Method and system for measuring urban forest carbon sink based on remote sensing image and machine learning

Abstract

The application provides a kind of urban forest carbon sink measurement method and system based on remote sensing image and machine learning, comprising: collecting remote sensing image of forest to be estimated in urban area;The collected remote sensing image is preprocessed to form multi-band remote sensing image data;Based on the multi-band remote sensing image data, the vegetation index under different bands is constructed, the correlation and importance of the vegetation index under different bands are analyzed, and the vegetation index under the optimal band is obtained;Based on the vegetation index under the optimal band, a leaf chlorophyll content estimation model of urban forest is constructed by combining a machine learning method;Obtain the vegetation stand factor and meteorological factor in each sample area;Based on the estimated leaf chlorophyll content, vegetation stand factor and meteorological factor, an urban forest carbon sink measurement model is constructed by combining a machine learning model.The application realizes small-scale urban forest carbon sink measurement with high precision, high intelligence and strong applicability based on remote sensing image and machine learning model.

Description

Technical Field

[0001] The present invention relates to the technical field of urban forest carbon sink calculation, and in particular to a method and system for urban forest carbon sink calculation based on remote sensing imagery and machine learning. Background Art

[0002] Cities are major energy consumers in my country, contributing 85% of the country's carbon dioxide emissions and playing a key role in climate change. Forests are the backbone of terrestrial ecosystems. Increasing forest carbon sinks through afforestation and forest conservation, absorbing atmospheric carbon dioxide and sequestering it within trees, is an effective way to mitigate rising atmospheric carbon dioxide concentrations and achieve carbon neutrality. Urban forests offer the dual ecological benefits of increasing urban carbon sinks and reducing carbon emissions. Therefore, urban forests have become crucial for promoting low-carbon, circular urban development. In recent years, the carbon storage and sequestration capacity of urban forests has garnered widespread attention. Therefore, developing a highly accurate, intelligent, and applicable method for estimating urban forest carbon sinks has become a pressing need. This approach is also crucial for my country's sustainable urban development and its green economic transition.

[0003] With the advancement of remote sensing technology, the emergence of digital satellite products such as hyperspectral and multispectral data has provided new technical and data support for forest carbon sink estimation, offering a new approach to research. However, current research on forest carbon sink estimation based on remote sensing imagery has mostly focused on large-scale forest areas. Research on small-scale urban forest carbon sink estimation is relatively limited. Furthermore, in forest carbon sink estimation models, many studies directly use vegetation indices such as the Normalized Difference Vegetation Index (NDVI), the Difference Vegetation Index (DVI), and the Ratio Vegetation Index (RVI), as well as spectral factors, as model feature variables. These variables not only suffer from information redundancy and correlation, but also make it difficult to more intuitively and accurately represent the carbon sequestration capacity of forests. Regarding the data dimensions used for forest carbon sinks, most studies have only considered remote sensing imagery data, which, however, is acquired from high altitudes and therefore represents horizontal data.

[0004] The Chinese invention patent document with publication number CN104820065A discloses a method for calculating the carbon sequestration of single urban trees. First, the crown light energy utilization rate G, crown width D, root respiration ratio B, leaf shading rate f and instantaneous photosynthetic rate Pi are measured to calculate the annual net carbon amount Wa per unit leaf area. Then, the management coefficient H and the total leaf area LAT = S × LAI, S = π × D2 / 4, and the average light intensity per unit leaf area Y = 1 / (f × LAI) are substituted into the formula I, that is, CL = H × G × (1-B) × Wa × LAT × Y, to calculate the carbon sequestration value of the single tree.

[0005] Regarding the above-mentioned related technologies, the inventors believe that due to the complexity, nonlinearity, high dimensionality and other characteristics of remote sensing image data, many traditional measurement methods find it difficult to accurately capture high-value information from it. Summary of the Invention

[0006] In view of the defects in the existing technology, the purpose of the present invention is to provide a method and system for measuring urban forest carbon sequestration based on remote sensing imagery and machine learning.

[0007] According to the present invention, a method for calculating urban forest carbon sequestration based on remote sensing images and machine learning includes the following steps:

[0008] Remote sensing image acquisition steps: Collect remote sensing images of the forest to be assessed within the urban area;

[0009] Remote sensing image preprocessing step: preprocess the collected remote sensing images to form multi-band remote sensing image data;

[0010] Vegetation index selection steps: Based on multi-band remote sensing image data, vegetation indices in different bands are constructed. The correlation coefficient method and machine learning fusion model are used to analyze the correlation and importance of vegetation indices in different bands to obtain the vegetation index in the optimal band.

[0011] Chlorophyll content estimation steps: Based on the vegetation index in the optimal band and combined with machine learning methods, a chlorophyll content estimation model for urban forests is constructed; chlorophyll content is estimated using the chlorophyll content estimation model;

[0012] Factor acquisition steps: select sample plots in the urban forest area to be assessed, and obtain the vegetation stand factors and meteorological factors in each sample plot;

[0013] Carbon sink calculation steps: Based on the estimated chlorophyll content, vegetation stand factors and meteorological factors, combined with the machine learning model, an urban forest carbon sink calculation model is constructed, and the urban forest carbon sink calculation model is used to calculate urban forest carbon sinks.

[0014] Preferably, in the remote sensing image acquisition step, the sources of the remote sensing images of the forests to be estimated in the urban area include satellite digital products, multispectral remote sensing image products and elevation image products in the data platform; the forests to be estimated include urban woodlands in the urban area.

