Method for evaluating carbon storage of mangrove ecosystem based on remote sensing and DeepLabV3+

Abstract

A method for assessing carbon storage in mangrove ecosystems based on remote sensing and DeepLabV3+ includes (1) analysis of spectral characteristics of mangrove communities and construction of a sample library based on remote sensing; (2) extraction of mangrove extent based on DeepLabV3+; and (3) determination of mangrove community categories by combining field surveys, thereby achieving rapid estimation of mangrove carbon storage. This invention, based on remote sensing feature analysis of mangroves and utilizing the DeepLabV3+ deep learning method, achieves rapid extraction of the spatiotemporal distribution characteristics and extent of mangroves, as well as rapid calculation of large-scale mangrove area. By combining remote sensing, deep learning, field surveys, and other methods, it enables rapid estimation of mangrove carbon storage over large spatial spans, providing a reference for economic development and ecological restoration.

Description

Technical Field

[0001] This invention belongs to the field of marine carbon sink technology, specifically a method for assessing the carbon storage capacity of ecosystems. Background Technology

[0002] Marine carbon sinks are the process by which marine activities and marine life absorb carbon dioxide from the atmosphere and fix and store it in the ocean. This is mainly achieved through marine carbon pumps (solution pumps, carbonate pumps, biopumps, etc.) to realize the vertical and horizontal migration and form conversion of carbon in the ocean.

[0003] Coastal wetlands, as an important type of coastal blue carbon ecosystem, possess enormous carbon absorption capacity. Current research on coastal wetlands is receiving increasing attention, especially on the carbon storage capacity of mangrove wetlands. For example, Cole et al. estimated mangrove aboveground biomass as early as 1999 using a mangrove biomass equation; subsequently, Proisy et al. used synthetic aperture radar data in 2003 to study mangrove vegetation biomass in Australia and French Guiana; and in recent years, Guo Yanru extracted vegetation indices from remote sensing images and established a soil carbon storage estimation model through regression analysis to estimate soil carbon storage in the Qinglan Port mangrove wetland. Various remote sensing data have limitations in spatial, spectral, and radiometric resolution, which restricts the accuracy of remote sensing estimation of mangrove aboveground biomass. Furthermore, sampling errors in survey plots can easily lead to inconsistencies in the accuracy of mangrove aboveground biomass estimations using different remote sensing data. By combining ecological factors, environmental factors, and multi-source remote sensing data such as high-resolution imagery and UAV LiDAR imagery, a multi-source remote sensing data model for estimating mangrove aboveground biomass can be constructed, which can mitigate the impact of these factors. Research indicates that enhancing the carbon sequestration capacity of coastal wetlands can be achieved through species selection and integrated planting, habitat modification techniques, planting techniques, and management techniques to improve the survival rate and carbon sequestration efficiency of coastal plant communities. To improve the utilization of ecological resources by organisms, efficient carbon sequestration planting methods using mixed communities of trees, shrubs, and herbaceous plants can be adopted. Simultaneously, vegetation density must be reasonable; too sparse vegetation results in low carbon sequestration, while too dense vegetation leads to poor ventilation, releasing more CO2 than it is fixed, thus having a counterproductive effect. In conclusion, enhancing the carbon sequestration potential of coastal wetlands requires focused research on species selection and planting, improving sediment carbon sequestration, and reducing carbon emissions. Management techniques and conservation efforts are also crucial, as research shows that the restoration of damaged wetlands is a lengthy process.

