = ρ NIR -ρ Red , the value of any pixel in the enhanced vegetation index image is recorded as EVI, EVI = 2.5 × ((ρ NIR -ρ Red ) / (ρ NIR +6ρ Red -7.5ρ Blue +1)), the value of any pixel in the vegetation moisture index image is recorded as LSWI, LSWI = (ρ NIR -ρ SWIR1 ) / (ρ NIR +ρ SW1R1 ), the value of any pixel in the ratio vegetation index image is recorded as SR, SR = ρ NIR / ρ Red , the value of any pixel in the soil adjustment index image is recorded as SAVI, SAVI = ((ρ NIR -ρ Red )(1+0.5)) / (ρ NIR +ρ Red +0.5), and the value of any pixel in the modified soil conditioning index image is recorded as MSAVI, Then, the point value extraction method is used to obtain each vegetation index corresponding to the new urban green space sample plot. Any vegetation index is the value of the corresponding position of the new urban green space sample plot in the corresponding vegetation index image; where ρ NIR Represents the spectral reflectance of the corresponding position in the narrow near-infrared band of the remote sensing image, ρ Red Represents the spectral reflectance of the corresponding position in the red light band of the remote sensing image, ρ Green Represents the spectral reflectance of the corresponding position in the green light band of the remote sensing image, ρ Blue Represents the spectral reflectance of the corresponding position in the blue light band of the remote sensing image, ρ SWIR1 Indicates the spectral reflectance of the corresponding position in the short-side infrared band of the remote sensing image.
[0034] The specific process of step 8 is as follows:
[0035] Step 8.1: Using the carbon storage of each new urban green space sample in the remote sensing image and its corresponding 20 remote sensing parameters, the random forest regression model is trained to fit the nonlinear relationship between the carbon storage of all new urban green space samples and their corresponding 20 remote sensing parameters. This nonlinear relationship is defined as the preliminary assessment model of urban green space carbon storage. The MDA parameters corresponding to each remote sensing parameter are obtained during the training.
[0036] Step 8.2: According to the MDA parameters corresponding to the 20 remote sensing parameters, the 20 remote sensing parameters are sorted in descending order, and the remote sensing parameters corresponding to the top 5 MDA parameters are selected as the optimal remote sensing parameters;
[0037] Step 8.3: Using the carbon storage of each new urban green space sample in the remote sensing image and the five optimal remote sensing parameters selected, a random forest regression model is used for training to fit the nonlinear relationship between the carbon storage of all new urban green space samples and the five optimal remote sensing parameters. This nonlinear relationship is defined as the optimized urban green space carbon storage assessment model;
[0038] Step 8.4: Use the goodness of fit parameters of the random forest regression model in step 8.3 to detect the robustness of the optimized urban green space carbon stock assessment model. If the robustness is good, the optimized urban green space carbon stock assessment model is used as the final urban green space carbon stock assessment model. If the robustness is poor, first adjust the ntree parameter of the random forest regression model in step 8.3, then return to step 8.3 and re-test the robustness. If the robustness still does not meet the experimental requirements, adjust the ntree parameter of the random forest regression model in step 8.1 and re-execute. Perform step 8.1 until the final goodness of fit parameter of the urban green space carbon storage assessment model meets the experimental requirements; wherein, the initial value of the ntree parameter when the random forest regression model is used for training in step 8.1 is 800, and the initial value of the ntree parameter when the random forest regression model is used for training in step 8.3 is 800. When the goodness of fit parameter of the random forest regression model in step 8.3 is greater than or equal to 85%, it is considered that the robustness is good, and when the goodness of fit parameter of the random forest regression model in step 8.3 is less than 85%, it is considered that the robustness is poor.
[0039] Compared with the prior art, the advantages of the present invention are:
[0040] (1) The method of the present invention can monitor the temporal dynamic changes of carbon storage in urban green spaces and has good timeliness.
[0041] (2) The method of the present invention can identify the carbon storage of urban green spaces in highly urbanized areas with high assessment accuracy.
[0042] (3) The method of the present invention can evaluate the carbon storage of urban green spaces at the local to regional scales and has the characteristics of large-scale monitoring, thus being able to scientifically evaluate the carbon sink efficiency of urban ecological spaces. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is the overall implementation flow chart of the method of the present invention;
[0044] Figure 2 This is a schematic diagram of the spatial distribution of urban green spaces in the main built-up areas of Wuhan;
[0045] Figure 3 This is a schematic diagram of the geographical locations of 121 urban green space plots surveyed in the main built-up areas of Wuhan;
[0046] Figure 4 The results of a field survey of 121 urban green space plots in the main built-up areas of Wuhan;
[0047] Figure 5 Schematic diagram of the spatial shift, expansion and carbon stock redistribution process of an urban green space plot;
[0048] Figure 6 This is a schematic diagram of the urban green space carbon storage assessment model based on the random forest regression model;
[0049] Figure 7 The schematic diagram of the descending arrangement of the MDA parameters corresponding to the 20 remote sensing parameters;
[0050] Figure 8 This is a schematic diagram of the spatial assessment results of urban green space carbon storage in the main built-up areas of Wuhan in 2022. DETAILED DESCRIPTION
[0051] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.
[0052] The present invention integrates remote sensing technology, field investigation, machine learning and other model methods, combines Sentinel-2 remote sensing images with sample carbon storage survey data, and uses a random forest algorithm to construct a carbon storage assessment model to obtain an urban green space carbon storage assessment method based on high-resolution images and machine learning.
[0053] The present invention proposes a method for evaluating urban green space carbon storage based on high-resolution images and machine learning. The overall implementation flow chart is as follows: Figure 1 As shown, it includes the following steps:
[0054] Step 1: Obtain multiple original high-resolution remote sensing images covering the same field monitoring area (such as the main built-up area of Wuhan City); then perform image preprocessing on all the original high-resolution remote sensing images to obtain a remote sensing image of the field monitoring area of interest; wherein, the remote sensing image retains 12 bands, namely, the coast / aerosol band, namely the B1 band, the blue light band, namely the B2 band, the green light band, namely the B3 band, the red light band, namely the B4 band, the red edge band 1, namely the B5 band, the red edge band 2, namely the B6 band, the red edge band 3, namely the B7 band, the wide near-infrared band, namely the B8 band, the narrow near-infrared band, namely the B8A band, the water vapor band, namely the B9 band, the short-edge infrared band 1, namely the B11 band, and the short-edge infrared band 2, namely the B12 band.