[0015] Preferably, in the remote sensing image preprocessing step, the preprocessing operations performed on the collected remote sensing images include radiometric calibration, atmospheric correction, terrain correction, image stitching, masking, and cropping;

[0016] After removing redundant images, multi-band remote sensing image data of the forest area to be assessed are obtained.

[0017] Preferably, in the vegetation index selection step, based on multi-band remote sensing image data, image bands representing vegetation chlorophyll are selected, and vegetation indices are constructed based on different band information;

[0018] The constructed vegetation indices include normalized difference vegetation index, ratio vegetation index and chlorophyll index;

[0019] Select a preset number of fixed-size plots within the urban forest area to be assessed. The plots are evenly distributed throughout the urban forest area, and the trees within the plots represent the average growth conditions of trees within the urban forest area to be assessed. Select a number of leaves within the plots and measure the target chlorophyll content using an ultraviolet spectrophotometer.

[0020] The correlation coefficient and machine learning model were used to conduct correlation analysis between vegetation indices in different bands, and the importance analysis was conducted between vegetation indices in different bands and chlorophyll content target values ​​to obtain the vegetation index in the optimal band.

[0021] Preferably, in the chlorophyll content estimation step, a data set is constructed based on the optimal vegetation index and chlorophyll content target value of each sample plot in the urban forest; the optimal vegetation index of each sample plot in the urban forest is used as a characteristic variable for model input, and the chlorophyll content target value of each sample plot forest is used as the model output. Based on the data set combined with the particle swarm optimization least squares support vector machine machine learning model, after model training and parameter optimization and adjustment, a chlorophyll content estimation model for the urban forest is finally constructed.

[0022] Preferably, in the factor acquisition step, a preset number of fixed-size plots are selected in the urban forest area to be estimated, the stand factors of all trees in the plots are measured by the tree-per-plot method, and the forest biomass in the plots is calculated based on the stand factors; and the meteorological factors in each plot in the forest area to be estimated are obtained based on relevant meteorological environment sensors.

[0023] Preferably, in the carbon sink calculation step, an urban forest carbon sink calculation model is constructed based on the chlorophyll content, vegetation stand factors and meteorological factors of the forest to be estimated in the urban area, combined with a deep belief network.

[0024] The present invention provides an urban forest carbon sink calculation system based on remote sensing imagery and machine learning, including the following modules:

[0025] Remote sensing image acquisition module: collects remote sensing images of forests to be assessed within urban areas;

[0026] Remote sensing image preprocessing module: preprocesses the collected remote sensing images to form multi-band remote sensing image data;

[0027] Vegetation index selection module: Constructs vegetation indices in different bands based on multi-band remote sensing image data, uses the correlation coefficient method and machine learning fusion model to analyze the correlation and importance of vegetation indices in different bands, and obtains the vegetation index in the optimal band;

[0028] Chlorophyll content estimation module: Based on the vegetation index in the optimal band and combined with machine learning, a chlorophyll content estimation model for urban forests is constructed; chlorophyll content is estimated using the chlorophyll content estimation model;

[0029] Factor acquisition module: select sample plots in the urban forest area to be assessed and obtain vegetation stand factors and meteorological factors in each sample plot;

[0030] Carbon sink calculation module: Based on the estimated chlorophyll content, vegetation stand factors and meteorological factors, combined with the machine learning model, an urban forest carbon sink calculation model is constructed, and the urban forest carbon sink calculation model is used to calculate urban forest carbon sinks.

[0031] Preferably, in the remote sensing image acquisition module, the sources of remote sensing images of the forests to be estimated in the urban area include satellite digital products, multispectral remote sensing image products and elevation image products in the data platform; the forests to be estimated include urban woodlands in the urban area.

[0032] Preferably, in the remote sensing image preprocessing module, the preprocessing operations performed on the collected remote sensing images include radiometric calibration, atmospheric correction, terrain correction, image stitching, masking, and cropping;

[0033] After removing redundant images, multi-band remote sensing image data of the forest area to be assessed are obtained.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This invention uses remote sensing images and machine learning models to implement a highly accurate, intelligent, and applicable method for calculating small-scale urban forest carbon sinks.

[0036] 2. This paper uses multi-band remote sensing imagery to extract chlorophyll content information, which reflects the carbon sequestration capacity of vegetation under photosynthesis, as a characteristic factor (optimal band vegetation index). It also proposes a particle swarm optimization least squares support vector machine machine learning method as a chlorophyll content estimation model for urban forests. This method improves the model's information representation capability in the urban forest carbon sequestration process and enhances the model's accuracy and precision.

[0037] 3. This invention integrates multi-dimensional data from forest horizontal dimensions represented by multi-band remote sensing imagery with vertical dimensions represented by vegetation stand factors and meteorological factors, improving the reliability and robustness of the urban forest carbon sink estimation model.

[0038] 4. In the urban forest carbon sink calculation model, this invention uses a deep belief network method to effectively process high-dimensional and nonlinear data. Compared with traditional methods and shallow network models, it has stronger feature learning and expression capabilities for the diversity and multi-dimensional characteristics of remote sensing data, measured data, horizontal data, and vertical data, thereby further improving the accuracy and effectiveness of urban forest carbon sink calculation results.

