Remote sensing monitoring method for plateau surface vegetation coverage and carbon sink function

Abstract

The application discloses a kind of highland surface vegetation coverage and carbon sink function remote sensing monitoring method, it is related to remote sensing ecological monitoring technical field, comprising the following steps: the highland remote sensing image is divided into multiple image blocks, and the multispectral feature vector of each image block is extracted;Image block is input into the spatial-semantic structure recognition model based on the structure constructed by Transform;Based on spatial dependence relationship and semantic association relationship, the structure recognition result and vegetation type classification result of image block are generated;According to structure recognition result and vegetation type classification result, the carbon absorption parameter corresponding to vegetation type and structure is combined, and the carbon sink function index of corresponding image area is calculated;The monitoring result including surface vegetation coverage distribution map, vegetation structure diagram and carbon sink function spatial distribution map is output.The application solves the problem of insufficient carbon sink function remote sensing evaluation precision under the complex structure environment of highland area.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing ecological monitoring, more particularly to a method for remote sensing monitoring of plateau surface vegetation coverage and carbon sink function. BACKGROUND

[0002] The plateau region has important carbon sink potential and ecological regulation function in the global ecosystem, and the surface vegetation plays an irreplaceable role in maintaining ecological stability, regulating climate and realizing regional carbon neutralization. With the development of remote sensing technology, using satellite remote sensing data to monitor the surface vegetation coverage and carbon sink function has become a mainstream method. Especially in the ecological sensitive area of the plateau, constructing a continuous and accurate carbon sink monitoring system has practical significance for guiding ecological protection engineering and promoting the realization of ecological product value. Remote sensing images can provide large-scale, high-temporal-resolution surface observation data, and through image classification, vegetation index estimation and carbon flux model inversion, the spatial evaluation of regional vegetation conditions and carbon absorption capacity can be realized.

[0003] However, in the existing technical system, the methods of vegetation identification and carbon sink estimation generally have weak spatial semantic understanding ability. The traditional method usually classifies remote sensing images through fixed threshold vegetation index (such as NDVI, EVI), which is simple and easy to use, but ignores the differences in structure continuity and semantic distribution between different regions, and is difficult to effectively identify the edge transition characteristics of ecological patches. In addition, although the common convolutional neural network structure can improve the classification accuracy, its perception range is limited to the local area of the image, and it cannot establish the ecological structure relationship between remote image blocks, which is not suitable for the plateau region with complex terrain and irregular distribution of vegetation patches. In multi-temporal remote sensing monitoring, cloud cover is also a problem that cannot be ignored. The plateau region is significantly affected by monsoon, and cloud cover is frequent, resulting in a large number of missing areas in remote sensing images. The linear interpolation or simple temporal smoothing method used for image reconstruction has limited effect in preserving image details and spatial continuity. In addition, the existing carbon sink model is mostly based on stationary climate data, and lacks the ability to simulate sudden extreme climate events, making it difficult to accurately evaluate the carbon sink elasticity and steady-state fluctuations of the ecological system in the plateau environment where extreme weather occurs frequently, and the prediction ability of vegetation response under sudden climate impact is insufficient. Therefore, how to construct a remote sensing monitoring method that can understand image structure and semantics, has image completion ability, and can simulate extreme climate disturbance and be used for carbon sink function calculation has become a problem that needs to be broken through in the current technical field. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a method for remote sensing monitoring of plateau surface vegetation coverage and carbon sink function to solve the problems mentioned in the background.