[0055] Specifically, in step 1, the original high-resolution remote sensing image is obtained based on Sentinel-2 remote sensing data provided by the European Space Agency, and the level of the original high-resolution remote sensing image is L2A or L1C.
[0056] Currently, Sentinel-2 is the multispectral imager with the highest resolution among open-source image data. Compared with other open-source remote sensing data, Sentinel-2 remote sensing data has better spatial resolution (10 meters), richer spectral information (13 bands, see Table 1) and better radiation characteristics (12-bit radiometric resolution, see Table 1), and has been applied in many fields such as global change monitoring, ecology, agriculture, forestry, regional planning, monitoring and education. In addition, the precise red edge (RE) and near-infrared narrow (NIRn) spectral information and ultra-high temporal frequency (5-day recurrence period) in Sentinel-2MSI provide strong support for the spatial identification of urban green spaces and the precise monitoring of green space carbon reserves.
[0057] Table 1 Spectral band information of Sentinel-2 remote sensing data
[0058]
[0059] Specifically, in step 1, the process of image preprocessing for all the original high-resolution remote sensing images obtained is as follows: if the level of all the original high-resolution remote sensing images obtained is L2A, then all the bands of all the original high-resolution remote sensing images are resampled to 10 meters using the SNAP software, and then all the 10-meter bands are mosaicked in ENVI 5.3, and the mosaicked results are cropped according to the scope of the field monitoring area to obtain a remote sensing image containing the field monitoring area of interest, and the spatial resolution of the remote sensing image is 10 meters; if there are images of levels L2A and L1C among all the original high-resolution remote sensing images obtained, then the images of level L1C are first subjected to radiation calibration and atmospheric correction based on the Sen2Cor software provided by the European Space Agency. The level of the image obtained after atmospheric correction is L2A, and then all the bands of all the images of level L2A are resampled to 10 meters using the SNAP software, and then in ENVI In 5.3, all 10-meter bands are mosaicked, and the mosaicked results are cropped according to the scope of the field monitoring area to obtain a remote sensing image containing the field monitoring area of interest, and the spatial resolution of the remote sensing image is 10 meters. Here, radiation calibration and atmospheric correction, image mosaicking and image cropping are all existing technologies.
[0060] Step 2: In the remote sensing image obtained in step 1, multiple evenly distributed areas of interest are manually marked as samples for various land cover types on the remote sensing image, and the same number of samples are randomly selected from all samples of various land cover types as training samples, and the remaining samples are used as verification samples; then, based on the reflectivity of the narrow near-infrared band and the red light band of the remote sensing image, the normalized difference vegetation index image of the remote sensing image is calculated; then, based on the training samples, the remote sensing image and its normalized difference vegetation index image, a random forest classifier is used to accurately identify the urban green space in the remote sensing image; among them, the land cover types are urban green space, water body, bare land and impervious surface, and the verification samples are used to evaluate the recognition accuracy of the acquired urban green space. Figure 2 The spatial distribution of urban green spaces in the main built-up areas of Wuhan City is given.
[0061] Further defined, the specific process of step 2 is:
[0062] Step 2.1: Before identifying urban green space, it is necessary to divide different land cover types according to the actual situation of the field monitoring area, that is, to divide the field monitoring area in the remote sensing image into four different land cover types, namely urban green space, water body, bare land and impervious surface; then based on the remote sensing image, for each land cover type, a manual annotation method is used to evenly select multiple regions of interest as samples on the remote sensing image. In order to eliminate the negative impact of the unbalanced data set, the number of samples of the four land cover types is required to be the same, for example, 1000 regions of interest for each land cover type are manually annotated, and each region of interest contains multiple pixels; finally, the stratified sampling method is used to randomly select the same number of samples from all samples of the four land cover types as training samples, and the remaining samples of various land cover types are used as verification samples.
[0063] Preferably, the number of samples of each land cover type is Y, and there are 4Y samples in total for the four land cover types. 2 / 3 of the Y samples of each land cover type are randomly selected as training samples, and the remaining 1 / 3 of the samples of each land cover type are used as verification samples, where Y>200. The number of randomly selected samples for each land cover type can be determined according to actual conditions. Generally, the more samples, the better. Usually, 200, 300 or 400 areas of interest can be randomly selected as samples for each land cover type.
[0064] Step 2.2: Calculate the Normalized Difference Vegetation Index (NDVI) image of the remote sensing image. The value of any pixel in the Normalized Difference Vegetation Index image of the remote sensing image is recorded as NDVI. NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ), where ρ NIR Represents the spectral reflectance of the corresponding position in the narrow near-infrared band of the remote sensing image, ρ Red Represents the reflectivity of the corresponding position in the red light band of the remote sensing image. Here, NDVI is the best indicator factor for evaluating the growth status of green vegetation and the spatial distribution density of vegetation. Due to its excellent vegetation characterization performance, NDVI has become one of the most important indicators in the field of remote sensing. This indicator quantifies vegetation by measuring the difference between near-infrared (strong reflection of vegetation) and red light (absorption of vegetation). The value range of NDVI is between -1 and 1. The higher the value, the higher the vegetation coverage. Therefore, based on the red and near-infrared bands of Sentinel-2, the present invention calculates the NDVI image of the study area to enhance the separability between different land cover categories and thus improve the accuracy of urban green space identification.
[0065] Step 2.3: Use the training samples, remote sensing images and their normalized difference vegetation index images to train a random forest classifier to obtain a trained random forest classifier; then input the remote sensing image into the trained random forest classifier to classify each land cover type in the field monitoring area in the remote sensing image to obtain a classified result image, in which the spatial distribution of the land cover types of all pixels in the field monitoring area in the remote sensing image is clearly defined, and the spatial distribution of urban green space is extracted, that is, the urban green space in the remote sensing image is identified; wherein, when the random forest classifier is used for training, the initial value of the random forest number (ntree) is 500. Here, the random forest classification technology uses a self-service sampling aggregation method (i.e., bagging method) to extract multiple training samples from the original samples and perform decision tree modeling, of which about 2 / 3 of the samples (i.e., in-bag samples) are used for model training, and the remaining 1 / 3 of the samples (i.e., out-of-bag samples) are used for internal cross-validation of the model (i.e., out-of-bag error, out-of-bag-error). Each decision tree is generated independently without any pruning, and each node is split based on the number of features defined by the user (Mtry); then, the decision tree is grown to a user-defined random forest tree (ntree); finally, the prediction results generated by multiple decision trees are combined, and the final classification result is obtained by voting. The random forest classifier is processed based on the "randomforest" plug-in in the ENVI53 version.
