A method and apparatus for estimating aboveground carbon storage of a stand

Abstract

The embodiment of the present application provides a method and device for estimating aboveground carbon storage of forest stand, which can realize the estimation of aboveground carbon storage of forest stand. The method comprises the following steps: obtaining an image on the ground in the forest; obtaining a carbon storage parameter from the image on the ground in the forest; and estimating the aboveground carbon storage of the forest stand by using a carbon storage model according to the carbon storage parameter. The carbon storage parameter is the percentage x of the tree foreground image in the whole image. According to the present application, the estimation of the aboveground carbon storage of the forest stand can be realized by using a common camera and a computer. Therefore, the present application can estimate the aboveground carbon storage of the forest stand without adding extra equipment. The present application technology and its application in the dynamic monitoring of the aboveground carbon storage of the forest stand in the forest field, long-term fixed sample field and other scenes.

Description

Technical Field

[0001] This invention relates to a technique for estimating aboveground carbon storage in forest stands, and more particularly to a method and apparatus for estimating aboveground carbon storage in forest stands. Background Technology

[0002] Forests are the main body of terrestrial ecosystems and the largest carbon sink within them. Their role in the global carbon cycle is crucial because carbon flows between forests and the atmosphere through natural exchanges such as photosynthesis, respiration, decomposition, and combustion. Understanding aboveground carbon storage not only enhances our understanding of the role of forests in the carbon cycle but also has significant implications for the sustainable management of existing forests. Therefore, it is essential to employ an accurate and reasonable method for a precise spatial estimation of carbon storage in forests.

[0003] The most accurate way to estimate carbon storage is to fell trees and weigh them by dry weight, but this method is not only time-consuming and labor-intensive, but also damages the ecological environment. Methods based on satellite remote sensing and other means also have problems such as the need for ground verification, expensive equipment, and high technical difficulty. Summary of the Invention

[0004] The embodiments of the present invention provide a method and apparatus for estimating aboveground carbon storage in forest stands, which can realize the estimation of aboveground carbon storage in forest stands.

[0005] The methods for estimating aboveground carbon storage in forest stands include:

[0006] Acquire images of the forest floor;

[0007] Carbon storage parameters were obtained from images of the forest floor.

[0008] The aboveground carbon storage of forest stands is estimated using a carbon storage model based on carbon storage parameters;

[0009] The carbon storage parameter is: percentage of the foreground image of trees in the entire image x;

[0010] The carbon storage model is any one of the following three equations:

[0011] y = a0 + (a1 + u)x (4)

[0012]

[0013]

[0014] in:

[0015] y: Aboveground carbon storage of the forest stand;

[0016] x: Percentage of the foreground image of trees in the entire image;

[0017] a0, a1: Parameters to be determined;

[0018] u: Random effects parameter;

[0019] The parameters to be determined were obtained through experiments.

[0020] The acquisition of images of the forest floor specifically includes:

[0021] Establish rectangular sample plots;

[0022] Set up shooting locations;

[0023] Take images at each shooting point.

[0024] The specific methods for obtaining carbon storage parameters from images of the forest floor include:

[0025] Step 51: Load the acquired image into a fully convolutional neural network model;

[0026] Step 52: The fully convolutional neural network model traverses the image pixel by pixel, starting from the beginning position of the image;

[0027] Step 53: Calculate the feature value of the pixel according to the weight file, and compare the feature value with the learned target range. If it is within the target value range, the current pixel value is determined as the foreground; otherwise, the pixel value is set as the background.

[0028] Step 54: Traverse all pixels in order from top to bottom and from left to right, and determine if it is the last pixel. If it is, proceed to step 55; otherwise, return to step 53.

[0029] Step 55: Extract the foreground portion and calculate the carbon storage parameter using the foreground portion: The carbon storage parameter is the percentage of the forest foreground image in the entire image (x).

[0030] The method further includes:

[0031] The specific steps to obtain the weight file are as follows:

[0032] Construct training image samples;

[0033] The image samples are input into a fully convolutional network model, which consists of 13 convolutional layers and 5 downsampling layers. Downsampling uses a 2×2 max pooling method with a stride of 2. Non-linear activation is performed using rectified linear unit functions. All convolutional kernels are 3×3 in size with a stride of 2 and padding of 1.

[0034] The fully convolutional network model is trained so that the input image samples are identified pixel by pixel by the network structure. Through the label samples drawn from a large number of image samples, the network model gradually learns that in the same type of image, pixels with a gray value within a certain range are target pixels, while pixels in other ranges are non-target pixels. Then, the range enclosed by the edges of all target pixels or the area of ​​the target pixels is the segmented image. The criteria for these target pixels are the weight file generated by training the network structure.