[0005] In order to achieve the above object, the present application adopts the following technical solutions:

[0006] A highland surface vegetation coverage and carbon sink function remote sensing monitoring method, comprising the following steps:

[0007] Obtaining a highland region surface image by using a remote sensing image, dividing the remote sensing image into a plurality of image blocks, and extracting a multispectral feature vector of each image block;

[0008] Inputting the image blocks into a spatial-semantic structure recognition model constructed based on a Transformer structure, and establishing spatial dependency relationships and semantic association relationships between the image blocks through a self-attention mechanism;

[0009] Generating a structure recognition result and a vegetation type classification result of the image blocks based on the spatial dependency relationships and the semantic association relationships;

[0010] According to the structure recognition result and the vegetation type classification result, combining carbon absorption parameters corresponding to the vegetation type and the structure, and calculating a carbon sink function index of a corresponding image region;

[0011] Outputting a monitoring result including a surface vegetation coverage distribution map, a vegetation structure map, and a carbon sink function spatial distribution map.

[0012] In an optional embodiment, when there is a region in the remote sensing image in which image information is missing due to cloud cover shielding, an image reconstruction method based on a time sequence is used for image completion.

[0013] In an optional embodiment, the image reconstruction method comprises:

[0014] Image registration is performed on historical multi-temporal remote sensing images of the missing region, and multispectral feature vectors of the corresponding region at adjacent time nodes are extracted;

[0015] A spectral prediction model taking a time sequence as input is constructed, and the time variation trend of the feature vectors is used to predict feature values at the missing time;

[0016] The spatial features of the edge image blocks of the missing region are combined in space, and a spatial interpolation algorithm is used to further correct the predicted feature values;

[0017] The predicted results are reconstructed into complete image blocks, and are used to replace the original missing region for subsequent vegetation structure recognition and carbon sink function calculation.

[0018] In an optional embodiment, when monitoring the carbon sink function, a modeling process based on enhanced extreme climate disturbance is introduced, and the modeling process comprises:

[0019] constructing a climate disturbance sample set containing historical meteorological data, the meteorological data including any one or more of the following: temperature, humidity, precipitation, wind speed, wind direction, solar radiation multi-dimensional climate factors;

[0020] applying a disturbance function to the climate factor data to simulate short-term or sudden extreme climate events that may occur in plateau regions;

[0021] inputting the disturbed climate samples into a disturbance enhancement training module to jointly train a vegetation response prediction model with the vegetation feature vectors extracted from the remote sensing images;

[0022] adjusting and correcting the carbon sink stability of the vegetation region based on the disturbance response scores output by the vegetation response prediction model, and outputting the prediction results of the changes in carbon sink function of the disturbance-sensitive regions.

[0023] In an optional embodiment, the remote sensing images include multi-temporal and multi-spectral remote sensing images, each image containing red light, near-infrared, short-wave infrared and red edge bands.

[0024] In an optional embodiment, the multi-spectral feature vector includes a combination of average reflectivity, standard deviation, normalized vegetation index, enhanced vegetation index, chlorophyll red edge index, and terrain-related variables of each image block at multiple bands.

[0025] In an optional embodiment, the Transformer structure in the spatial-semantic structure recognition model includes an image block embedding module, a multi-head self-attention module, a feedforward neural network module, and a position encoding module, which learns the spatial and semantic context dependencies of the image blocks to generate structure discrimination outputs.

[0026] In an optional embodiment, the semantic association relationship between the image blocks is obtained by constructing an ecological semantic atlas, which is constructed based on plant type similarity, functional ecological zone adjacency, and historical evolution trend.

[0027] In an optional embodiment, the structure recognition result includes the ecological structure label, structure integrity score, and ecological boundary type of each image block, and the vegetation type classification result includes any one or more of the following: shrubs, alpine meadow, plateau grassland, rocky bare land, and permafrost region.

[0028] In an optional embodiment, the calculation of the carbon sink function index is based on the unit carbon flux parameter corresponding to the vegetation type and the structure integrity factor, and different weight carbon sink function estimation models are set for regions with the same vegetation type but different structure integrity.