[0066] Step 2.4: Based on the verification samples and the classified result images, the confusion matrix method is used to select four indicators to evaluate the recognition accuracy of the acquired urban green space. The four indicators are overall accuracy, kappa coefficient, user accuracy and mapping accuracy. If all four indicators meet the accuracy requirements, the identified urban green space will be used as the final recognition result. If some of the four indicators do not meet the accuracy requirements, adjust the ntree parameters in the random forest classifier and return to step 2.3 to continue training. Figure 2 The overall accuracy of the urban green space shown is 95.58%, the kappa coefficient reaches 0.9297, the user accuracy is 96.7%, and the user accuracy of the impervious surface is 97.1%, which can meet the accuracy requirements of land cover evaluation in the study area.
[0067] Step 3: After identifying the urban green space in the remote sensing image, determine the actual urban green space; then use the random sampling method to randomly select multiple urban green space plots to be investigated from the actual urban green space; then based on expert knowledge and visual interpretation, delete all randomly selected urban green space plots with unstable growth status; then conduct field investigations and on-site surveys on all retained urban green space plots, record the longitude and latitude of each retained urban green space plot, and clarify the actual geographical location of each urban green space plot. Here, the random sampling method can eliminate the influence of human factors on the results; the interval between green space plots is at least 1 km.
[0068] Specifically, in step 3, the number of randomly selected urban green space plots is 200. To ensure a certain number of green space plots, the influence of accidental factors on the results can be eliminated. In order to adapt to the high heterogeneity of the main environmental pressures affecting biomass, combined with the plant composition and canopy density in the green space, the size of the urban green space plot is set to 30 meters × 30 meters; the urban green space plot with unstable growth status refers to the urban green space plot with large-scale construction activities or obvious human interference in the past two or three years; the actual geographical location of the urban green space plot refers to the actual latitude and longitude coordinates marked by the BIGMAP software during the field investigation and on-site survey. Here, it is assumed that there are 79 urban green space plots with unstable growth status and 121 urban green space plots for field survey.
[0069] It is necessary to lay out green space plots in the urban green space in the study area. The purpose of laying out urban green space plots is to sample the vegetation in the urban green space, obtain basic data for the study, and use this as a basis for quantitatively estimating the carbon storage of the urban green space in the study area. When selecting urban green space plots, in order to ensure the typicality of the plots and to be able to represent the study area, the following two principles should be followed when selecting urban green space plots: (1) random sampling should be used to exclude the influence of human factors on the results; (2) a certain number of urban green space plots should be guaranteed to exclude the influence of accidental factors on the results. Based on the random sampling method, 200 urban green space plots were randomly collected from the built-up area of Wuhan City, and the geographical location and spatial information of each urban green space plot were obtained. Based on visual interpretation, urban green space plots with unstable status were deleted to ensure that there were no large-scale construction activities and obvious human interference in the urban green space plots in recent years, and the approximate location of the urban green space plots was preliminarily determined. Finally, after conducting field surveys, the actual latitude and longitude coordinates of urban green space plots were obtained based on the BIGMAP map, and 121 urban green space plots were determined, such as Figure 3 shown.
[0070] Step 4: During the field survey and on-site investigation of each retained urban green space plot, conduct a tree survey of all vegetation types of interest in each urban green space plot, and record the type and growth parameters of each plant; the vegetation types of interest include trees and shrubs, the growth parameters of trees include DBH, crown width and plant height, and the growth parameters of shrubs include base diameter, crown width and plant height. Here, the DBH of the tree is measured at 1.3 meters above the ground, and there will be a large uncertainty in the measurement of the tree height, so the value here is an estimate.
[0071] From September to October 2022, the project team conducted a detailed survey of urban green spaces in the main built-up areas of Wuhan. Based on the 121 pre-set urban green space plots, the field survey was carried out, and each tree in all urban green space plots was surveyed, recording in detail the types, number of plants and growth parameters of all plants. Figure 4 The results of field survey of urban green space plots in major built-up areas of Wuhan City are given.
[0072] Step 5: As vegetation in cities cannot be destroyed at will and plants cannot be harvested for laboratory biomass and carbon storage measurements, the estimation of urban green space vegetation carbon storage will be mainly based on urban vegetation inventory data. According to the growth parameters of each plant in each urban green space sample plot, the total biomass of each plant in each urban green space sample plot is calculated; then the total biomass of each plant in each urban green space sample plot is multiplied by the corresponding carbon content coefficient to obtain the carbon storage of each plant in each urban green space sample plot; and then the total carbon storage of all vegetation in each urban green space sample plot is calculated.
[0073] Further defined, in step 5, the process of calculating the total biomass of each plant in each urban green space plot is:
[0074] Step 5.1: Based on the allometric growth equation of each plant, calculate the aboveground biomass of each plant in each urban green space plot, and record the aboveground biomass of any plant in any urban green space plot as W above , if the vegetation is a tree and its height is difficult to measure accurately, then W above =aD b , if the vegetation is a tree and its height is accurately measured or if the vegetation is a shrub, then W above =a(D 2 H) b ; Among them, W aboveThe unit is kg. When the vegetation is a tree, D represents the diameter at breast height of the vegetation. When the vegetation is a shrub, D represents the basal diameter of the vegetation. The unit of D is cm. H represents the plant height of the vegetation. The unit of H is m. Both a and b are regression constants. Vegetation (especially trees and shrubs) has site adaptability. The biomass models of different vegetation of the same type in different regions are different, which is reflected in the different specific values of a and b. If there is no corresponding biomass model for the vegetation, then the corresponding biomass model of other vegetation of the same genus and family is considered as a substitute. For example, if this plant belongs to an arbor, the corresponding arbor biomass model is used, such as The willow model is used for willows, the Robinia model is used for Robinia, and the Osmanthus model is used for Osmanthus. If the model for each vegetation cannot be found, look for the biomass model of the same genus. For example, the same genus of Robinia is Sophora japonica. If the biomass model of Sophora japonica is found, use the biomass model of Sophora japonica to calculate the biomass of Robinia japonica in the green space sample plot; if even the biomass model of the same genus cannot be found, look for the biomass model of the tree of the Leguminosae family, such as the Delonix regia, and use the biomass model of the Delonix regia for calculation; if even the same family cannot be found, use the biomass model of deciduous trees.