[0035] The method further includes a step of obtaining parameters to be determined, specifically including:

[0036] A certain number of forest stand aboveground images and corresponding measured values ​​of aboveground carbon storage were obtained. The images were segmented to obtain the carbon storage parameters of the corresponding forest stands, and the maximum likelihood estimation method was used to estimate the undetermined parameters.

[0037] The method further includes:

[0038] The steps for setting the random effects parameter are as follows: the sample plots are divided into two levels based on their actual altitude, namely u<750m and u≥750m. The random effects for u<750m are set to 0, and the random effects for u≥750m are set to 1.

[0039] This invention also provides a device for determining the aboveground carbon storage of a forest stand, the device comprising:

[0040] Image acquisition device, used to acquire images of the forest stand ground;

[0041] The computing unit is used to obtain carbon storage parameters from color images of forest stands;

[0042] The unit is defined to estimate the aboveground carbon storage of a forest stand using a carbon storage model based on carbon storage parameters.

[0043] According to the present invention, the aboveground carbon storage of a forest stand can be estimated using a common camera and a computer. Therefore, the present invention can estimate the aboveground carbon storage of a forest stand without the need for additional equipment. The technology of the present invention is applicable to the dynamic monitoring of aboveground carbon storage in forest farms, long-term fixed sample plots, and other similar scenarios. Attached Figure Description

[0044] Figure 1 This invention illustrates a method for estimating aboveground carbon storage in forest stands according to an embodiment of the present invention;

[0045] Figure 2 This invention illustrates a method for acquiring images of the forest floor according to an embodiment of the present invention;

[0046] Figure 3 This invention illustrates one method for setting up shooting points according to an embodiment of the invention;

[0047] Figure 4 This invention illustrates a shooting angle at the shooting point according to an embodiment of the invention;

[0048] Figure 5 This invention illustrates a method for obtaining carbon storage parameters from images of the forest floor according to an embodiment of the present invention;

[0049] Figure 6 An artificial neural network architecture is shown;

[0050] Figure 7 This is an embodiment of the present invention for determining the aboveground carbon storage of a forest stand. Detailed Implementation

[0051] To facilitate understanding and implementation of the present invention by those skilled in the art, embodiments of the present invention are now described in conjunction with the accompanying drawings.

[0052] Example 1

[0053] Since images of the forest floor reveal its basic characteristics, they provide a basis for obtaining stand factors without damaging the forest habitat. The simplification and affordability of digital image acquisition, along with the widespread adoption of deep learning frameworks, have objectively promoted image-based stand factor research and hold promise for automating and accelerating the estimation of aboveground carbon storage. In the study of aboveground carbon storage, given that pines account for approximately 65% ​​of the forest area in northern China and dominate the forest ecosystem, the aboveground carbon storage of a pure stand of Dahurian larch (Larix gmelinii (Rupr.) Kuzen) in the Greater Khingan Mountains is investigated as an example.

[0054] Research on aboveground carbon storage in Dahurian larch forests mainly focuses on two aspects: image segmentation of the forest floor and carbon storage model construction. Image segmentation can be performed using deep learning methods. With the significant improvement in computing power, deep learning frameworks have developed considerably. Given a certain amount of labeled samples, deep learning can effectively segment target images. This application uses a fully convolutional network (FCNN) model for aboveground image segmentation. Most carbon storage model construction research focuses on least squares models, but this can lead to biases in parameter and variance estimation. Furthermore, carbon storage estimation may suffer from spatial autocorrelation issues. Using a mixed-effects model can address these problems. Therefore, constructing a mixed-effects model for aboveground carbon storage using the aboveground image segmentation results is appropriate.

[0055] like Figure 1 As shown in the figure, this embodiment provides a method for estimating aboveground carbon storage in forest stands, which includes the following steps.

[0056] Step 1: Obtain an image of the forest floor. This image can be a color image or a grayscale image. For example... Figure 2-4 As shown, the steps for acquiring an image of the forest floor are described below:

[0057] Step 11: Set up rectangular sample plots. For large forest stands, it is necessary to set up sample plots, such as rectangular sample plots, so as to set up shooting points within the forest stand and obtain more comprehensive images of the trees in the forest.

[0058] Step 12: Set up shooting points. Shooting points can be set up using any method, such as the traditional sample plot survey method. Figure 2 This paper illustrates a method for setting up shooting points, showing the path points for image capture. The path for image capture is a winding shape, and multiple shooting points are set along this path.