[0039] 5. The present invention provides a method with practical application value for urban carbon cycle and sustainable development. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0041] Figure 1 Schematic diagram of a flow chart of an embodiment of a method for calculating urban forest carbon sequestration based on remote sensing images and machine learning according to the present invention;

[0042] Figure 2 Schematic diagram of the specific steps of acquiring and preprocessing remote sensing images described in the embodiment;

[0043] Figure 3 This is a schematic diagram of the specific steps for constructing and selecting the vegetation index described in the embodiment;

[0044] Figure 4 This is a schematic diagram of the specific steps of the vegetation chlorophyll content estimation model described in the embodiment;

[0045] Figure 5 This is a schematic diagram of the specific steps for obtaining the forest stand factors / meteorological factors described in the embodiment;

[0046] Figure 6 This is a schematic diagram of the specific steps for constructing the urban forest carbon sink calculation model described in the embodiment. DETAILED DESCRIPTION

[0047] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0048] The embodiment of the present invention discloses a method for calculating urban forest carbon sinks based on remote sensing images and machine learning. Figure 1 As shown, the following steps are included:

[0049] S1. Remote sensing image acquisition: Collect remote sensing images of the forest to be assessed within the urban area.

[0050] Based on the embodiment, the acquisition of the remote sensing image, such as Figure 2 As shown, the specific implementation steps include:

[0051] Specifically, remote sensing images can be obtained from data platforms such as Google Earth Engine (GEE), Geospatial Data Cloud, and the Chinese Academy of Sciences Data Cloud, but are not limited to these. Remote sensing images of forests to be assessed in urban areas can come from data platforms such as Landsat8 OLI_TIRS satellite digital products, multispectral remote sensing image products such as Sentinel, and digital elevation model (DEM) image products, but are not limited to these. Forests to be assessed mainly include large areas of urban woodlands such as national forest parks / gardens within urban areas.

[0052] S2. Remote sensing image preprocessing: preprocessing operations are performed on the collected remote sensing images of the forest to be estimated in the urban area, and multi-band remote sensing image data of the urban forest area to be estimated (i.e., forest remote sensing spectral data) are generated.

[0053] Based on the embodiment, the remote sensing image preprocessing is as follows: Figure 2 As shown, the specific implementation steps include:

[0054] Specifically, due to the characteristics of satellite remote sensing images, it is impossible to obtain remote sensing images of the area to be studied. Only large or multiple remote sensing images that include the study area can be obtained. Therefore, it is necessary to perform preprocessing operations on the collected remote sensing images of the forest to be estimated in the urban area, including radiation calibration, atmospheric correction, terrain correction, image stitching, masking, and cropping. After removing redundant images, the multi-band remote sensing images of the forest area to be estimated are finally obtained. The above preprocessing process is implemented using ENVI and ArcGIS software, but is not limited to this. The full name of ENVI in English is The Environment for Visualizing Images, and its Chinese translation is a complete remote sensing image processing platform. ArcGIS is a geographic information system software.

[0055] Radiation calibration: Eliminate the error of the sensor itself and determine the accurate radiation value at the sensor entrance.

[0056] Atmospheric correction: Eliminate the effects of light and factors such as water vapor, oxygen, and carbon dioxide in the atmosphere on the reflection of ground objects.

[0057] Terrain correction: Based on DEM elevation images, eliminate or mitigate radiation errors caused by terrain undulations.

[0058] Image stitching: stitching multiple remote sensing images covering the study area into a single image.

[0059] Image cropping and masking: After removing redundant images from a single overall image, a multi-band remote sensing image of a specific study area is obtained.

[0060] S3. Construction and selection of vegetation index: Based on the preprocessed multi-band remote sensing image data, the vegetation index in different bands is inverted, and the correlation coefficient method + machine learning fusion model is used to perform correlation and importance analysis on the vegetation index in different bands to obtain the vegetation index in the optimal band.

[0061] Specifically, based on multi-band remote sensing imagery data of the forest area to be assessed, image bands that can represent vegetation chlorophyll are selected. Based on the information from different bands, vegetation indices are constructed, including, but not limited to, the Normalized Difference Vegetation Index (NDVI), the Ratio Vegetation Index (RVI), and the Chlorophyll Index (CI). A correlation coefficient method combined with a machine learning fusion model is used to test the correlation between vegetation indices in different bands and the importance of vegetation indices in different bands and the target chlorophyll content. Ultimately, the vegetation index for the optimal band is obtained.

[0062] Image bands that can characterize vegetation chlorophyll were selected, including blue, green, red, red-edge, and near-infrared bands. The correlation coefficient method (Pearson correlation coefficient, R) was used as the correlation test, and the interpretable XGBOOST machine learning model was used to test importance. The target chlorophyll content was obtained by selecting a preset number of fixed-size plots within the urban forest area to be assessed. The plots should be evenly distributed throughout the urban forest area, and the trees within the plots should be representative of the average growth of trees within the urban forest area to be assessed. Several healthy, intact, and regular leaves were selected within the plots, and their chlorophyll content was measured using a UV spectrophotometer.

[0063] Based on the embodiment, the construction and selection of the vegetation index are as follows: Figure 3 As shown, the specific implementation steps include:

[0064] Based on multi-band remote sensing imagery at corresponding locations within each plot of the forest area to be assessed, the S31 selects image bands that characterize vegetation chlorophyll, including blue, green, red, red-edge, and near-infrared bands. The red-edge band is the most prominent feature of the green vegetation spectral curve and is closely correlated with various physical and chemical parameters of vegetation. It is an important indicator of plant chlorophyll content and health. Chlorophyll has strong absorption peaks in the blue and red bands and a valley in the green band, but minimal absorption in the near-infrared band, so it is not considered a primary band here. Because different satellites carry different sensors, different remote sensing image products provide different numbers and types of bands. Therefore, when selecting bands, select image bands that characterize vegetation chlorophyll based on the specific remote sensing image product.