[0075] The biomass model of a single tree is usually constructed with the diameter at breast height or the diameter at breast height + plant height as variables. In actual measurements, trees are generally tall and it is difficult to accurately measure their plant height. Therefore, in order to minimize the impact of plant height errors on the calculation results, the diameter at breast height of the tree can be used as a variable to calculate vegetation biomass and carbon storage. Since vegetation is site-specific, the currently constructed biomass models are mostly targeted at specific tree species in specific areas. Therefore, when selecting a model, a model similar or close to the site conditions in Wuhan should be selected as much as possible. If there is no corresponding model, a unified model of other vegetation species of the same genus and family should be considered instead. Table 2 lists the biomass models of the aboveground parts such as trunks, branches, and leaves of some trees and shrubs in Wuhan.
[0076] Table 2 Main biomass models of trees and shrubs
[0077]
[0078] Note: S , W B , W L They represent stem biomass, branch biomass and leaf biomass respectively; DBH represents the breast height diameter of trees (cm); D and H represent the base diameter (cm) and plant height (m) of shrubs respectively.
[0079] Step 5.2: Calculate the underground biomass of each plant in each urban green space sample plot, and record the underground biomass of any plant in any urban green space sample plot as W under , W above With W under The ratio is 4.28.
[0080] Step 5.3: Calculate the total biomass of each plant in each urban green space plot, and record the total biomass of any plant in any urban green space plot as W total If the vegetation is trees or unpruned shrubs, then W total =W above +W under , if the vegetation is shrubs that are often pruned, then W total =(W above +W under )×n%; n% is the total biomass reduction ratio of the vegetation, which is generally 20%. If the shrubs are pruned manually, the shrubs can be set to reduce a certain percentage of biomass according to the actual situation, such as 20%. Pruning of trees is not common in Wuhan, but shrubs are pruned more frequently. Therefore, it is set that shrubs will reduce 20% of underground biomass.
[0081] Here, based on the published literature, the carbon content coefficient of each plant in each urban green space sample can be found. Table 3 shows the carbon content coefficient of common vegetation in Wuhan.
[0082] Table 3 Carbon content coefficient of common vegetation in Wuhan
[0083]
[0084] Further, in step 5, the process of calculating the total carbon storage of all vegetation in each urban green space sample plot is as follows: the total carbon storage of all vegetation in any urban green space sample plot is recorded as C total , Where N represents the number of vegetation in any urban green space sample, C i Represents the carbon storage of the i-th plant in any urban green space plot.
[0085] Step 6: According to the actual geographical location recorded in each urban green space plot, the corresponding coordinate position is matched in the remote sensing image; then, with the coordinate position of each urban green space plot matched in the remote sensing image as the center, the spatial offset method is used to spatially offset and expand the urban green space plot in the surrounding areas in the remote sensing image to obtain 9 times the new urban green space plots. The spatial resolution of each new urban green space plot is consistent with the spatial resolution of the remote sensing image; finally, the total carbon storage of all vegetation in the urban green space plot is evenly distributed according to the number of new urban green space plots corresponding to each urban green space plot, and the carbon storage of each new urban green space plot is obtained. Figure 5 A schematic diagram of the process of offset, expansion and average distribution of carbon storage in an urban green space plot is given.
[0086] Here, the size of the new urban green space plot is set to 10 meters × 10 meters. For any urban green space plot surveyed in the field, 9 new urban green space plots are obtained after spatial offset and expansion in the remote sensing image. Since the size of the urban green space plot is 30 meters × 30 meters, and the spatial resolution of the remote sensing image is 10 meters, in order to match the spatial resolution, the urban green space plot needs to be offset and the carbon storage is redistributed according to the number of new urban green space plots obtained by offset expansion. For example, assuming that the total carbon storage of all vegetation in an urban green space plot is 100g, it is expanded into 9 new urban green space plots, then the total carbon storage of all vegetation in each new urban green space plot is 100 / 9 = 11.11g. A total of 1089 new urban green space plots were obtained by expanding 121 urban green space plots.
[0087] Step 7: Extract 20 remote sensing parameters corresponding to each new urban green space plot in the remote sensing image, including 12 spectral information and 8 vegetation indices corresponding to the new urban green space plot; the 12 spectral information corresponding to the new urban green space plot is the spectral reflectance of the corresponding position in the 12 bands of the remote sensing image, and the 8 vegetation indices corresponding to the new urban green space plot are the values of the corresponding positions in the 8 vegetation index images calculated based on the remote sensing image. The 8 vegetation index images are Normalized Difference Vegetation Index (NDVI) image, Normalized Green Difference Vegetation Index (GNDVI) image, Difference Vegetation Index (DVI) image, Enhanced Vegetation Index (EVI) image, Land Surface Water Index (LSWI) image, Simple Ratio Vegetation Index (SWI) image, and 12 spectral information corresponding to the new urban green space plot. Here, the remote sensing parameters of each new sample plot of urban green space can be extracted through the "Multi-value Extraction to Point" tool in ArcGIS 10.8.
[0088] Further defined, in step 7, the process of extracting 20 remote sensing parameters corresponding to a new urban green space plot is:
[0089] Step 7.1: Extract 12 types of spectral information corresponding to the new urban green space sample: Use the point value extraction method to obtain each type of spectral information corresponding to the new urban green space sample. Any type of spectral information is the spectral reflectance of the corresponding position of the new urban green space sample in the corresponding band of the remote sensing image.
[0090] Step 7.2: Extract the eight vegetation indices corresponding to the new sample plot of urban green space: record the value of any pixel in the normalized difference vegetation index image as NDVI, NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ), the value of any pixel in the normalized green difference vegetation index image is recorded as GNDVI, GNDVI = (ρ NIR -ρ Green ) / (ρ NIR +ρ Green ), the value of any pixel in the difference vegetation index image is recorded as DVI, DVI = ρ NIR -ρ Red , the value of any pixel in the enhanced vegetation index image is recorded as EVI, EVI = 2.5 × ((ρ NIR -ρ Red ) / (ρ NIR +6ρ Red -7.5ρ Blue +1)), the value of any pixel in the vegetation moisture index image is recorded as LSWI, LSWI = (ρ NIR -ρ SWIR1 ) / (ρ NIR +ρ SW1R1 ), the value of any pixel in the ratio vegetation index image is recorded as SR, SR = ρ NIR / ρ Red , the value of any pixel in the soil adjustment index image is recorded as SAVI, SAVI = ((ρ NIR -ρ Red )(1+0.5)) / (ρ NIR +ρ Red +0.5), and the value of any pixel in the modified soil conditioning index image is recorded as MSAVI, Then, the point value extraction method is used to obtain each vegetation index corresponding to the new urban green space sample plot. Any vegetation index is the value of the corresponding position of the new urban green space sample plot in the corresponding vegetation index image; where ρ NIR Represents the spectral reflectance of the corresponding position in the narrow near-infrared band of the remote sensing image, ρ Red Represents the spectral reflectance of the corresponding position in the red light band of the remote sensing image, ρ GreenRepresents the spectral reflectance of the corresponding position in the green light band of the remote sensing image, ρ Blue Represents the spectral reflectance of the corresponding position in the blue light band of the remote sensing image, ρ SWIR1 Indicates the spectral reflectance of the corresponding position in the short-side infrared band of the remote sensing image.
