Method and system for dynamic monitoring of urban green carbon sink based on multi-source remote sensing data

Abstract

The application discloses a method for monitoring urban green land carbon sink based on multi-source remote sensing data. The method comprises the following steps: obtaining spectral data, solar azimuth and elevation angle corresponding to the spectral data, meteorological data and vegetation state data; processing the spectral data by using a double-channel convolution model to obtain a vegetation parameter vector; processing the spectral data and the solar azimuth and elevation angle corresponding to the spectral data by using a photosynthetically active radiation dynamic model to obtain photosynthetically active radiation; and calculating carbon sink intensity by using a double-time-scale LSTM network according to the meteorological data, the photosynthetically active radiation, the vegetation state data and the vegetation parameter vector. The method has strong timeliness and high calculation accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data processing technology, and more specifically, to a method and system for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data. Background Technology

[0002] Nowadays, in order to strengthen the monitoring of the urban ecological environment, it is often necessary to monitor the carbon sink intensity (carbon flux) data of urban green spaces in real time.

[0003] Traditional methods typically employ the Monte Carlo radiative transfer model to calculate carbon sink intensity (carbon flux). However, the inversion process of the Monte Carlo radiative transfer model presents a contradiction between high computational load and real-time requirements. The Monte Carlo method approximates the problem through extensive stochastic simulations. When inverting PAR values, massive photon path tracing simulations are required to account for the effects of various atmospheric factors on solar radiation scattering and absorption. However, in the traditional serial calculation mode, the inversion process of the Monte Carlo radiative transfer model is too time-consuming, failing to meet the timeliness requirements of dynamic monitoring. Furthermore, because the data is constantly changing, the data used in traditional calculations does not update in a timely manner with environmental changes, nor does it consider the fusion of multiple data sources for comprehensive judgment, resulting in low calculation accuracy.

[0004] Existing technology discloses a phenological-based multi-source remote sensing method for estimating ecosystem carbon flux, comprising the following steps: using multi-source remote sensing data to calculate direct and scattered photosynthetically active radiation in the target area; acquiring regional biometric data of the same time series, extracting vegetation phenological characteristics using nonlinear least squares fitting with double S-curves, setting thresholds based on phenological characteristics, and constructing phenological constraint factors; establishing an ecosystem carbon flux model based on a two-leaf light energy utilization efficiency framework and in conjunction with the phenological constraint factors; and optimizing the model parameters using the Markov chain Monte Carlo method. The optimized model is then used to generate global gross primary productivity data. However, this method has low real-time performance and cannot meet the requirements for real-time calculation of carbon sink intensity. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies, such as poor timeliness and low calculation accuracy, by providing a method and system for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data. This method can calculate carbon sequestration intensity in real time, with strong timeliness and high accuracy.

[0006] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows:

[0007] Methods for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data include:

[0008] S1: Acquire spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data;

[0009] S2: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector;

[0010] S3: The spectral data, the solar azimuth angle and elevation angle corresponding to the spectral data are processed using the dynamic model of photosynthetically active radiation to obtain photosynthetically active radiation;

[0011] S4: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, the carbon sink intensity is calculated using a dual-timescale LSTM network.

[0012] Methods for dynamic monitoring of carbon sequestration in urban green spaces under extreme weather conditions include:

[0013] S01: Acquire spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data;

[0014] S02: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector;

[0015] S03: Using the Monte Carlo model, process the spectral data and the corresponding solar azimuth and elevation angles to obtain the first photosynthetically active radiation.

[0016] S04: The spectral data, the solar azimuth angle and elevation angle corresponding to the spectral data are processed using the dynamic model of photosynthetically active radiation to obtain the second photosynthetically active radiation;

[0017] S05: Integrate the first photosynthetically active radiation and the second photosynthetically active radiation to obtain photosynthetically active radiation;

[0018] S06: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, the carbon sink intensity is calculated using a dual-timescale LSTM network.

[0019] Furthermore, the spectral data includes visible light spectral data and near-infrared spectral data;

[0020] The dual-channel convolutional model includes: a first feature extraction channel, a second feature extraction channel, a first feature fusion unit, and a first output unit;

[0021] Visible light band spectral data is input to the input terminal of the first feature extraction channel, and near-infrared band spectral data is input to the input terminal of the second feature extraction channel. The output terminals of the first feature extraction channel and the second feature extraction channel are connected to the input terminal of the first feature fusion unit. The output terminal of the first feature fusion unit is connected to the input terminal of the first output unit. The output terminal of the first output unit outputs a vegetation parameter vector.

[0022] Furthermore, the first feature fusion unit includes: a first feature concatenation layer, a first convolutional layer, a batch normalization layer, a first activation layer, and a first global average pooling layer;

[0023] The outputs of the first feature extraction channel and the second feature extraction channel are connected to the input of the first feature splicing layer. The output of the first feature splicing layer is connected to the input of the first convolutional layer. The output of the first convolutional layer is connected to the input of the batch normalization layer. The output of the batch normalization layer is connected to the input of the first activation layer. The output of the first activation layer is connected to the input of the first global average pooling layer. The output of the first global average pooling layer is connected to the input of the first output unit.