[0059] Step 13: Take images at each shooting point. Figure 3 The shooting angles at each shooting point are shown. Specifically, images can be taken in eight directions from each shooting point. To obtain more comprehensive forest stand image information, the images should be of the area of ​​interest. Specifically, any camera device can be used to capture images of the forest floor, such as a handheld camera or mobile phone. Observe the viewfinder or LCD screen, move the camera or adjust the focus, and ensure that the ray passing through the center of the image plane intersects the tree trunk at half its length, guaranteeing that the entire target forest stand area appears in the image.

[0060] Step 2: Obtain carbon storage parameters from images of the forest floor, such as... Figure 5 As shown, the specific method is as follows:

[0061] Step 51: Load the acquired image into the FCN network model. The deep learning architecture used in this invention is a Fully Convolutional Network (FCN) model, which is an end-to-end semantic segmentation model. In this embodiment, FCN is preferably used for image segmentation.

[0062] FCN is a deep learning architecture. Deep learning generally refers to classifying and regressing unknown data by training multi-layered network structures. These multi-layered network structures typically refer to Artificial Neural Networks (ANNs). An ANN is an algorithmic mathematical model that mimics the behavioral characteristics of animal neural networks, performing distributed parallel information processing. Figure 5As shown. This type of network relies on the complexity of the system to adjust the interconnections between a large number of internal nodes, thereby achieving the purpose of information processing, and possesses self-learning and adaptive capabilities. In recent years, with the significant improvement in computer computing power, artificial neural networks have gradually become widespread, and deep learning network models based on this have also made great strides. A major application area of ​​deep learning includes image classification, that is, classifying or recognizing entire images. The classification method first involves manually segmenting a portion of the image to be classified to construct a training set; then, image features, such as RGB spectral values, are learned based on the manually segmented image. When processing the unsegmented entire image, automatic segmentation can be performed based on the learned features. The deep learning model used in this invention generates a weight file as a feature weight file after training and learning. Using this file, image segmentation can be achieved in forest stands of the same type and environment without the need for repeated model training.

[0063] The fully convolutional neural network model used in this invention differs from an ANN (Advanced Neural Network) in that the fully connected layers at the ends of the ANN are replaced with convolutional layers, allowing the convolutional network to slide across a larger input image. The output will capture a heatmap of the target image, rather than a category. Simultaneously, an upsampling method is used to restore the image size, solving the problem of image size reduction caused by convolution and pooling. However, the initial FCN model suffered from slow training speed. To improve the model's convergence speed and make the hidden layer output feature distribution more stable, this invention adds a batch normalization layer to the input part of the hidden layers in the neural network.

[0064] Step 52: The FCN network model traverses the image pixel by pixel, starting from the beginning of the image.

[0065] Step 53: Calculate the feature values ​​of the image pixels. According to the weight file, compare the feature values ​​with the target range of the learned feature values. If the value is within the target range, the current pixel is determined to be a target pixel. Otherwise, the pixel is a non-target pixel. All target pixels form the segmented part, which is the foreground of the image. All non-target pixels form the unsegmented part, which is the background of the image.

[0066] The weight file serves as the basis for determining target pixels. To obtain the weight file, training image samples are first constructed, which is equivalent to creating labeled samples in deep learning. Then, a fully convolutional network model is constructed, consisting of 13 convolutional layers (COLs) and 5 downsampling layers. Downsampling uses a 2×2 max pooling method with a stride of 2. Non-linear activation is performed using rectified linear unit functions. All convolutional kernels are 3×3 in size, with a stride of 2 and padding of 1. Finally, the network is trained. In simple terms, the input image is identified pixel by pixel by the network structure. Through the labeled samples drawn from a large number of images, the fully convolutional network model gradually learns that pixels with gray values ​​within a certain range in images of the same type are target pixels, while pixels in other ranges are non-target pixels. The area enclosed by the edges of all target pixels, or the area of ​​the target pixels, is the segmented image. The basis for these target pixel discriminations is the weight file generated by training the network structure. Using this weight file, the same target can be segmented in images of the same type.

[0067] Step 54: Traverse all pixels in order from top to bottom and from left to right, and determine if it is the last pixel. If it is, proceed to step 55; otherwise, return to step 53.

[0068] Step 55: Extract the foreground portion and calculate the carbon storage parameter using the foreground portion: The carbon storage parameter is the percentage of the forest foreground image in the entire image (x).