[0091] Step 8: Based on the carbon stock of each new urban green space sample in the remote sensing image and its corresponding 20 remote sensing parameters, the random forest regression model was used to preliminarily fit the nonlinear relationship between the carbon stock of the new urban green space sample and the 20 remote sensing parameters, and the preliminary urban green space carbon stock assessment model and the MDA parameters corresponding to each remote sensing parameter were obtained; then, the MDA parameters were arranged in descending order, and the top five remote sensing parameters were selected as the optimal remote sensing parameters; based on the carbon stock of each new urban green space sample in the remote sensing image and its corresponding five optimal remote sensing parameters, the random forest regression model was used again to obtain the optimized urban green space carbon stock assessment model; finally, after the model robustness test, the final urban green space carbon stock assessment model was determined.
[0092] Further defined, the specific process of step 8 is:
[0093] Step 8.1: Use the carbon storage of each new urban green space sample in the remote sensing image and its corresponding 20 remote sensing parameters, and train the random forest regression model to fit the nonlinear relationship between the carbon storage of all new urban green space samples and their corresponding 20 remote sensing parameters. This nonlinear relationship is defined as the preliminary assessment model of urban green space carbon storage. The training also obtains the MDA (mean decrease accuracy) parameter corresponding to each remote sensing parameter. Here, the random forest regression model is run based on the "randomForest" package in the R language (version 4.2.1).
[0094] The random forest regression model is a machine learning model and an ensemble learner. It is a combination algorithm that uses multiple decision trees for prediction and can be used to process or regression problems. In remote sensing data processing, random forests have been widely used in many fields such as urban research, land research, and ecological research due to their fast processing speed, easy parameterization, and robustness. The specific method is to construct the vegetation carbon storage of the sample plot and define the remote sensing parameters corresponding to the same geographical location of the sample plot. The two are analyzed by parameters to establish a nonlinear model of vegetation carbon storage. Therefore, based on the random forest regression model, the processing steps for constructing the Wuhan urban green space carbon storage assessment model include the following steps, see Figure 6 :
[0095] (1) Data preparation: Collect and organize the data of the regression problem to ensure that there are no missing values and outliers in the data, namely, multi-band image data and sampling point data;
[0096] (2) Random sampling: A bootstrap sampling aggregation method (i.e., bagging method) is used to extract multiple training samples from the original sample for decision tree modeling, of which about 2 / 3 of the samples (i.e., in-bag samples) are used for model training, and the remaining 1 / 2 of the samples (i.e., out-of-bag samples) are used for internal cross-validation of the model (i.e., out-of-bag error);
[0097] (3) Feature selection: For each node of each decision tree, not all features are used during training. Instead, a certain number of features are randomly selected from all features for splitting;
[0098] (4) Building decision trees: Use randomly sampled training data and randomly selected features to build multiple decision trees. Each tree uses a certain feature to split the data at the node until a stopping condition is reached, such as the depth of the tree reaches a predetermined value or the number of samples at the node is less than a certain threshold;
[0099] (5) Ensemble prediction: For regression problems, each decision tree outputs a prediction value. Random forests obtain the final prediction result by averaging (or weighted averaging) the prediction results of all decision trees;
[0100] (6) Model evaluation: The performance of the model was evaluated using the test data set. The prediction accuracy of the urban green space carbon storage assessment model was determined based on the goodness of fit of the random forest.
[0101] (7) Tuning parameters: Random forest has some key hyperparameters, such as the number of decision trees, the maximum depth of each tree, the number of feature selections, etc.
[0102] The random forest regression model used in the present invention is implemented based on the "randomForest" package (version 4.6-14) of the R environment software (version 4.1.2). The number of regression trees (ntree) is set to 800, and the number of features used on each node (mtry) is set to the square root of the total number of input features.
[0103] Step 8.2: According to the MDA parameters corresponding to the 20 remote sensing parameters, the 20 remote sensing parameters are screened in descending order, and the remote sensing parameters corresponding to the top 5 MDA parameters are screened as the optimal remote sensing parameters.
[0104] For urban green space carbon stock assessment, general users will face large amounts of data and complex parameter information. Especially for high-resolution remote sensing data such as Sentinel-2, they will face tens of millions of data volumes. When constructing urban green space carbon stocks, the initial remote sensing parameters include 12 types of spectral information and 8 vegetation indices. Therefore, in order to avoid data redundancy and save model prediction time to improve the universality of the model, the present invention will perform feature screening based on the importance of multiple remote sensing parameters during the construction of the urban green space carbon stock assessment model, and select the TOP5 features to finally optimize the construction of the urban green space carbon stock assessment model. When explaining the dependent variable, the random forest regression model can evaluate the importance of the independent variable through feature permutation, that is, the mean accuracy decrease (MDA). Simply put, in order to overcome the problem of feature space autocorrelation, random forest will introduce randomness when constructing the regression model, thereby effectively reducing the correlation between variables. MDA is used to evaluate the importance of 20 remote sensing parameters (see Figure 7 ). The larger the MDA, the more important the feature. Finally, the top five remote sensing parameters in terms of feature importance will be used in the construction of the urban green space carbon storage assessment model and for prediction. Figure 7 The remote sensing parameters selected for the sample data are B9, B1, B11, B5, and B12, for the water vapor band, coast / aerosol band, shortwave infrared band 1, red edge band 1, and shortwave infrared band 2 of the Sentinel-2 image.