[0029] The advantages of the present application over the prior art are that the present application realizes the learning of the spatial dependence relationship and the semantic association relationship between the image blocks in the remote sensing image by introducing a spatial-semantic structure recognition model based on the Transformer structure, and breaks through the limitations of the traditional remote sensing classification model that strongly depends on local features and lacks long-distance context understanding. By dividing the remote sensing image into multiple image blocks, extracting the multispectral feature vectors thereof, and identifying the global dependence relationship between the image blocks by means of the self-attention mechanism of the Transformer, precise modeling and type discrimination of the ecological structure are realized, and a more realistic spatial basis is provided for subsequent carbon sink function calculation. Through the fusion of the structure recognition result and the vegetation classification result, a carbon sink function index map with precision and ecological semantic interpretation is constructed by combining the carbon absorption parameters corresponding to the vegetation types and structures, thereby improving the spatial resolution and ecological rationality of carbon sink monitoring.

[0030] On this basis, the present application further introduces a plurality of auxiliary mechanisms. When the remote sensing image is blocked by clouds, causing information loss, a time series-based image reconstruction method is constructed to perform registration processing on historical multi-temporal remote sensing images, extract multispectral features of adjacent time nodes, and combine time trends and spatial edge information to complete the reconstruction of the spectral values of the missing areas, thereby significantly improving the spatio-temporal integrity and continuity of the remote sensing data, providing stable input for the model, and avoiding interruption of carbon sink evaluation. In addition, in the dynamic modeling process of the carbon sink function, an extreme climate disturbance sample set is constructed, and a disturbance function is applied to the climate factor data to simulate short-term mutation events in plateau regions, and then the vegetation response prediction model is trained jointly with the vegetation feature vectors extracted from the remote sensing images, realizing the carbon sink stability analysis of the ecological system under high-intensity climate pressure. The output disturbance response score is used to adjust and correct the carbon sink function of the vegetation area, so that the monitoring result can reflect the uncertainty fluctuations in reality, and has the ability to predict and respond to extreme climate. These beneficial technical solutions work together to provide a solid support for building a complete, dynamic and real scene-oriented plateau carbon sink remote sensing monitoring system. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is the overall flowchart of the method of the present application;

[0032] Figure 2 is the flowchart of the time series-based image reconstruction mechanism of the present application;

[0033] Figure 3 is the flowchart of the carbon sink steady-state perception modeling enhanced based on extreme climate disturbance of the present application. DETAILED DESCRIPTION

[0034] The specific embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0035] In the remote sensing monitoring task of plateau ecological system, due to the complex terrain, variable climate, sparse vegetation and diverse types, the traditional image recognition algorithm often cannot realize the fine identification and high-precision classification of the ground vegetation, especially in the process of calculating the carbon sink function, there is error accumulation and spatial distortion. Therefore, constructing an algorithm model which can comprehensively perceive the image content, understand the spatial pattern of vegetation and predict its ecological function has become an important technical breakthrough point of the current plateau ecological remote sensing research.

[0036] Under this background, the present application proposes a monitoring method fusing remote sensing image structure recognition and ecological semantic understanding, which not only pays attention to whether there is green vegetation on the ground, but further identifies the specific type, distribution structure and ecological continuity of the green vegetation, and finally establishes a carbon sink ability expression model at the pixel level.

[0037] As shown in Figure 1 More specifically, the present application comprises the following steps:

[0038] The plateau region ground surface image is obtained by using a remote sensing image, the remote sensing image is divided into a plurality of image blocks, and a multi-spectral feature vector of each image block is extracted;

[0039] The image block is input into a spatial-semantic structure recognition model constructed based on a Transformer structure, and the spatial dependency relationship and the semantic association relationship between the image blocks are established through a self-attention mechanism;

[0040] Based on the spatial dependency relationship and the semantic association relationship, a structure recognition result and a vegetation type classification result of the image block are generated;

[0041] According to the structure recognition result and the vegetation type classification result, the carbon sink function index of the corresponding image region is calculated by combining the carbon absorption parameters corresponding to the vegetation type and the structure;

[0042] The monitoring result including the ground vegetation coverage distribution map, the vegetation structure map and the carbon sink function spatial distribution map is output.