[0065] S32 constructs vegetation indices based on information from different bands, including, but not limited to, the Normalized Difference Vegetation Index (NDVI), the Ratio Vegetation Index (RVI), and the Chlorophyll Index (CI). Vegetation indices are combinations of related spectral signals, formed by combining multispectral or hyperspectral remote sensing data from two or more bands through certain mathematical transformations. The construction process of vegetation indices based on different bands is shown in Table 1, using Sentinel-2A remote sensing imagery as an example:

[0066] Table 1 Construction process of vegetation index in different bands

[0067] Vegetation Index Calculation formula Vegetation Index Calculation formula Vegetation Index Calculation formula <![CDATA[NDVI 红 ]]> <![CDATA[(ρ8-ρ4) / (ρ8+ρ4)]]> <![CDATA[RVI 红 ]]> <![CDATA[ρ8 / ρ4]]> <![CDATA[CI 红 ]]> <![CDATA[ρ8 / ρ4-1]]> <![CDATA[NDVI 蓝 ]]> <![CDATA[(ρ3-ρ2) / (ρ3+ρ2)]]> <![CDATA[RVI 蓝 ]]> <![CDATA[ρ3 / ρ2]]> <![CDATA[CI 蓝 ]]> <![CDATA[ρ8 / ρ2-1]]> <![CDATA[NDVI 绿 ]]> <![CDATA[(ρ3-ρ4) / (ρ3+ρ4)]]> <![CDATA[RVI 绿 ]]> <![CDATA[ρ3 / ρ4]]> <![CDATA[CI 绿 ]]> <![CDATA[ρ8 / ρ3-1]]> <![CDATA[NDVI 红边 ]]> <![CDATA[(ρ 8A -ρ7) / (ρ 8A +p7)]]> <![CDATA[RVI 红边 ]]> <![CDATA[ρ 8A / p7]]> <![CDATA[CI 红边 ]]> <![CDATA[ρ 8A / p7-1]]>

[0068] Among them, ρ2, ρ3, ρ4, ρ7, ρ8, ρ 8A They represent the surface reflectance of blue band-2, green band-3, red band-4, red edge band-7, and near-infrared band-8 / 8A respectively.

[0069] The high correlation between S33 spectral bands and the addition of bands that are insensitive or important to chlorophyll content will reduce the accuracy of the model in estimating chlorophyll content. Therefore, the present invention performs a correlation test + an importance test to select the vegetation index under the optimal band. The specific steps include:

[0070] Correlation test analysis: The correlation coefficient method (Pearson correlation coefficient, R) was used to test the correlation between vegetation indices in different bands. The formula is as follows:

[0071]

[0072] Among them, R ij is the correlation coefficient of vegetation index in bands i and j; is the covariance of vegetation index in bands i and j; σ ii and σ jjare the standard deviations of vegetation indices in bands i and j, respectively.

[0073] Importance Detection Analysis: The interpretable XGBOOST algorithm was used to test the importance of vegetation index and chlorophyll content target values ​​in different bands of urban forest remote sensing images. XGBOOST is a gradient boosting algorithm that assigns corresponding weights to different feature variables according to their importance scores and brings them into the model regressor. The parameter β of the next decision tree of the gradient boosting algorithm is determined by minimizing the loss function. k , the formula is as follows:

[0074]

[0075] Among them, β k is the parameter of the base learner of the characteristic variable of the k-th urban forest chlorophyll content; L(y m ,f k (x m )) is the loss function of the boosting tree algorithm for the characteristic variable of urban forest chlorophyll content; y m is the true value of chlorophyll content in urban forests; f k (x m ) is the predicted value of chlorophyll content in urban forests; x m is the characteristic variable, i.e., the vegetation index in different bands; M is the number of quadrats (samples) of urban forests; and m is the training quadrats.

[0076] Currently, there are two main aspects of calculating feature importance: one is to calculate feature importance during training; the other is to use OOB (Out of Bag) data to calculate feature importance after training. This invention calculates feature importance during training. The specific idea is: the more times a feature is selected as a split point, the more important the feature is. Therefore, during training, the total number of feature splits is recorded, and the number of times a feature appears as a split point / the total number of splits is used as the feature importance for quantification.

[0077] Finally, considering the results of correlation test and importance test comprehensively, characteristic variables with low correlation between spectral bands and high importance to chlorophyll content were selected as vegetation indices under the optimal band.

[0078] S4. Establishment of chlorophyll content estimation model: A chlorophyll content estimation model for urban forests was constructed based on the vegetation index under the optimal band combined with a machine learning method (particle swarm optimization least squares support vector machine, PSO-LSSVM).

[0079] Specifically, the vegetation index under the optimal band of each sample plot in the urban forest is used as the characteristic variable for model input, and the target value of the chlorophyll content of each sample plot is used as the model output. Combined with the particle swarm optimization least squares support vector machine (PSO-LSSVM) machine learning model, after model training and parameter optimization and adjustment, a chlorophyll content estimation model for urban forests is finally constructed.

[0080] Based on the embodiment, the chlorophyll content estimation model is established, such as Figure 4 As shown, the specific implementation steps include:

[0081] The vegetation index under the optimal band of each sample plot of S41 urban forest and the target value of chlorophyll content constitute the data set for model construction, among which the vegetation index under the optimal band is used as the characteristic variable for model input, and the target value of chlorophyll content is used as the model output. At the same time, 80% of the sample data samples are used as training sets and 20% of the sample data are used as test sets.