[0105] Step 8.3: Using the carbon storage of each new urban green space sample in the remote sensing image and the five optimal remote sensing parameters selected, a random forest regression model is trained to fit the nonlinear relationship between the carbon storage of all new urban green space samples and the five optimal remote sensing parameters. This nonlinear relationship is defined as the optimized urban green space carbon storage assessment model.
[0106] Step 8.4: Use the goodness of fit parameters of the random forest regression model in step 8.3 to detect the robustness of the optimized urban green space carbon stock assessment model. If the robustness is good, the optimized urban green space carbon stock assessment model is used as the final urban green space carbon stock assessment model. If the robustness is poor, first adjust the ntree parameter of the random forest regression model in step 8.3, then return to step 8.3 and re-test the robustness. If the robustness still does not meet the experimental requirements, adjust the ntree parameter of the random forest regression model in step 8.1 and re-execute. Perform step 8.1 until the goodness of fit parameter of the final urban green space carbon storage assessment model meets the experimental requirements; wherein, the initial value of the ntree parameter is 800 when the random forest regression model is used for training in step 8.1, and the initial value of the ntree parameter is 800 when the random forest regression model is used for training in step 8.3. When the goodness of fit parameter of the random forest regression model in step 8.3 is greater than or equal to 85%, it is considered to have good robustness, and when the goodness of fit parameter of the random forest regression model in step 8.3 is less than 85%, it is considered to have poor robustness. The goodness of fit of the random forest regression model in Wuhan is 94.48%, which has a good fitting effect. Generally, if the goodness of fit is >80%, it is considered that the model has high reliability and can be used to assess the carbon storage of regional urban green spaces. If the fit is too low, the user needs to retrain the model until the fit requirement is met.
[0107] Step 9: Synthesize the five images corresponding to the five optimal remote sensing parameters into a remote sensing image of urban green space carbon storage to be predicted; then input the remote sensing image to be predicted into the final urban green space carbon storage assessment model for simulation and prediction, and obtain the carbon storage prediction result image of the remote sensing image to be predicted; then, based on the urban green space space in the remote sensing image identified in step 2, mask the carbon storage prediction result image of the remote sensing image to be predicted, and obtain the predicted carbon storage of each pixel in the urban green space space in the remote sensing image to be predicted. Here, the 12 bands of the remote sensing image and the 8 vegetation index images are all a sub-image, and the synthesis of the five images can use spectral synthesis technology.
[0108] Taking the main built-up areas of Wuhan as an example, the carbon storage of urban green spaces was simulated and predicted based on the final urban green space carbon storage assessment model. The results are as follows: Figure 8As shown. The areas with higher carbon storage in Wuhan's urban green spaces are mainly distributed in the Jiufeng Mountain Forest Reserve in the central part of the city, the Longquan Mountain Scenic Area in the southeast, the Baiyun Cave Scenic Area and the Qinglong Mountain Forest Reserve in the south, and the Jiuzhen Mountain Scenic Area in the west. The vegetation types in these areas are mainly forests, with dense trees and rich vegetation resources. Most of them are located in ecological protection areas. Urban vegetation can obtain better ecological protection conditions, thereby absorbing and storing more CO2, providing stronger carbon sink capacity, and is the main carbon source for Wuhan's green spaces. Low-carbon areas often border non-vegetated surfaces, and are mostly located in the transition area between the central area and suburbs of Wuhan, or in farm areas where facility agriculture is concentrated around the city. Overall, the carbon storage of urban green spaces in Wuhan mainly shows a trend of high in the south and low in the north.
Claims
1. A method for assessing urban green space carbon storage based on high-resolution images and machine learning, characterized in that The following steps are involved: Step 1: obtain multiple original high-resolution remote sensing images covering the same field monitoring area; then perform image preprocessing on all the original high-resolution remote sensing images obtained to obtain a remote sensing image of the field monitoring area of interest; wherein, the remote sensing image retains 12 bands, namely, the coast / aerosol band, i.e., the B1 band, the blue light band, i.e., the B2 band, the green light band, i.e., the B3 band, the red light band, i.e., the B4 band, the red edge band 1, i.e., the B5 band, the red edge band 2, i.e., the B6 band, the red edge band 3, i.e., the B7 band, the wide near-infrared band, i.e., the B8 band, the narrow near-infrared band, i.e., the B8A band, the water vapor band, i.e., the B9 band, the short-edge infrared band 1, i.e., the B11 band, and the short-edge infrared band 2, i.e., the B12 band; Step 2: In the remote sensing image obtained in step 1, multiple evenly distributed areas of interest are manually marked as samples for various land cover types on the remote sensing image, and the same number of samples are randomly selected from all samples of various land cover types as training samples, and the remaining samples are used as verification samples; then, based on the reflectivity of the narrow near-infrared band and the red light band of the remote sensing image, the normalized difference vegetation index image of the remote sensing image is calculated; then, based on the training samples, the remote sensing image and its normalized difference vegetation index image, a random forest classifier is used to accurately identify the urban green space in the remote sensing image; wherein the land cover types are urban green space, water body, bare land and impervious surface, and the verification samples are used to evaluate the recognition accuracy of the acquired urban green space; Step 3: After identifying the urban green space in the remote sensing image, determine the actual urban green space; then use the random sampling method to randomly select multiple urban green space plots to be investigated from the actual urban green space; then, based on expert knowledge and visual interpretation, delete all the randomly selected urban green space plots with unstable growth status; then conduct field investigations and on-site surveys on all the retained urban green space plots, record the latitude and longitude of each retained urban green space plot, and clarify the actual geographical location of each urban green space plot; Step 4: During the field survey and on-site investigation of each retained urban green space plot, conduct a tree-by-tree survey of all vegetation types of interest in each urban green space plot, and record the type and growth parameters of each plant; the vegetation types of interest include trees and shrubs, the growth parameters of trees include breast diameter, crown width and plant height, and the growth parameters of shrubs include base diameter, crown width and plant height; Step 5: Calculate the total biomass of each plant in each urban green space sample plot based on the growth parameters of each plant in each urban green space sample plot; then multiply the total biomass of each plant in each urban green space sample plot by the corresponding carbon content coefficient to obtain the carbon storage of each plant in each urban green space sample plot; and then calculate the total carbon storage of all vegetation in each urban green space sample plot; Step 6: According to the actual geographical location recorded in each urban green space sample plot, the corresponding coordinate position is matched in the remote sensing image; then, with the coordinate position of each urban green space sample plot matched in the remote sensing image as the center, the spatial offset method is used to perform spatial offset and