[0043] More specifically, in the above steps, the first step is to obtain a remote sensing image data source covering the target plateau region, usually a satellite remote sensing image with medium-high spatial resolution and multi-spectral components is selected, such as Sentinel-2, Landsat-8, etc. These images contain multiple bands, covering visible light, red edge, near-infrared, short-wave infrared and other multiple spectral regions closely related to vegetation.

[0044] After obtaining the original remote sensing image, it needs to be pre-processed, including radiation correction, atmospheric correction, image registration, cropping and cloud shadow removal, to ensure the spatial and temporal consistency and quality stability of the data. After completing the basic processing, the whole remote sensing image is divided into multiple image blocks according to the rule sliding window method, for example, using a division size of 16x16 pixels, and covering extraction with a step size of 8 pixels. After division, the spectral features of each image block are extracted, including but not limited to the average reflectivity of red, near-infrared and red edge bands, normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), red edge vegetation index, etc. At the same time, structural features such as image texture, color histogram, brightness contrast, spatial gray level co-occurrence matrix, etc. are introduced, and then a multi-dimensional feature vector is formed.

[0045] These image blocks and their corresponding feature vectors will be input into the spatial-semantic structure recognition model constructed. Unlike traditional convolutional neural networks, the Transformer model does not rely on local receptive fields, but uses a self-attention mechanism to model the global relationship between image blocks. Under this mechanism, each image block not only obtains classification basis from its own features, but also determines its spatial structure position and ecological significance through information interaction with all other image blocks.

[0046] For example, a seemingly ordinary green block may be determined to be in the transition zone between meadow and shrub after establishing semantic connections with its neighboring blocks. This semantic information can only be revealed in the context of the entire image. In addition, to enhance the ecological rationality of semantic modeling, prior ecological semantic graphs are further embedded in the Transformer structure, such as constructing dependency graphs between plant types based on historical classification data, regional vegetation succession rules and neighborhood spatial topology, so that the model can more accurately identify complex structures such as ecological transition zones, vegetation patches and fracture zones.

[0047] After model processing, the system outputs the structure recognition result and vegetation type classification result of each image block. The structure recognition result is not simply labeled as grassland or shrub, but further indicates whether the block is in the core area, edge area or broken transition area of the grassland, and calculates the structure integrity score.

[0048] The structure integrity score is used to comprehensively reflect the area proportion of the core area, edge area and transition area with high continuity in the image block. Specifically, for the i-th image block, its structure integrity score s i is calculated by the following formula:

[0049]

[0050] where Acore,i denotes the area of image i that is identified as intact grassland core region, A e,i denotes the area of image i that is identified as edge region, A tr,i denotes the area of image i that is identified as fragmented or transition region, A to，i is the total area of image i. Parameters β and γ are the weight coefficients of edge region and transition region, respectively, whose values satisfy 0 < γ < β ≤ 1.

[0051] In this formula, the core region is given the highest weight (regarded as 1), and the edge region and the transition region are given appropriate reductions according to their lower contribution to ecological continuity.

[0052] Through the above formula, even within the same vegetation type, the more complete the structure and the higher the continuity of the region, the closer the s i value to 1, while the more fragmented and more serious the edge of the region, the s i value is significantly reduced, thus quantitatively reflecting the stability of the ecological structure in the region.

[0053] Vegetation types are classified into alpine meadow, shrub, rock bare land, permafrost region, etc. according to the prior labeling system. Based on the two output results, the system combines the existing vegetation type-carbon flux control parameter library to calculate the carbon sink capacity of each image block. The mapping function not only determines the carbon absorption capacity per unit area according to the vegetation type, but also introduces the structural integrity factor as a regulating variable. For example, the same as alpine meadow, the region with high integrity can be given a higher weight of carbon absorption efficiency, while the region with serious fragmentation is given a weakening adjustment.