[0069] Step 3: Estimate the aboveground carbon storage of the forest stand based on the carbon storage model and carbon storage parameters. There are three types of carbon storage models, which can be represented by the following three formulas:

[0070] y = a0 + (a1 + u)x (1)

[0071]

[0072]

[0073] in:

[0074] y: Aboveground carbon storage of forest stand

[0075] x: Percentage of the image occupied by the foreground of trees.

[0076] a0, a1: Parameters to be determined;

[0077] u: Random effects parameter;

[0078] According to the literature, there is a certain relationship between stand aboveground carbon storage and stand altitude. Therefore, this invention introduces a random effect parameter u when estimating stand aboveground carbon storage, dividing altitude into several levels to more accurately estimate stand aboveground carbon storage. The specific random effect parameter is determined according to different stand altitudes, as detailed below.

[0079] According to the growth patterns of trees, within a certain range of tree growth, diameter at breast height (DBH) and tree height are positively correlated. Therefore, the more mature the forest stand, the larger the percentage of the foreground trees occupying in the entire image. According to the carbon storage allometric growth equation, carbon storage is positively correlated with tree DBH. Therefore, the larger the percentage of the foreground trees occupying in the entire image, the higher the aboveground carbon storage of the forest stand. This is consistent with common knowledge in forestry.

[0080] The Pinaceae family has many tree species and is widely distributed in northern China. If the aboveground carbon storage of forest stands is estimated directly using the example model, it will produce a large error. Therefore, before using the present invention to estimate the aboveground carbon storage of forest stands, it is necessary to extract images and data that can represent the tree species to be used to confirm the calibration parameters a0, a1, and u, and then use the model with local parameters to estimate the aboveground carbon storage of forest stands over a large area.

[0081] Since there is a difference between the measured aboveground carbon storage and the model-predicted aboveground carbon storage, the relationship between the two can be determined by analyzing the correlation between the measured and model-predicted values ​​of aboveground carbon storage, thereby enabling a comparison of the model's fit to the data and a test of the model's applicability.

[0082] Let C i and These are the measured aboveground carbon storage and the predicted aboveground carbon storage of the forest stand in the i-th measurement (i = 1, ..., n), where n is the total number of observations. If it is the average of observed carbon storage values, then...

[0083]

[0084] The following section uses Dahurian larch as an example to calibrate the image estimation model for aboveground carbon storage in forest stands.

[0085] To determine the calibration parameters a0, a1, and u of the forest stand aboveground carbon storage estimation model, it is first necessary to obtain a certain number of images of the forest stand and the corresponding measured values ​​of aboveground carbon storage. After obtaining the carbon storage parameters of the corresponding forest stand by segmenting the images using the method of this invention, the model parameters can be estimated using the maximum likelihood estimation method, thus enabling the practical application of the model. This process of determining the model parameters is called calibration. During calibration, the number of sample plots is generally required to be ≥30. As an example, we randomly selected 32 sample plots and used... Figure 1 and Figure 2Images of the forest stand were acquired using a method that avoids destructive means. The camera used was a Canon EOS 700D, with an image resolution of 3456×5184, ISO 100, shutter speed of 1 / 60s, and lens aperture of F / 3.5. Since it was impossible to obtain measured values ​​of forest stand aboveground carbon storage through destructive methods, this embodiment adopted a commonly used remote sensing method for obtaining forest stand aboveground carbon storage values. Specifically, the aboveground carbon storage of a single tree was calculated based on the measured diameter at breast height (DBH) of individual trees using the allometric growth equation in the method for measuring the biomass and carbon content of Dahurian larch (DB23 / T 2650-2020) developed by the Heilongjiang Provincial Forestry and Grassland Bureau. The total aboveground carbon storage of the forest stand was then calculated. The random effect parameters in the model are divided into two levels according to the actual altitude of the sample plots, namely u<750m and u≥750m. The random effect of u<750m is set to 0, and the random effect of u≥750m is set to 1. This is done to distinguish the two types of data when constructing the covariance matrix. Based on this, the model parameters in (1)-(3) can be calculated according to the obtained carbon storage parameters and the aboveground carbon storage of the corresponding sample plots. The model parameters can also be solved using existing statistical software such as SPSS, SAS, and R language. The model in this example is fitted with data from 32 sample plots. The applicability of the model is tested using leave-one-out cross-validation. Based on the comprehensive data test results, model (3) has the highest accuracy. The coefficient of determination of the training set is 0.8256, and the coefficient of determination of the validation set is 0.8044. It can be used to estimate the aboveground carbon storage of the forest stand. The aboveground carbon storage estimation model of the forest stand is as follows:

[0086] y = 0.5277e ((-a+u)x)

[0087] u~N(0,1056) (4)

[0088] The following table shows the calculation results of the three carbon storage models mentioned above. It is clear that the third model represented by formula (3) has the best effect.