[0082] S42 builds an urban forest chlorophyll content estimation model based on the dataset samples and combined with the PSO-LSSVM model. The specific details are as follows:

[0083] LSSVM is a nonlinear machine learning method that improves and expands on the theory of structural risk minimization of support vector machines (SVM). LSSVM integrates the least squares method based on the SVM theory, replaces the insensitive loss function in SVM with a quadratic loss function, and changes the inequality constraints to equality constraints. The optimization problem then becomes:

[0084]

[0085] Among them, c is the regularization parameter, which is used to adjust the penalty intensity of the error between the predicted value of chlorophyll content obtained by the vegetation index under the optimal band of urban forest in the training sample and the target value of chlorophyll content. The larger the error, the greater the penalty intensity for the model. Through this mechanism, the model error is continuously optimized and the model accuracy is improved; b is the bias; ω is the weight vector; ξ m is the error vector; J(ω,ξ) is the loss function; is a nonlinear mapping function; m is a training quadrat; M is the number of urban forest quadrat; x m is the characteristic variable, that is, the vegetation index under the optimal band in the training sample; y m is the target value of chlorophyll content; st is the abbreviation of constraint condition.

[0086] In order to solve the above optimization problem, the quadratic programming problem can be transformed into a problem of solving a system of linear equations. In this process, the LSSVM nonlinear regression model can be obtained:

[0087]

[0088] Among them, K(x,x k ) is the radial basis kernel function; x is the characteristic variable of the input model; x k is the center of the radial basis sum function; σ is the radial basis kernel parameter. y(x) is the predicted value of chlorophyll content.

[0089] The regularization parameter c and the radial basis kernel parameter σ are important factors affecting the accuracy and generalization ability of the LSSVM model. Therefore, optimizing c and σ becomes the primary task of building an optimal LSSVM model. Therefore, the present invention uses the particle swarm optimization (PSO) to optimize the model parameters. PSO is a global optimization algorithm that uses a group of parallel search engines to search for the optimal solution of multiple parameter combinations through repeated iterative recursion. It has the advantages of fewer parameters and faster convergence.

[0090] Based on the PSO-LSSVM model, combined with the training sample set, the model is trained and the parameters are optimized. In order to test the stability and accuracy of the model, the coefficient of determination (R 2 ) and root mean square error (RMSE) to evaluate the model training results, R 2 The larger the value, the higher the model accuracy, and the smaller the RMSE, the smaller the model error. The formula is as follows:

[0091]

[0092]

[0093] Among them, y m is the target value of chlorophyll content; is the model-predicted value of chlorophyll content; is the average value of the target chlorophyll content; M is the number of quadrats in the training set.

[0094] After training and parameter optimization, the optimal chlorophyll content estimation model was obtained, and then the model was verified using the test set samples. The model verification results were still evaluated using formulas (5) and (6).

[0095] Based on the optimal chlorophyll content estimation model obtained above, the chlorophyll content of the entire forest area to be estimated within the urban area is estimated.

[0096] S5. Obtaining vegetation stand factors / meteorological factors: In the urban forest area to be assessed, select a preset number of fixed-size sample plots, obtain the vegetation stand factors in each sample plot through field measurements, and calculate the forest biomass in the sample plot based on the stand factors; obtain the meteorological factors in each sample plot in the forest area to be assessed based on relevant sensor technology.

[0097] Specifically, a preset number of fixed-size plots are selected in the urban forest area to be assessed (Note: the plots obtained from the stand factors / meteorological factors are consistent with the plots obtained from the target chlorophyll content values). The stand factors such as tree height, diameter at breast height (DBH), forest species, and vegetation density of all trees in the plot are measured using the tree ruler method. The DBH is the diameter of the tree trunk at 1.3 m from the ground surface. Trees less than 1.3 m in height are replaced by the base diameter. The average stand height and average DBH of the plots are calculated. The average stand height of the plots is the average of the tree heights of all trees in the plots, and the average DBH of the plots is the average of the DBH of all trees in the plots. The forest biomass in the plots is calculated based on the stand factors. The meteorological factors in each plot within the forest area to be assessed are obtained based on relevant meteorological environment sensors, including temperature, relative humidity, precipitation, etc.

[0098] Based on the embodiment, the vegetation stand factor / meteorological factor is obtained, such as Figure 5 shown.

[0099] Specifically, in the measurement of urban forest carbon sinks, considering only remote sensing image data (horizontal dimension) is not enough to provide a more comprehensive and accurate analysis of the measurement results. In view of this, the present invention also incorporates stand factor / meteorological factor data (vertical dimension) that can affect forest carbon sink capacity into the urban forest measurement, achieving the accuracy and comprehensiveness of urban forest carbon sink measurement from a multi-dimensional perspective. Specifically, it includes:

[0100] In the urban forest area to be assessed, a preset number of fixed-size plots are selected, and the tree height, diameter at breast height (DBH), forest species, vegetation density and other stand factors of all trees in the plot are measured using the tree ruler method. The DBH is the diameter of the tree trunk at 1.3m from the ground surface. If the tree height is less than 1.3m, the base diameter is used instead. The average stand height and average DBH of the plot are calculated. The average stand height of the plot is the average of the tree heights of all trees in the plot, and the average DBH of the plot is the average of the DBH of all trees in the plot. Based on the average stand height and average DBH, the forest biomass in the plot is calculated using the binary allometric growth model of standing trees. Based on relevant meteorological environment sensors, the meteorological factors in each plot of the forest range to be assessed, including temperature, relative humidity, precipitation, etc., are obtained.