[0054] Specifically, let y i denote the vegetation type (such as alpine meadow, shrub, rock bare land, permafrost region, etc.) to which image i belongs, and a corresponding unit carbon absorption parameter library is established for each vegetation type in advance, where denotes the unit area carbon absorption coefficient corresponding to the vegetation type y i . This coefficient reflects the amount of carbon that can be absorbed per square meter of this vegetation under ideal conditions. The mapping function combines the above two factors to calculate the carbon sink capacity C i of image i, and the calculation formula is:

[0055]

[0056] In this formula, C i is the carbon sink capacity of image i, is the unit carbon absorption coefficient corresponding to the vegetation type y i , and s iis the structural integrity score calculated above. Through this mapping function, even in the same vegetation type, due to the difference in ecological structure, the carbon absorption efficiency of different regions will also be different, and the structural integrity region (s i with higher carbon sink capacity will be enhanced accordingly, while the carbon sink capacity of the structural fragmentation region (s i with lower carbon sink capacity will be weakened, thus making the carbon sink assessment more refined and scientific.

[0057] Finally, the remote sensing monitoring system will output three spatialized result layers:

[0058] The first layer is the surface vegetation coverage distribution map, which reflects the overall greening level of the region.

[0059] The second layer is the vegetation structure map, which shows the structural complexity and connectivity of the ecological space, including ecological boundaries, fragmented areas, patch centers, etc. It can also be the core area, edge area and transition area mentioned above.

[0060] The third layer is the carbon sink function spatial distribution map, which intuitively reflects the carbon sink contribution level of each region, which can be further superimposed with administrative boundaries, land use boundaries and other layers for decision support.

[0061] In actual landing scenarios, for example, in a certain alpine meadow restoration area in the southern Qinghai-Tibet Plateau, the system can output the changes in vegetation coverage, structure reconstruction trends and total carbon sink changes before and after restoration by comparing remote sensing images of different years, providing quantitative evaluation evidence for ecological restoration projects, and can real-time early warning of risk areas of vegetation degradation or carbon sink capacity decline.

[0062] In the actual remote sensing monitoring process, especially when observing surface vegetation and carbon sink function in plateau areas, a very prominent problem is that remote sensing images are often affected by cloud cover. Due to its unique topography and climate characteristics, the plateau region has a high frequency and long duration of cloud cover, resulting in missing areas in the acquired images. Especially in multi-temporal observation, in order to capture the dynamic process of vegetation growth, a continuous and stable sequence of remote sensing images is needed, but the interference of clouds makes some image data at certain time nodes partially or completely unavailable, thus breaking the monitoring chain and affecting the accuracy and continuity of subsequent model structure identification, time series analysis and carbon sink calculation. Therefore, an image reconstruction method based on time series can be used for image completion.

[0063] As shown in Figure 2 , the image reconstruction method includes:

[0064] image registration is performed on the historical multi-temporal remote sensing images of the missing area, and a multi-spectral feature vector of the corresponding area at adjacent time nodes is extracted;

[0065] A spectral prediction model is constructed with time series as input, and the time trend of the feature vector is used to predict the feature value at the missing time point.

[0066] The spatial features of the missing area edge image block are combined in space, and the predicted feature value is further corrected by using the spatial interpolation algorithm.

[0067] The prediction result is reconstructed into a complete image block and replaces the original missing area for subsequent vegetation structure identification and carbon sink function calculation.

[0068] The above method is based on a core assumption that the spectral change of vegetation within a certain time scale is predictable, that is, although the image at a certain time point is blocked by clouds, there may be cloud-free images at the previous and subsequent time points. By modeling the spectral features of these adjacent time images, the spectral performance of the blocked time point image can be reasonably inferred.