[0004] Currently, blue carbon-related policies promoted in various countries mainly focus on the integration of carbon emission trading markets and blue carbon programs, policy guarantees for the implementation of blue carbon programs, and the establishment of regional blue carbon trading markets. Scientific research, on the other hand, focuses on the processes and mechanisms of marine carbon sink activities, as well as the quantification of marine ecosystems. Research on marine carbon sinks is still in its early stages. Research on marine carbon sink accounting and management policies has not yet formed a scientific system and urgently needs further exploration. Establishing a scientific methodology and assessment standards for marine carbon sinks can lay a solid foundation for their development, and a reliable marine carbon sink capacity accounting system is a prerequisite for its entry into China's carbon trading market and for improving the allocation of marine resources, which has significant theoretical and practical implications. Summary of the Invention

[0005] The purpose of this invention is to propose a method for monitoring and assessing carbon storage in mangrove ecosystems based on remote sensing and DeepLabV3+. Targeting the main mangrove community species, this invention comprehensively utilizes multiple methods, including remote sensing inversion, deep learning, and field survey sampling. Through remote sensing spectral feature analysis and deep learning, a mangrove classification model is constructed and its range extracted to create a spatial distribution map of mangroves. Combining survey sampling, formula calculation, and component analysis, a method and model for estimating mangrove carbon storage and its changes are established, achieving the goal of "large-scale, rapid, multi-dimensional, dynamic, and intelligent" assessment of mangrove carbon sequestration capacity.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, the specific process of which is as follows: Figure 1 As shown.

[0007] This invention discloses a method for assessing carbon storage in mangrove ecosystems based on remote sensing and DeepLabV3+. It mainly comprises three parts: (1) analysis of spectral characteristics of mangrove communities and construction of a sample library based on remote sensing; (2) extraction of mangrove extent based on DeepLabV3+; and (3) determination of mangrove community categories through field surveys, thereby achieving rapid estimation of mangrove carbon storage. Specifically, it includes the following steps:

[0008] Step S1: Remote Sensing Spectral Feature Analysis and Sample Library Construction for Mangroves. First, high-resolution satellite and UAV remote sensing image data at high and low tide levels are collected. The selected hyperspectral data undergoes preprocessing such as radiometric calibration and atmospheric correction to convert radiance to surface reflectance and remove atmospheric path radiation interference. Then, through band operations, Normalized Difference Water Index (NDWI) and Normalized Difference Vegetation Index (NDVI) are extracted from the preprocessed images, generating corresponding vectors for water-land and vegetation-land regions, respectively. This comprehensive analysis determines the mask range for selecting mangrove samples. Based on the mask range, a deep learning sample library for semantic segmentation of mangrove remote sensing images is constructed, and the samples are labeled. For example... Figure 2 As shown.

[0009] Furthermore, the normalized water index (NDWI) extraction process involves the following steps: Land features are distinguished by differences in reflectance in the visible and near-infrared bands. Water reflectance gradually decreases from the visible to the infrared bands, resulting in significant differences in wetland types around flooded areas between these two bands. The near-infrared band exhibits strong absorption, while vegetation shows strong reflectance in this band. However, the application of NDWI is affected by soil background interference. ρG reen ρNIR represents the reflectance of green light, while ρNIR represents the reflectance of near-infrared light.

[0010]

[0011] Furthermore, the normalized density vegetation index (NDVI) extraction is implemented through the following steps: based on the reflectance characteristics of vegetation, a combination of multiple wavelength bands is used for calculation. The NDVI uses the red and near-infrared wavelengths, and the calculation formula is as follows. Compared to other vegetation indices, NDVI is less affected by the atmosphere and also enhances the identification of soil disturbances, making it one of the most widely used vegetation indices to date. It can be used for vegetation extraction, calculating vegetation cover, retrieving chlorophyll content, estimating biomass, and measuring photosynthetically active radiation absorption, among other things. ρNIR represents near-infrared reflectance, and ρR... e d represents the reflectivity of red light.

[0012]

[0013] Step S2: Based on the spectral image features of mangroves in remote sensing images, this invention constructs a semantic segmentation model for mangrove images using the DeepLabV3+ network. This model segments the mangrove area in large-scale remote sensing images and calculates the mangrove area using a statistical pixel method. Figure 3 As shown.