expansion of the urban green space sample plot in the surrounding areas in the remote sensing image to obtain 9 times the new urban green space sample plots. The spatial resolution of each new urban green space sample plot is consistent with the spatial resolution of the remote sensing image; finally, the total carbon storage of all vegetation in the urban green space sample plot is evenly distributed according to the number of new urban green space sample plots corresponding to each urban green space sample plot, and the carbon storage of each new urban green space sample plot is obtained; Step 7: extract 20 remote sensing parameters corresponding to each new urban green space plot in the remote sensing image, including 12 spectral information and 8 vegetation indices corresponding to the new urban green space plot; wherein, the 12 spectral information corresponding to the new urban green space plot is the spectral reflectance of the corresponding position in the 12 bands of the remote sensing image, and the 8 vegetation indices corresponding to the new urban green space plot are the values of the corresponding positions in 8 vegetation index images calculated based on the remote sensing image, and the 8 vegetation index images are normalized difference vegetation index image, normalized green difference vegetation index image, difference vegetation index image, enhanced vegetation index image, vegetation moisture index image, ratio vegetation index image, soil adjustment index image, and modified soil adjustment index image; Step 8: Based on the carbon storage of each new urban green space sample in the remote sensing image and its corresponding 20 remote sensing parameters, the random forest regression model is used to preliminarily fit the nonlinear relationship between the carbon storage of the new urban green space sample and the 20 remote sensing parameters, and the preliminary urban green space carbon storage assessment model and the MDA parameters corresponding to each remote sensing parameter are obtained; then, the top five remote sensing parameters are selected as the optimal remote sensing parameters according to the descending order of the MDA parameters; then, based on the carbon storage of each new urban green space sample in the remote sensing image and its corresponding five optimal remote sensing parameters, the random forest regression model is used again to obtain the optimized urban green space carbon storage assessment model; finally, after the model robustness test, the final urban green space carbon storage assessment model is determined; Step 9: The five images corresponding to the five optimal remote sensing parameters are synthesized into a remote sensing image of urban green space carbon storage to be predicted; the remote sensing image to be predicted is then input into the final urban green space carbon storage assessment model for simulation and prediction to obtain the carbon storage prediction result image of the remote sensing image to be predicted; based on the urban green space space in the remote sensing image identified in step 2, the carbon storage prediction result image of the remote sensing image to be predicted is masked to obtain the predicted carbon storage of each pixel in the urban green space space in the remote sensing image to be predicted.
2. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 1 is characterized in that In the step 1, the original high-resolution remote sensing image is obtained based on Sentinel-2 remote sensing data, and the level of the original high-resolution remote sensing image is L2A or L1C.
3. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 2 is characterized in that In the step 1, the process of image preprocessing for all the original high-resolution remote sensing images obtained is as follows: if the level of all the original high-resolution remote sensing images obtained is L2A, then all the bands of all the original high-resolution remote sensing images are resampled to 10 meters, and then all the 10-meter bands are mosaicked, and the mosaicked results are cropped according to the scope of the field monitoring area to obtain a remote sensing image containing the field monitoring area of interest; if there are images with levels of L2A and L1C among all the original high-resolution remote sensing images obtained, then the image with level L1C is firstly subjected to radiation calibration and atmospheric correction, and the level of the image obtained after the atmospheric correction is L2A, and then all the bands of all the images with level L2A are resampled to 10 meters, and then all the 10-meter bands are mosaicked, and the mosaicked results are cropped according to the scope of the field monitoring area to obtain a remote sensing image containing the field monitoring area of interest.
4. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 1 is characterized in that The specific process of step 2 is: Step 2.1: Divide the field monitoring area in the remote sensing image into four different land cover types, namely urban green space, water body, bare land and impervious surface; then, based on the remote sensing image, use the manual annotation method to evenly select multiple areas of interest as samples for each land cover type on the remote sensing image, and the number of samples of the four land cover types is the same; finally, use the stratified sampling method to randomly select the same number of samples from all samples of each of the four land cover types as training samples, and use the remaining samples of each land cover type as verification samples; Step 2.2: Calculate the normalized difference vegetation index image of the remote sensing image, and record the value of any pixel in the normalized difference vegetation index image of the remote sensing image as NDVI, NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ); Among them, ρ NIR Represents the spectral reflectance of the corresponding position in the narrow near-infrared band of the remote sensing image, ρ Red Indicates the reflectance of the corresponding position in the red light band of the remote sensing image; Step 2.3: Use the training samples, remote sensing images and their normalized difference vegetation index images to train a random forest classifier to obtain a trained random forest classifier; then input the remote sensing image into the trained random forest classifier to classify each land cover type in the field monitoring area in the remote sensing image to obtain a classified result image, in which the spatial distribution of the land cover types of all pixels in the field monitoring area in the remote sensing image is clearly defined, and the spatial distribution of urban green space is extracted, that is, the urban green space in the remote sensing image is identified; Step 2.4: Based on the verification samples and the classified result images, the confusion matrix method is used to select four indicators to evaluate the recognition accuracy of the acquired urban green space. The four indicators are overall accuracy, kappa coefficient, user accuracy and mapping accuracy. If all four indicators meet the accuracy requirements, the identified urban green space will be used as the final recognition result. If some of the four indicators do not meet the accuracy requirements, adjust the ntree parameters in the random forest classifier and return to step 2.3 to continue training.
5. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 4 is characterized in that In the step 2.1, the number of samples of each land cover type is Y, of which 2 / 3 of the samples are used as training samples and the remaining 1 / 3 of the samples are used as verification samples, where Y>200; in the step 2.3, when the random forest classifier is used for training, the initial value of the random forest number is 500.
6. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 1 is characterized in that In step 3, the number of randomly selected urban green space plots is 200, and the size of the urban green space plot is set to 30 meters × 30 meters; the urban green space plot with unstable growth status refers to the urban green space plot that has had large-scale construction activities or obvious human interference in the past two to three years; the actual geographical location of the urban green space plot refers to the actual latitude and longitude coordinates marked during the field investigation and on-site survey.
7. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 1 is characterized in that In step 5, the process of calculating the total biomass of each plant in each urban green space plot is as follows: Step 5.1: Based on the allometric growth equation of each plant, calculate the aboveground biomass of each plant in each urban green space plot, and record the aboveground biomass of any plant in any urban green space plot as W above , if the vegetation is a tree and its height is difficult to measure accurately, then W above =aD b , if the vegetation is a tree and its height is accurately measured or if the vegetation is a shrub, then W above =a(D 2 H) b ; Among them, W above The unit is kg. When the vegetation is a tree, D represents the diameter at breast height of the vegetation. When the vegetation is a shrub, D represents the base diameter of the vegetation. The unit of D is cm. H represents the plant height. The unit of H is m. a and b are both regression constants. Step 5.2: Calculate the underground biomass of each plant in each urban green space sample plot, and record the underground biomass of any plant in any urban green space sample plot as W under , W above With W under The ratio is 4.28; Step 5.3: Calculate the total biomass of each plant in each urban green space plot, and record the total biomass of any plant in any urban green space plot as W total If the vegetation is trees or unpruned shrubs, then W total =W above +W under , if the vegetation is shrubs that are often pruned, then W total =(W above +W under )×n%; where n% is the reduction ratio of the total biomass of the vegetation; In step 5, the process of calculating the total carbon storage of all vegetation in each urban green space sample plot is as follows: the total carbon storage of all vegetation in any urban green space sample plot is recorded as C total , Where N represents the number of vegetation in any urban green space sample, C i Represents the carbon storage of the i-th plant in any urban green space plot.
8. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 6 is characterized in that In step 6, the size of the new urban green space plot is set to 10 meters × 10 meters. For any urban green space plot surveyed on the spot, 9 new urban green space plots are obtained after spatial offset and expansion in the remote sensing image.
9. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 1 is characterized in that In step 7, the process of extracting 20 remote sensing parameters corresponding to a new urban green space plot is as follows: Step 7.1: Extract 12 types of spectral information corresponding to the new urban green space sample: Use the point value extraction method to obtain each type of spectral information corresponding to the new urban green space sample. Any type of spectral information is the spectral reflectance of the position corresponding to the new urban green space sample in the band corresponding to the remote sensing image; Step 7.2: Extract the eight vegetation indices corresponding to the new sample plot of urban green space: record the value of any pixel in the normalized difference vegetation index image as NDVI, NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ), the value of any pixel in the normalized green difference vegetation index image is recorded as GNDVI, GNDVI = (ρ NIR -ρ Green ) / (ρ NIR +ρ Green ), the value of any pixel in the difference vegetation index image is recorded as DVI, DVI = ρ NIR -ρ Red , the value of any pixel in the enhanced vegetation index image is recorded as EVI, EVI = 2.5 × ((ρ NlR -ρ Red ) / (ρ NIR +6ρ Red -7.5ρ Blue +1)), the value of any pixel in the vegetation moisture index image is recorded as LSWI, LSWI = (ρ NIR -ρ SWIR1 ) / (ρ NIR +ρ SW1R1 ), the value of any pixel in the ratio vegetation index image is recorded as SR, SR = ρ NIR / ρ Red , the value of any pixel in the soil adjustment index image is recorded as SAVI, SAVI = ((ρ NIR -ρ Red )(1+0.5)) / (ρ NIR +ρ Red +0.5), and the value of any pixel in the modified soil conditioning index image is recorded as MSAVI, Then, the point value extraction method is used to obtain each vegetation index corresponding to the new urban green space sample plot. Any vegetation index is the value of the corresponding position of the new urban green space sample plot in the corresponding vegetation index image. Among them, ρ NIR Represents the spectral reflectance of the corresponding position in the narrow near-infrared band of the remote sensing image, ρ Red Represents the spectral reflectance of the corresponding position in the red light band of the remote sensing image, ρ Green Represents the spectral reflectance of the corresponding position in the green light band of the remote sensing image, ρ Blue Represents the spectral reflectance of the corresponding position in the blue light band of the remote sensing image, ρ SWIR1 Indicates the spectral reflectance of the corresponding position in the short-side infrared band of the remote sensing image.
10. The urban green space carbon storage assessment method based on high-resolution imaging and machine learning according to claim 1, characterized in that The specific process of step 8 is as follows: Step 8.1: Using the carbon storage of each new urban green space sample in the remote sensing image and its corresponding 20 remote sensing parameters, the random forest regression model is trained to fit the nonlinear relationship between the carbon storage of all new urban green space samples and their corresponding 20 remote sensing parameters. This nonlinear relationship is defined as the preliminary assessment model of urban green space carbon storage. The MDA parameters corresponding to each remote sensing parameter are obtained during the training. Step 8.2: According to the MDA parameters corresponding to the 20 remote sensing parameters, the 20 remote sensing parameters are sorted in descending order, and the remote sensing parameters corresponding to the top 5 MDA parameters are selected as the optimal remote sensing parameters; Step 8.3: Using the carbon storage of each new urban green space sample in the remote sensing image and the five optimal remote sensing parameters selected, a random forest regression model is used for training to fit the nonlinear relationship between the carbon storage of all new urban green space samples and the five optimal remote sensing parameters. This nonlinear relationship is defined as the optimized urban green space carbon storage assessment model; Step 8.4: Use the goodness of fit parameters of the random forest regression model in step 8.3 to detect the robustness of the optimized urban green space carbon stock assessment model. If the robustness is good, the optimized urban green space carbon stock assessment model is used as the final urban green space carbon stock assessment model. If the robustness is poor, first adjust the ntree parameter of the random forest regression model in step 8.3, then return to step 8.3 and re-test the robustness. If the robustness still does not meet the experimental requirements, adjust the ntree parameter of the random forest regression model in step 8.1 and re-execute. Perform step 8.1 until the final goodness of fit parameter of the urban green space carbon storage assessment model meets the experimental requirements; wherein, the initial value of the ntree parameter when the random forest regression model is used for training in step 8.1 is 800, and the initial value of the ntree parameter when the random forest regression model is used for training in step 8.3 is 800. When the goodness of fit parameter of the random forest regression model in step 8.3 is greater than or equal to 85%, it is considered that the robustness is good, and when the goodness of fit parameter of the random forest regression model in step 8.3 is less than 85%, it is considered that the robustness is poor.
Citation Information
Patent Citations
Urban forest carbon reserve landscape scale deducting method
CN108009384A
Forest interannual phenology monitoring method based on multi-source remote sensing
CN113850139A
Urban forest monitoring method based on cooperation of high-resolution remote sensing data and ground survey data
CN114166842A
Deep network model-based marsh wetland information extraction method and device, medium and terminal
CN115019166A
Urban forest carbon sink measuring and calculating method and system based on remote sensing image and machine learning
CN117036928A