[0069] Specifically, the first step of image reconstruction is to collect remote sensing images of the missing area at multiple time nodes, and to perform registration processing on these images. The purpose of registration is to ensure that the images in the time series are accurately aligned in space, so that the position of a pixel at different time points is consistent, thereby facilitating the analysis of the change of the pixel in the time dimension. After registration, for an image area blocked by clouds, its performance in historical images can be traced back, and the multispectral reflectance features of the area at multiple time points can be extracted. These historical features can be described as a sequence of vectors that change over time, reflecting the growth trend, change rhythm and seasonal response of the vegetation at that location.

[0070] After mastering the time series features, the modeling stage is entered. At this time, a time series prediction model can be used to learn the spectral change trend. The input of the model is the feature vector at the historical time point, and the output is the predicted vector at the current blocked time point. The form of the prediction model can be flexibly selected, such as traditional regression model, sliding window average, weighted interpolation, or more advanced machine learning methods such as long short-term memory network or time convolution network, etc. During the model training process, the correlation of the data in the time dimension is used to fit the historical change trajectory, capture the typical time evolution pattern of the plateau vegetation in the region, and then infer the possible spectral value at the missing time point.

[0071] Considering that relying solely on time series data for spectral prediction may still introduce certain uncertainties, spatial information needs to be incorporated for correction. At the spatial level, unoccluded image patches surrounding the missing region can be selected as references to extract their spatial structure features and spectral distribution patterns. Then, spatial interpolation methods, such as bilinear interpolation, inverse distance weighting, or prediction methods based on local spatial clustering, are used to fine-tune the temporal prediction results. This step guides the predicted values ​​to better fit the existing spatial distribution pattern in the image at the current time, preventing the temporal prediction model from introducing systematic errors due to neglecting spatial context.

[0072] Finally, the predictions obtained through temporal modeling and spatial correction are recombined into complete image patches, which are then used to fill in missing areas in the original remote sensing image caused by cloud cover. These reconstructed areas are subsequently input along with other areas of the original image into subsequent spatial structure recognition and carbon sequestration function estimation models, achieving complete temporal and spatial compensation of image information and ensuring the continuity and integrity of the entire remote sensing monitoring chain. This method is particularly suitable for special scenarios such as plateau regions where image loss is frequent, and can significantly improve image availability and the spatiotemporal consistency of model output. Actual tests show that the areas filled by this reconstruction method have smaller errors in carbon sequestration estimation accuracy compared to the actual observed images, demonstrating practical application value.

[0073] Furthermore, a critical and complex challenge in remote sensing monitoring of carbon sequestration in plateau regions is the highly unstable climate, prone to sudden and extreme weather events. Such climate disturbances significantly impact the inherently fragile plateau vegetation system, particularly in terms of carbon absorption capacity, exhibiting high sensitivity and long recovery periods. However, traditional remote sensing models are typically based on assumptions of stable climate or uniform time series, neglecting the disruptive effects of sudden weather events on ecosystems. This leads to overly smooth or lagging model predictions, hindering effective early warning of ecological risks in practical applications and impeding accurate estimation of carbon sequestration dynamics in different regions under climate pressure. Therefore, introducing an extreme climate disturbance enhancement modeling mechanism into the monitoring scheme can more realistically simulate the highly variable environment of the ecosystem and improve its sensitivity to sudden conditions through model training, thereby enhancing the robustness and responsiveness of carbon sequestration predictions.

[0074] like Figure 3 As shown, the modeling process of this invention includes:

[0075] Construct a climate disturbance sample set containing historical meteorological data, including temperature, humidity, precipitation, wind speed, wind direction, and solar radiation, among other multi-dimensional climate factors.

[0076] A disturbance function is applied to the climate factor data to simulate short-term or sudden extreme climate events that may occur in plateau regions.

[0077] The disturbed climate samples are input into a disturbance enhancement training module, which trains a vegetation response prediction model in conjunction with the vegetation feature vectors extracted from the remote sensing images.