[0014] Furthermore, the implementation steps of the semantic segmentation model for mangrove images based on the DeepLabV3+ network are as follows:

[0015] (1) Based on the sample library data constructed by S1, features are extracted from the input image through the backbone network ResNet and low-level semantic feature maps and high-level semantic feature maps are generated;

[0016] (2) Based on the high-level semantic feature map obtained in (1), multi-scale sampling is performed using the Atrous Spatial Pyramid Pooling (ASPP) module to generate multi-scale feature maps;

[0017] (3) Combine the multi-scale high-level semantic feature maps obtained in (2) in the channel dimension, and perform channel dimensionality reduction through 1×1 convolution;

[0018] (4) The low-level semantic feature map generated in (1) is reduced in channel dimension by 1×1 convolution to maintain the proportion when it is concatenated with the high-level semantic feature map and enhance the network learning ability.

[0019] (5) Upsample the high-level semantic feature map generated in (3) by 4 times;

[0020] (6) Connect the feature maps in (4) and (5) to form a new feature map;

[0021] (7) Extract features using a 3×3 convolution, then upsample by 4 times to generate the final semantic segmentation map.

[0022] Step S3: Based on field surveys, determine the mangrove community categories, and use methods such as litter collection and formula calculations to estimate the aboveground biomass and carbon storage of mangroves, constructing carbon storage estimation models for different mangrove ecosystems. For example... Figure 4 As shown.

[0023] Furthermore, the implementation steps of the aforementioned mangrove ecosystem carbon storage estimation model are as follows:

[0024] (1) Selecting quadrats: Based on the community composition type, one quadrat is randomly set for each community. Since the communities are mostly distributed in strips, the size of the woody plant community quadrat is set to 10m×5m, and the size of the herbaceous plant community quadrat is 2m×2m. The number, height, average spacing and diameter at breast height of trees (grasses) in the quadrats are counted.

[0025] (2) Community structure survey: For plants below 2m, the direct measurement method was used, and the readings were taken directly using a ruler; for trees above 2m, the trigonometric leveling method was used for measurement.

[0026] (3) Calculation of aboveground biomass: Each tree in the sample plot was measured to obtain its height and diameter at breast height, as well as the plot's cover, leaf area index, and soil salinity. Based on various empirical models of mangrove biomass, the biomass per tree and the biomass of each aboveground component (bark, trunk, leaves, bare branches, and branches with leaves) were calculated. Furthermore, the total aboveground biomass and pixel biomass of the sample plot were calculated.

[0027] (4) Calculation of underground biomass: Since mangroves are generally located in protected areas and cannot be damaged or excavated, this invention uses ground-penetrating radar to scan the underground roots of mangroves and analyzes the biomass of the underground roots of mangroves based on the transmitted waveform.

[0028] (5) Litter biomass calculation: The collected litter was dried in an 85℃ drying oven and weighed. The decomposition rate (decomposition constant) was calculated according to the Olson index model:

[0029] L r =ae -kt =(X i / X0)×100% (3)

[0030] Where: L r denoted as residual litter (g); a is a correction factor; k is the decomposition constant; X i X0 represents the dry weight (g) of the litter sample taken from the i-th litter; X0 represents the initial dry weight (g) of the litter.

[0031] The biomass accumulated on the ground by litter is calculated based on the monthly collection amount and monthly decomposition rate, using the following formula:

[0032]

[0033] In the formula: R i The total amount of litter remaining from the i-th plant species; i represents the i-th plant species; t represents the decomposition time; t i The time required for 95% decomposition of the i-th plant; k i Let a be the decomposition constant of the litter of the i-th plant species; i is the correction coefficient for litter of the i-th plant species.

[0034] (6) Calculation of total biomass of plant communities

[0035] The total biomass of a mangrove plant community includes the aboveground parts, belowground parts, and litter. The biomass of each community can be determined by combining its area with that of the individual communities.

[0036] (7) Methods for calculating plant carbon storage

[0037] Based on the photosynthetic reaction equation, it is estimated that 1.62g of CO2 is needed to produce 1g of dry matter. According to the molecular formula and the corresponding atomic mass, the relative atomic mass of carbon is 12, and that of carbon dioxide is 44. That is, the mass fraction of carbon in carbon dioxide is 27.28%, and 0.44g of C is needed to produce 1g of dry matter.