[0101] S6 Construction of urban forest carbon sink estimation model: Based on the chlorophyll content of the forest to be estimated in the urban area and vegetation stand factors / meteorological factors, combined with machine learning methods (deep belief network (DBN)), a forest carbon sink estimation model to be estimated in the urban area is constructed.

[0102] Specifically, based on the chlorophyll content of each sample plot of the forest to be estimated in the urban area and the fusion data of vegetation stand factors / meteorological factors, the correlation coefficient method + machine learning fusion model in step 3 is used to test the correlation between the chlorophyll content and vegetation stand factors / meteorological factors in the sample plot, and the importance test between the chlorophyll content and vegetation stand factors / meteorological factors and the forest biomass in the sample plot is performed. The optimal characteristic factors are selected and combined with the deep belief network (DBN) learning model. After model training and parameter optimization and adjustment, the urban forest biomass is calculated. Finally, the urban forest carbon stock is obtained by combining the biomass-carbon stock conversion factor. Then, the urban forest carbon stock in the preset baseline period is subtracted to obtain the forest carbon sink to be estimated in the urban area.

[0103] Based on the embodiment, the construction of the urban forest carbon sink calculation model is as follows: Figure 6 The specific steps are as follows:

[0104] S61 constructs a data set by taking the chlorophyll content of each sample plot in the urban area and the stand factor / meteorological factor of the vegetation as characteristic variables and the forest biomass in each sample plot as the target value.

[0105] S62 is based on the correlation test (R) + importance test (XGBOOST) method process described in S33. The implementation process of the method will not be described in detail here. A correlation test analysis is performed between the chlorophyll content and the vegetation stand factor / meteorological factor in each sample plot, and an importance test analysis is performed between the chlorophyll content and the vegetation stand factor / meteorological factor in each sample plot and the forest biomass target value in the sample plot to select the optimal characteristic variable.

[0106] S63 establishes a forest biomass estimation model based on the optimal characteristic variables within each quadrat extracted in S62 and the forest biomass target value within the quadrat, combined with the deep belief network (DBN) learning method. Specifically, it includes:

[0107] The DBN machine learning model is a generative deep learning model based on a stochastic restricted Boltzmann machine (RBM). It is composed of several RBMs stacked from bottom to top. It has strong feature analysis capabilities, a simple model structure, fast convergence, and strong generalization capabilities.

[0108] RBM is a probabilistic modeling method based on energy function. Its structure consists of visible layer and hidden layer, which can generate joint distribution P(V,h1,h2,....,h n ), the formula is as follows:

[0109]

[0110] Among them, P(h i |hi+1 ) represents the conditional distribution between hidden layers; V is the visible layer; h is the hidden layer; n is the number of hidden layers. i represents the i-th hidden layer.

[0111] The visible layer and hidden layer of each RBM in the DBN model are fully connected through weights and biases. The weights and biases between the visible layer and hidden layer output by the current RBM will become the input of the next RBM. The DBN model uses the contrastive divergence algorithm to train and update the weights and biases. The steps are as follows:

[0112] Hidden layer initialization:

[0113] h0=f(v0·w T +b0) (8)

[0114] Where w=[w0,w1,w2,....,w n ] is the initial weight matrix; b0 is the initial bias vector; f() is the activation function; v0 is the first visible layer, i.e., the initial visible layer; h0 is the initial hidden layer. T represents transpose.

[0115] Reconstruct subsequent visible layers With hidden layer

[0116]

[0117]

[0118] Where B=[b1,b2,b3,...,b n ] is the new bias vector in the reconstruction process. After reconstruction, the new visible layers v1, v2, v3..., v n And the hidden layers h1,h2,h3,...,h n ;.

[0119] Calculate the weight difference ΔW:

[0120]

[0121] Where ζ is the amount of data input each time. The above process is repeated until convergence, that is, the weight difference is minimized.

[0122] The present invention constructs a DBN model dataset based on the optimal characteristic variables of each urban forest plot to be estimated and the target forest biomass values ​​of the plots, obtained in S62. 80% of the plot data is used as a training set, and 20% of the plot data is used as a test set. The optimal characteristic variables of the plots are used as model input, and the target forest biomass values ​​of the plots are used as model output. The DBN learning process is divided into two stages: bottom-up unsupervised pre-training and top-down supervised tuning. With the goal of minimizing the difference between the predicted values ​​and the target values ​​of the urban forest biomass plot model, the network connection parameters are trained and optimized through pre-training and tuning, ultimately obtaining an optimal forest biomass estimation model.

[0123] During the training process, the difference between the model predicted value and the target value of the sample quadrat forest biomass was evaluated using formulas (5) and (6).

[0124] The optimal forest biomass estimation model obtained through training was tested using the test set samples, and the test results were evaluated using formulas (5) and (6).

[0125] S64 is based on the optimal forest biomass estimation model described in S63 and is combined with the optimal characteristic variables of the entire range of the forest to be estimated in the urban area (i.e., chlorophyll, vegetation stand factor\meteorological factor of the entire region. The acquisition of characteristic variables of the entire region is based on the sample data acquisition method, and the remote sensing image and the measured range cover the entire urban forest land). The forest biomass of the entire area of ​​the forest area to be estimated is calculated, and the forest carbon stock is calculated through the biomass-carbon stock conversion factor (i.e., the tree species carbon content coefficient), and the forest carbon stock of the entire area of ​​the forest area to be estimated in the urban area is generated.