[0078] Based on the disturbance response score output by the model, the carbon sink stability of the vegetation area is adjusted and corrected, and the prediction result of the change in carbon sink function of the disturbance-sensitive area is output.

[0079] In the above process, more specifically:

[0080] The first step is to systematically build a sample set containing historical meteorological information, which includes multiple key climate factors such as daily average temperature, relative humidity, cumulative precipitation, wind speed, wind direction, and solar radiation. These factors not only fluctuate strongly in plateau environments, but also often determine whether vegetation can maintain normal photosynthesis efficiency and physiological activity levels. In the process of establishing the sample set, high-frequency observation data sources or resampling of low-frequency data should be prioritized to obtain more complete time series coverage.

[0081] After obtaining the original climate sample data, the next key is how to apply disturbances to these data to simulate the occurrence of extreme events. The disturbance is not random destruction, but a guided "functional deformation", such as introducing short-term drops in temperature sequences, inserting consecutive drought periods in precipitation, and superimposing strong wind fluctuations in wind speed sequences, to generate synthetic climate samples with extreme properties. Such samples not only closely resemble real observations, but also cover a wider range of climate variation types at a lower cost, allowing the subsequent training model to be exposed to a wider range of ecological stress scenarios based on limited meteorological observations.

[0082] The climate samples with disturbance characteristics are sent to the disturbance enhancement training module, which trains these climate inputs in conjunction with the vegetation feature vectors extracted from the remote sensing images. In this stage, the vegetation reflectance, multispectral features, texture features, and other content contained in the image data at each time point are integrated into a feature vector reflecting the ecological state of the ground surface as one of the inputs to the model, while the disturbed climate factors are used as parallel inputs. The goal of training is to enable the model to learn the response patterns of vegetation to different climate pressures, i.e., when encountering a certain type of meteorological disturbance condition, the growth condition, coverage, and carbon absorption capacity of a certain type of vegetation may change trajectory. After such joint training, the model not only masters the ecological response under stable climate, but also has strong generalization ability for stress response under various extreme climates.

[0083] After the model is trained, a disturbance response score is generated for each target area, reflecting the degree of carbon sink stability of the vegetation system in the area under current or predicted climate conditions. The lower the score, the more likely the area is to lose stability when encountering similar disturbances, and the carbon absorption function may decline rapidly. A higher score indicates that the region's vegetation has good resistance to disturbances, and the carbon sink function is more stable. After obtaining the score, the preliminary estimated carbon sink value can be corrected based on stability, so that the final output carbon sink function distribution map not only considers the current ecological state, but also integrates dynamic predictions of future disturbance scenarios, making the monitoring results more forward-looking and warning.

[0084] In practical applications, this mechanism can provide important decision-making basis for ecological managers. For example, in a plateau watershed where vegetation restoration projects are being implemented, the model can identify areas that are currently well-vegetated but are extremely sensitive to sudden rainfall reduction, and recommend that they be given special attention or increased irrigation support in future management. At the same time, areas with strong resistance can be used for stable carbon sink estimation, providing data support for carbon trading or carbon sink measurement systems. Overall, the extreme climate disturbance enhancement modeling mechanism injects dynamic perception and resilience analysis capabilities into carbon sink monitoring in complex ecological regions of the plateau, and is an important technical support means for achieving multi-scenario usability and coping with future uncertainties.

[0085] In summary, the present application proposes three main innovations: 1) a Transformer-based image semantic structure recognition model, 2) a time series-based image reconstruction mechanism, and 3) a carbon sink stability perception modeling based on extreme climate disturbance enhancement. The Transformer structure provides superior recognition capabilities for identifying spatial patterns and ecological semantic features compared to traditional methods. The time series reconstruction mechanism effectively addresses the problem of observation interruption caused by the special climate of the plateau. The extreme climate disturbance enhancement modeling mechanism provides the system with the ability to perceive and respond to future risks. These three technical innovations complement each other and form an intelligent carbon sink remote sensing monitoring system suitable for the plateau environment.