[0038]

[0039] (8) Calculation method for standard deviation of carbon storage

[0040] The standard deviations of vegetation, sediment, and total carbon storage per unit area for each sample plot are calculated as follows:

[0041]

[0042]

[0043] X represents the average carbon storage per unit area of ​​each quadrat, in hectares (g / ha); n is the number of quadrats; X i This represents the carbon storage per unit area of ​​the i-th sample plot, expressed in g / ha.

[0044] Step S4: Based on the extraction of mangrove range and calculation of mangrove area in Step S2, and combined with the carbon storage estimation model of mangrove ecosystem for different communities in Step S3, the carbon storage of all mangrove ecosystem communities extracted in Step S1 is estimated, and the carbon storage of all mangrove ecosystems in the area to be evaluated is obtained.

[0045] This study compares the carbon storage capacity of different mangrove communities, considering the carbon storage capacity per unit area and its distribution area, and statistically analyzes the distribution characteristics (distribution, area, community, etc.) of different mangrove patches. It assesses the carbon storage of mangrove ecosystems and compiles an assessment report to provide computational basis and technical support for blue carbon accounting and carbon trading. Considering the key factors or environmental characteristics affecting carbon storage capacity within the same ecosystem type, it constructs ecosystem restoration projects aimed at increasing carbon storage, comprehensively optimizing the carbon storage capacity of coastal wetlands.

[0046] This invention analyzes the remote sensing spectral characteristics of different mangrove communities based on long-term satellite and UAV imagery to extract the mangrove range; it establishes a remote sensing sample library of mangrove communities and uses deep learning methods to construct an automatic range extraction model for different mangrove communities, enabling automatic calculation of mangrove community area; through field surveys, litter collection, and formula calculations, it establishes a mangrove carbon storage estimation model to estimate and analyze the composition and carbon sequestration capacity of mangrove ecosystems, providing technical support for promoting a "win-win" situation of marine economic development and ecological carbon sink.

[0047] Compared with the prior art, the present invention has the following beneficial technical effects:

[0048] 1) This method is based on remote sensing feature analysis of mangroves and uses DeepLabV3+ deep learning method to quickly extract the spatiotemporal distribution features and range of mangroves, as well as to quickly calculate the area of ​​mangroves over a large area.

[0049] 2) This method combines remote sensing, deep learning, field surveys and other means to quickly estimate the carbon storage of mangroves over a large spatial span, providing a reference for economic development and ecological restoration. Attached Figure Description

[0050] Figure 1 This is a flowchart of the technology of the present invention.

[0051] Figure 2 This document outlines the process for remote sensing spectral feature analysis and range extraction of mangroves.

[0052] Figure 3 This is a schematic diagram of the DeepLabV3+ network.

[0053] Figure 4 This is a semantic segmentation effect of mangroves based on remote sensing imagery.

[0054] Figure 5 A model for estimating carbon storage in mangrove ecosystems.

[0055] Figure 6 The distribution of mangrove biomass per unit area in each sample plot.

[0056] Figure 7 The distribution of biological vegetation biomass per unit area in different mangrove communities. Detailed Implementation

[0057] The present invention will now be further described with reference to the accompanying drawings.

[0058] A method for assessing carbon storage in mangrove ecosystems based on remote sensing and DeepLabV3+ includes the following steps:

[0059] S1. Collect high-resolution satellite and UAV remote sensing image data of high tide and low tide levels.

[0060] S2. Preprocess the remote sensing data, including radiometric calibration and atmospheric correction.

[0061] S3. Based on the preprocessed remote sensing images, extract NDWI and NDVI, generate corresponding vectors for water-land and vegetation-land regions, and comprehensively determine the mask range for mangrove samples.

[0062] S4. Based on remote sensing imagery, establish a mangrove sample database and label the samples. See Table 1.