[0126] The S65 forest carbon sink is the change in forest carbon storage over a period of time. Therefore, when calculating the urban forest carbon sink, it is necessary to first preset a certain period as the base period. Then, combined with the optimal urban forest biomass calculation method described above, the forest carbon storage in the base period is calculated. After the preset time interval to be estimated (year, season, month, day, etc.), the forest carbon storage after the corresponding interval is calculated according to the optimal urban forest biomass calculation method. Finally, the forest carbon storage after the preset time interval minus the forest carbon storage in the base period is the overall carbon sink of the forest to be estimated in the urban area.

[0127] The present invention relates to the technical field of remote sensing information extraction and urban forest carbon sink estimation, comprising the following steps: S1, remote sensing image acquisition; S2, remote sensing image preprocessing; S3, vegetation index construction and selection; S4, establishment of a chlorophyll content estimation model; S5, acquisition of vegetation stand factors and meteorological factors; and S6, construction of an urban forest carbon sink estimation model. The present invention utilizes remote sensing imagery and machine learning methods to achieve small-scale urban forest carbon sink estimation. Chlorophyll, which reflects the carbon sequestration capacity of vegetation through photosynthesis, is acquired using remote sensing imagery. Combined with vegetation stand factors and meteorological factors, machine learning methods are used to intelligently estimate small-scale urban forest carbon sinks in both horizontal and vertical dimensions. The present invention utilizes vegetation chlorophyll information derived from remote sensing imagery to more accurately capture vegetation carbon sequestration capacity. It also achieves data fusion in both horizontal and vertical dimensions, avoiding the drawback of large model estimation errors caused by insufficient characteristic factors. Combined with machine learning methods, it enables efficient and comprehensive analysis of multi-dimensional, nonlinear data. Consequently, the present invention boasts high accuracy, strong intelligence, and broad applicability.

[0128] Chlorophyll is the primary carrier of photosynthesis in green vegetation, while carbon sequestration primarily reflects the ability of green vegetation to sequester carbon through photosynthesis. Estimating chlorophyll content in vegetation provides a more intuitive and accurate representation of its carbon sequestration capacity. Therefore, forest chlorophyll content derived from remote sensing imagery holds significant research value in forest carbon sequestration estimation. Furthermore, most studies on forest carbon sequestration have focused solely on remote sensing imagery. However, remote sensing imagery, acquired from high altitudes, represents horizontal data. Vertical data such as tree height, diameter at breast height (DBH), temperature, and humidity also significantly influence forest carbon sequestration. Therefore, integrating these two dimensions offers novel research implications. Furthermore, with the advancement of machine learning technology, its application in intelligent processing and analysis of high-dimensional and nonlinear data is expanding. Therefore, developing a highly accurate, intelligent, and applicable small-scale urban forest carbon sequestration estimation method based on remote sensing imagery and machine learning techniques holds significant practical value and significance.

[0129] The present invention also provides an urban forest carbon sink calculation system based on remote sensing images and machine learning. The urban forest carbon sink calculation system based on remote sensing images and machine learning can be implemented by executing the process steps of the urban forest carbon sink calculation method based on remote sensing images and machine learning. That is, those skilled in the art can understand the urban forest carbon sink calculation method based on remote sensing images and machine learning as an optimal implementation of the urban forest carbon sink calculation system based on remote sensing images and machine learning.

[0130] The system includes the following modules:

[0131] Remote sensing image acquisition module: collects remote sensing images of forests to be assessed within urban areas.

[0132] Remote sensing image preprocessing module: preprocesses the collected remote sensing images to form multi-band remote sensing image data.

[0133] Vegetation index selection module: Constructs vegetation indices in different bands based on multi-band remote sensing image data, uses the correlation coefficient method and machine learning fusion model to analyze the correlation and importance of vegetation indices in different bands, and obtains the vegetation index in the optimal band.

[0134] Chlorophyll content estimation module: Based on the vegetation index in the optimal band and combined with machine learning, a chlorophyll content estimation model for urban forests is constructed; chlorophyll content is estimated through the chlorophyll content estimation model.

[0135] Factor acquisition module: select sample plots in the urban forest area to be assessed and obtain vegetation stand factors and meteorological factors in each sample plot.

[0136] Carbon sink calculation module: Based on the estimated chlorophyll content, vegetation stand factors and meteorological factors, combined with the machine learning model, an urban forest carbon sink calculation model is constructed, and the urban forest carbon sink calculation model is used to calculate urban forest carbon sinks.

[0137] A computer-readable storage medium stores a computer program, which, when executed, implements the steps of a method for calculating urban forest carbon sinks based on remote sensing images and machine learning.

[0138] An electronic device includes a memory, one or more processors and a computer program. When the processor executes the computer program stored in the memory, the steps of a method for calculating urban forest carbon sinks based on remote sensing images and machine learning are implemented.