[0086] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the scope of the disclosed technology, and such changes should be covered within the protection scope of the present application.

Claims

1. A method for monitoring highland surface vegetation coverage and carbon sink function by remote sensing, characterized in that, The method comprises the following steps: acquiring a highland region surface image by using remote sensing images, dividing the remote sensing images into a plurality of image blocks, and extracting a multispectral feature vector of each image block; inputting the multispectral feature vector of the image block into a spatial-semantic structure recognition model constructed based on a Transformer structure, and establishing spatial dependency and semantic association between the image blocks by using a self-attention mechanism; generating a structure recognition result and a vegetation type classification result of the image block based on the spatial dependency and the semantic association; the structure recognition result comprises an identification result of whether the surface is in a complete grassland core area, an edge area, or a broken transition area; calculating a carbon sink function index of a corresponding image region according to the structure recognition result and the vegetation type classification result, and combining a carbon absorption parameter corresponding to the vegetation type and the structure; outputting a monitoring result comprising a surface vegetation coverage distribution map, a vegetation structure map, and a carbon sink function spatial distribution map; when the carbon sink function is monitored, a modeling process based on an extreme climate disturbance enhancement is introduced, and the modeling process comprises the following steps: constructing a climate disturbance sample set comprising historical meteorological data, wherein the meteorological data comprises any one or more of the following: temperature, humidity, precipitation, wind speed, wind direction, and solar radiation; applying a disturbance function to the meteorological data to simulate a short-term or sudden extreme climate event that may occur in the highland region; inputting the disturbed climate sample into a disturbance enhancement training module, and training the disturbance enhancement training module in combination with a vegetation feature vector extracted from the remote sensing image; adjusting and correcting a carbon sink value of a vegetation region based on a disturbance response score output by the disturbance enhancement training module, and outputting a carbon sink function change prediction result of a disturbance sensitive region.

2. The method of claim 1, wherein, When there is a region in the remote sensing image in which image information is missing due to cloud cover, a time series-based image reconstruction method is used for image completion.

3. The method of claim 2, wherein, The image reconstruction method comprises the following steps: performing image registration on historical multi-temporal remote sensing images of the missing region, and extracting multispectral feature vectors of corresponding regions at adjacent time nodes; constructing a spectral prediction model with time series as input, and predicting feature values at the missing time by using the time variation trend of the feature vectors; further correcting the predicted feature values by using a spatial interpolation algorithm in combination with the spatial features of the image blocks at the edges of the missing region; reconstructing the predicted results into complete image blocks, and replacing the original missing region for subsequent vegetation structure recognition and carbon sink function calculation.

4. The method of claim 1, wherein, The remote sensing images comprise multi-temporal multispectral remote sensing images, and each image contains red light, near-infrared, short-wave infrared, and red edge bands.

5. The method of claim 1, wherein, The multispectral feature vector comprises a combination of average reflectivity, standard deviation, normalized vegetation index, enhanced vegetation index, and chlorophyll red edge index of each image block at multiple bands.

6. The method of claim 1, wherein, The Transformer structure in the spatial-semantic structure recognition model comprises an image block embedding module, a multi-head self-attention module, a feedforward neural network module, and a position encoding module, which generate structure discrimination output by learning the context dependency of the image blocks in space and semantics.

7. The method of claim 1 wherein, The semantic association relationship between the image blocks is obtained by constructing an ecological semantic graph, and the ecological semantic graph is constructed based on plant type similarity, functional ecological zone adjacency and historical evolution trend.

8. The method of claim 1, wherein, The structure recognition result includes an ecological structure label, a structure integrity score and an ecological boundary type of each image block, and the vegetation type classification result includes any one or more of shrub, alpine meadow, plateau steppe, rocky bare land and permafrost region.

Images