[0063] Table 1. Sample Library Categories and Quantities

[0064] Serial Number type Number of training samples Number of test samples 1 mangrove 150 50 2 Paulownia tree 150 50 3 Mulan 150 50 4 Red Sea Olive 150 50 5 White bone soil 150 50 6 water body 150 50 7 architecture 150 50 8 Other (roads, mudflats) 150 50

[0065] S5. Construct a semantic segmentation model for mangrove images and calculate the area of ​​the mangrove forest. After appropriate preprocessing of the sample database dataset, the input to the network is batch×3×256×256 images and batch×256×256 labels. The label values ​​range from 0 to 7, representing a total of 8 classes. Here, batch is the number of samples used for training in each batch.

[0066] (1) Feature extraction of the image was performed using the ResNet18 network, resulting in a 512×32x×32 feature map.

[0067] (2) The hollow spatial pyramid pooling network layer performs convolution extraction operations on the feature map obtained in the previous step from each of the 5 branches, and the resulting feature maps are all 256×32×32. Then, they are spliced ​​together to obtain a 1280×32×32 feature map, and then two convolutions are performed to obtain an 8×32×32 feature map.

[0068] (3) Finally, perform an upsampling operation on the 8×32×32 feature map to obtain an 8×256×256 feature map, which is the final output we need.

[0069] (4) The loss function directly calculates the cross-entropy loss for each pixel of the 8×256×256 feature map and the 256×256 label. Considering the class imbalance problem, the network also adds corresponding weights to the loss of each class. The final classification accuracy is shown in Table 2.

[0070] Table 2 Classification Accuracy Statistics

[0071] Serial Number type Error Quantity accuracy 1 mangrove 18 0.64 2 Paulownia tree 4 0.92 3 Mulan 6 0.88 4 Red Sea Olive 11 0.78 5 White bone soil 7 0.86

[0072] (5) Use the obtained semantic segmentation model to perform semantic segmentation on the remote sensing image and extract the range of the mangrove community, such as Figure 4 As shown; and the area of ​​mangroves was calculated using the statistical pixel method.

[0073] S6. Taking the Guangxi Shankou Mangrove Nature Reserve as an example, based on field investigation, the classification of mangrove communities was determined, and then 5 quadrats (specific locations are shown in Table 3) were selected to conduct biomass investigation and sampling analysis.

[0074] Table 3. Locations of 5 quadrats

[0075] Serial Number Station Number longitude latitude 1 Z001 109.6738 21.5722 2 Z002 109.7482 21.5575 3 Z003 109.6691 21.5562 4 Z004 109.7612 21.5256 5 Z005 109.7613 21.4951

[0076] S7. Conduct surveys and record data for each sample plot. The survey content and methods are shown in Tables 4 and 5. Construct carbon storage assessment models for different mangrove communities using methods such as surveying, weighing, and ground-penetrating radar. For example... Figure 5 .

[0077] Table 4. Survey content and methods for each sample plot

[0078]

[0079]

[0080] Table 5 Biomass estimation models for different mangrove communities

[0081]

[0082] S8. Based on the mangrove range and area extracted in S5, the survey results in S6, and the estimation model established in S7, estimate the biomass per unit area of ​​each quadrat and the biomass per unit area of ​​each community species, such as... Figure 5 , Figure 6 As shown in the figure. The estimation results show that the aboveground biomass per unit area of ​​different mangrove communities, from largest to smallest, is as follows: *Rhizophora stylosa*, *Rhizophora mukorossi*, *Symplocos edulis*, *Avicennia marina*, and *Kandelia candel*. Finally, the carbon storage of each mangrove community in the area to be evaluated was summed to obtain the total carbon storage: the total carbon storage of the Yamaguchi mangrove ecosystem is 3.45 × 10⁻⁶. 5 MgC is mainly composed of sedimentary carbon reserves, with mangrove vegetation accounting for 8.48%, sediments accounting for 91.46%, and litter accounting for only 0.05%.