[0139] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0140] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A method for estimating urban forest carbon sequestration based on remote sensing images and machine learning, characterized in that: The steps include: Remote sensing image acquisition steps: Collect remote sensing images of the forest to be assessed within the urban area; Remote sensing image preprocessing step: preprocess the collected remote sensing images to form multi-band remote sensing image data; Vegetation index selection steps: Based on multi-band remote sensing image data, vegetation indices in different bands are constructed. The correlation coefficient method and machine learning fusion model are used to analyze the correlation and importance of vegetation indices in different bands to obtain the vegetation index in the optimal band. Chlorophyll content estimation steps: Based on the vegetation index in the optimal band and combined with machine learning methods, a chlorophyll content estimation model for urban forests is constructed; chlorophyll content is estimated using the chlorophyll content estimation model; Factor acquisition steps: select sample plots in the urban forest area to be assessed, and obtain the vegetation stand factors and meteorological factors in each sample plot; Carbon sink calculation steps: Based on the estimated chlorophyll content, vegetation stand factors and meteorological factors, combined with the machine learning model, an urban forest carbon sink calculation model is constructed, and the urban forest carbon sink calculation model is used to calculate urban forest carbon sinks.

2. The urban forest carbon sink estimation method based on remote sensing imagery and machine learning according to claim 1 is characterized in that: In the remote sensing image acquisition step, sources of remote sensing images of the forests to be estimated in the urban area include satellite digital products, multispectral remote sensing image products, and elevation image products in the data platform; the forests to be estimated include urban woodlands in the urban area.

3. The urban forest carbon sink calculation method based on remote sensing imagery and machine learning according to claim 1 is characterized in that: In the remote sensing image preprocessing step, the preprocessing operations performed on the collected remote sensing images include radiometric calibration, atmospheric correction, terrain correction, image stitching, masking, and cropping; After removing redundant images, multi-band remote sensing image data of the forest area to be assessed are obtained.

4. The urban forest carbon sink calculation method based on remote sensing imagery and machine learning according to claim 1 is characterized in that: In the vegetation index selection step, based on the multi-band remote sensing image data, the image band representing the vegetation chlorophyll is selected, and the vegetation index is constructed according to the different band information; The constructed vegetation indices include normalized difference vegetation index, ratio vegetation index and chlorophyll index; Select a preset number of fixed-size plots within the urban forest area to be assessed. The plots are evenly distributed throughout the urban forest area, and the trees within the plots represent the average growth conditions of trees within the urban forest area to be assessed. Select a number of leaves within the plots and measure the target chlorophyll content using an ultraviolet spectrophotometer. The correlation coefficient method and machine learning model were used to conduct correlation analysis between vegetation indices in different bands, and the importance analysis was conducted between vegetation indices in different bands and chlorophyll content target values ​​to obtain the vegetation index in the optimal band.

5. The urban forest carbon sink estimation method based on remote sensing imagery and machine learning according to claim 1 is characterized in that: In the chlorophyll content estimation step, a data set is constructed based on the optimal vegetation index and chlorophyll content target value of each sample plot in the urban forest; the optimal vegetation index of each sample plot in the urban forest is used as a characteristic variable for model input, and the chlorophyll content target value of each sample plot is used as the model output. Based on the data set combined with the particle swarm optimization least squares support vector machine machine learning model, after model training and parameter optimization and adjustment, a chlorophyll content estimation model for the urban forest is finally constructed.

6. The urban forest carbon sink estimation method based on remote sensing imagery and machine learning according to claim 1 is characterized in that: In the factor acquisition step, a preset number of fixed-size plots are selected within the urban forest area to be assessed, the stand factors of all trees in the plots are measured using the tree-per-plot method, and the forest biomass in the plots is calculated based on the stand factors; and the meteorological factors in each plot within the forest area to be assessed are obtained based on relevant meteorological environment sensors.

7. The urban forest carbon sink calculation method based on remote sensing imagery and machine learning according to claim 1 is characterized in that: In the carbon sink calculation step, an urban forest carbon sink calculation model is constructed based on the chlorophyll content, vegetation stand factors and meteorological factors of the forest to be estimated in the urban area, combined with a deep belief network.

8. An urban forest carbon sink calculation system based on remote sensing images and machine learning, characterized by: Includes the following modules: Remote sensing image acquisition module: collects remote sensing images of forests to be assessed within urban areas; Remote sensing image preprocessing module: preprocesses the collected remote sensing images to form multi-band remote sensing image data; Vegetation index selection module: Constructs vegetation indices in different bands based on multi-band remote sensing image data, uses the correlation coefficient method and machine learning fusion model to analyze the correlation and importance of vegetation indices in different bands, and obtains the vegetation index in the optimal band; Chlorophyll content estimation module: Based on the vegetation index in the optimal band and combined with machine learning, a chlorophyll content estimation model for urban forests is constructed; chlorophyll content is estimated using the chlorophyll content estimation model; Factor acquisition module: select sample plots in the urban forest area to be assessed and obtain vegetation stand factors and meteorological factors in each sample plot; Carbon sink calculation module: Based on the estimated chlorophyll content, vegetation stand factors and meteorological factors, combined with the machine learning model, an urban forest carbon sink calculation model is constructed, and the urban forest carbon sink calculation model is used to calculate urban forest carbon sinks.

9. The urban forest carbon sink calculation system based on remote sensing images and machine learning according to claim 8 is characterized in that: In the remote sensing image acquisition module, the sources of remote sensing images of the forests to be estimated in the urban area include satellite digital products, multispectral remote sensing image products and elevation image products in the data platform; the forests to be estimated include urban woodlands in the urban area.

10. The urban forest carbon sink calculation system based on remote sensing images and machine learning according to claim 8 is characterized in that: In the remote sensing image preprocessing module, the preprocessing operations performed on the collected remote sensing images include radiometric calibration, atmospheric correction, terrain correction, image stitching, masking, and cropping; After removing redundant images, multi-band remote sensing image data of the forest area to be assessed are obtained.

Images