[0089]

[0090] Example 2

[0091] like Figure 7 As shown, this embodiment provides a device for determining the aboveground carbon storage of a forest stand, the device comprising:

[0092] Image acquisition device 61 is used to acquire images of the forest stand; the image acquisition device can be a camera or any device with a photographic function; calculation unit 62 is used to obtain carbon storage parameters from the images of the forest stand; determination unit 63 is used to estimate the carbon storage of the forest stand using a carbon storage model based on the carbon storage parameters.

[0093] The working principle of each unit in this embodiment can be found in the description of Embodiment 1.

[0094] According to the present invention, the aboveground carbon storage of a forest stand can be estimated using a common camera and a computer. Therefore, the present invention can estimate the aboveground carbon storage of a forest stand without the need for additional equipment. The technology of the present invention is applicable to the dynamic monitoring of aboveground carbon storage in forest farms, long-term fixed sample plots, and other similar scenarios.

[0095] Although the invention has been described by way of examples, those skilled in the art will recognize that many modifications and variations can be made to the invention without departing from its spirit and essence, the scope of which is defined by the appended claims.

Claims

1. A method for estimating aboveground carbon storage in forest stands, characterized in that, include: Acquire images of the forest floor; Carbon storage parameters were obtained from images of the forest floor. The aboveground carbon storage of forest stands is estimated using a carbon storage model based on carbon storage parameters; The carbon storage parameter is: the percentage of the foreground image of trees in the entire image. ; The carbon storage model is any one of the following three equations: in: Aboveground carbon storage in forest stands; : Percentage of the image occupied by the foreground trees; Parameters to be determined; Random effects parameters; The parameters to be determined were obtained through experiments; The method further includes: The steps for setting the random effects parameters are as follows, divided into two levels based on the actual altitude of the sample plots: The random effect is set to 1.

2. The method according to claim 1, characterized in that, The acquisition of images of the forest floor specifically includes: Establish rectangular sample plots; Set up shooting locations; Take images at each shooting point.

3. The method according to claim 1, characterized in that, The specific methods for obtaining carbon storage parameters from images of the forest floor include: Step 51: Load the acquired image into a fully convolutional neural network model; Step 52: The fully convolutional neural network model traverses the image pixel by pixel, starting from the beginning position of the image; Step 53: Calculate the feature value of the pixel according to the weight file, and compare the feature value with the learned target range. If it is within the target value range, the current pixel value is determined as the foreground; otherwise, the pixel value is set as the background. Step 54: Traverse all pixels in order from top to bottom and from left to right, and determine if it is the last pixel. If it is, proceed to step 55; otherwise, return to step 53. Step 55: Extract the foreground portion and calculate the carbon storage parameter using the foreground portion: The carbon storage parameter is the percentage of the forest foreground image in the entire image. .

4. The method according to claim 3, characterized in that, The method further includes: The specific steps to obtain the weight file are as follows: Construct training image samples; The image samples are input into a fully convolutional network structure, which consists of 13 convolutional layers and 5 downsampling layers. Downsampling uses a 2×2 max pooling method with a stride of 2. Non-linear activation is performed using a rectified linear unit function. All convolutional kernels are 3×3 in size with a stride of 2 and padding of 1. The fully convolutional network is trained so that the input image samples are identified pixel by pixel by the network structure. Through the label samples drawn from a large number of image samples, the network model gradually learns that in the same type of image, pixels with a gray value within a certain range are target pixels, while pixels in other ranges are non-target pixels. Then, the range enclosed by the edges of all target pixels or the area of ​​the target pixels is the segmented image. The criteria for these target pixels are the weight file generated by training the network structure.

5. The method according to claim 1, characterized in that, The method further includes a step of obtaining parameters to be determined, specifically including: A certain number of forest stand aboveground images and corresponding measured values ​​of aboveground carbon storage were obtained. The images were segmented to obtain the carbon storage parameters of the corresponding forest stands, and the maximum likelihood estimation method was used to estimate the undetermined parameters.

6. A device for estimating aboveground carbon storage in a forest stand, comprising the method described in any one of claims 1-5, characterized in that, The device includes: Image acquisition unit, used to acquire images of the forest stand's ground surface; The computing unit is used to obtain carbon storage parameters from color images of forest stands; The unit is defined to estimate the aboveground carbon storage of a forest stand using a carbon storage model based on carbon storage parameters.

Images