[0083] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. A method for assessing carbon storage in mangrove ecosystems based on remote sensing and DeepLabV3+, characterized by: Includes the following steps: Step S1: Analysis of remote sensing spectral features of mangroves and construction of a sample library: First, high-resolution satellite and UAV remote sensing image data at high tide and low tide are collected. Radiometric calibration and atmospheric correction preprocessing are performed on the selected hyperspectral data to complete the conversion of radiance to surface reflectance and the removal of atmospheric path radiation interference. Then, through band operations, the Normalized Difference Water Index (NDWI) and Normalized Difference Vegetation Index (NDVI) are extracted from the preprocessed images to generate corresponding vectors for water-land and vegetation-land regions, respectively. The mask range for selecting mangrove samples is determined comprehensively. Based on the mask range, a deep learning sample library for semantic segmentation of mangrove remote sensing images is constructed, and the samples are labeled. Step S2: Based on the spectral image features of mangroves in remote sensing images, construct a semantic segmentation model for mangrove images based on the DeepLabV3+ network to segment the mangrove area in a large-scale remote sensing image and calculate the mangrove area by statistical pixel method. Step S3: Based on the field survey, determine the mangrove community category, and use litter collection and formula calculation methods to estimate the aboveground biomass and carbon storage of mangroves, and construct carbon storage estimation models for different mangrove ecosystems. Step S4: Based on the extraction of mangrove range and calculation of mangrove area in Step S2, and combined with the carbon storage estimation model of mangrove ecosystem for different communities in Step S3, the carbon storage of all mangrove ecosystem communities extracted in Step S1 is estimated, and the carbon storage of all mangrove ecosystems in the area to be evaluated is obtained.

2. The method for assessing carbon storage in mangrove ecosystems based on remote sensing and DeepLabV3+ according to claim 1, characterized in that the construction of a deep learning sample library for semantic segmentation of mangrove remote sensing images in step S1 is implemented as follows: S11: Collect high-resolution satellite and UAV remote sensing image data of high tide and low tide levels, and perform preprocessing. S12: Extract the Normalized Water Index (NDWI) from the preprocessed image to generate the corresponding water-land range vector data; distinguish land features by the difference in reflectance in the green band and near-infrared band. The water reflectance category in the visible to infrared bands shows a gradually decreasing trend, resulting in significant differences in wetland categories around the flooded area in these two bands. Furthermore, the near-infrared band has strong absorption characteristics, while vegetation has strong reflectance in this band. The calculation formula is as follows. （1） in, Indicates the reflectivity of green light. Indicates the reflectivity of the near-infrared region; S13: Extract the Normalized Difference Vegetation Index (NDVI) from the preprocessed image and generate the corresponding vegetation-land area range vector data; based on the reflectance characteristics of vegetation, use a combination of multiple band expressions for calculation, where the expression bands used for the NDVI are the red light band and the near-infrared band, and the calculation formula is as follows. （2） in, Represents the reflectivity of the near-infrared region. Indicates the reflectivity of red light; S14: Vector overlay S12 and S13 to comprehensively determine the mask range for selecting mangrove samples, and crop and label the remote sensing images based on the mask range to construct a deep learning sample library.

3. The method for assessing carbon storage in mangrove ecosystems based on remote sensing and DeepLabV3+ according to claim 1, characterized in that the semantic segmentation model of mangrove images constructed based on the DeepLabV3+ network in step S2 is implemented as follows: S21: Based on the sample database built from S1, features are extracted from the input image through the backbone network ResNet to generate low-level semantic feature maps and high-level semantic feature maps; S22: Based on the high-level semantic feature map obtained in S21, multi-scale sampling is performed using the Spatial Pyramid Pooling Module (ASPP) to generate multi-scale feature maps. S23: Combine the multi-scale high-level semantic feature maps obtained in S22 along the channel dimension, and perform channel dimensionality reduction through 1×1 convolution; S24: The low-level semantic feature map generated in S21 is reduced in channel dimension by 1×1 convolution to maintain the weight when concatenated with the high-level semantic feature map and enhance the network's learning ability. S25: Upsample the high-level semantic feature map generated in S23 by a factor of 4; S26: Connect the feature maps in S24 and S25 to form a new feature map; S27: Extract features using a 3×3 convolution, then upsample by 4 times to produce the final semantic segmentation map.

Images