[0024] Furthermore, the photosynthetically active radiation dynamic model includes: an input unit, a second feature extraction unit, an attention mechanism unit, a second feature fusion unit, and a second output unit;

[0025] Spectral data, the corresponding solar azimuth angle and elevation angle are input to the input unit. The output of the input unit is connected to the input of the second feature extraction unit. The output of the second feature extraction unit is connected to the input of the attention mechanism unit. The output of the attention mechanism unit is connected to the input of the second feature fusion unit. The output of the second feature fusion unit is connected to the input of the second output unit. The output of the second output unit outputs photosynthetically active radiation.

[0026] Furthermore, the second feature extraction unit includes: a second convolutional layer, a first max pooling layer, a third convolutional layer, a second max pooling layer, a fourth convolutional layer, and a third max pooling layer;

[0027] The output of the input unit is connected to the input of the second convolutional layer; the output of the second convolutional layer is connected to the input of the first max pooling layer; the output of the first max pooling layer is connected to the input of the third convolutional layer; the output of the third convolutional layer is connected to the input of the second max pooling layer; the output of the second max pooling layer is connected to the input of the fourth convolutional layer; the output of the fourth convolutional layer is connected to the input of the third max pooling layer; and the output of the third max pooling layer is connected to the input of the attention mechanism unit.

[0028] The attention mechanism unit includes: a second global average pooling layer, a first fully connected layer, a second activation layer, a third activation layer, a weight generation layer, and a feature recalibration layer.

[0029] The output of the second feature extraction unit is connected to the input of the second global average pooling layer. The output of the second global average pooling layer is connected to the input of the first fully connected layer. The output of the first fully connected layer is connected to the input of the second activation layer. The output of the second activation layer is connected to the input of the second fully connected layer. The output of the second fully connected layer is connected to the input of the third activation layer. The output of the third activation layer is connected to the input of the weight generation layer. The output of the weight generation layer is connected to the input of the feature recalibration layer. The output of the feature recalibration layer is connected to the input of the second feature fusion unit.

[0030] Furthermore, the dual-timescale LSTM network includes: a short-term LSTM unit, a long-term LSTM unit, and a feature fusion output unit;

[0031] Meteorological data and photosynthetically active radiation are input to the input terminal of the short-term LSTM unit, vegetation status data and vegetation parameter vectors are input to the input terminal of the long-term LSTM unit, and the output terminals of the short-term LSTM unit and the long-term LSTM unit are connected to the input terminal of the feature fusion output unit; the output terminal of the feature fusion output unit outputs carbon sink intensity.

[0032] Furthermore, the short-term LSTM unit includes: a first preprocessing layer, a first LSTM layer, a first dropout layer, a second LSTM layer, and a first attention layer;

[0033] Meteorological data and photosynthetically active radiation are input to the input end of the first preprocessing layer. The output end of the first preprocessing layer is connected to the input end of the first LSTM layer. The output end of the first LSTM layer is connected to the input end of the first dropout layer. The output end of the first dropout layer is connected to the input end of the second LSTM layer. The output end of the second LSTM layer is connected to the input end of the first attention layer. The output end of the first attention layer is connected to the input end of the feature fusion output unit.

[0034] The long-term LSTM unit includes: a second preprocessing layer, a third LSTM layer, a second dropout layer, a fourth LSTM layer, and a second attention layer;

[0035] Vegetation state data and vegetation parameter vectors are input to the input of the second preprocessing layer. The output of the second preprocessing layer is connected to the input of the third LSTM layer. The output of the third LSTM layer is connected to the input of the second dropout layer. The output of the second dropout layer is connected to the input of the fourth LSTM layer. The output of the fourth LSTM layer is connected to the input of the second attention layer. The output of the second attention layer is connected to the input of the feature fusion unit.

[0036] Furthermore, the feature fusion output unit includes: a second feature splicing layer, a third attention layer, a third fully connected layer, a fourth activation layer, and a fourth fully connected layer;

[0037] The outputs of the short-term LSTM unit and the long-term LSTM unit are connected to the input of the second feature splicing layer. The output of the second feature splicing layer is connected to the input of the third attention layer. The output of the third attention layer is connected to the input of the third fully connected layer. The output of the third fully connected layer is connected to the input of the fourth activation layer. The output of the fourth activation layer is connected to the input of the fourth fully connected layer. The output of the fourth fully connected layer outputs the carbon sink intensity.

[0038] A dynamic monitoring system for urban green space carbon sequestration based on multi-source remote sensing data includes:

[0039] Data acquisition module: Acquires spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data;

[0040] Vegetation parameter vector module: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector;

[0041] Photosynthetically Active Radiation Module: The spectral data and the corresponding solar azimuth and elevation angles are processed using a dynamic model of photosynthetically active radiation to obtain photosynthetically active radiation;

[0042] Carbon sink intensity calculation module: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, the carbon sink intensity is calculated using a dual-timescale LSTM network.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention calculates carbon sink intensity using a dual-timescale LSTM network, thereby automatically adjusting the data weights of meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vectors based on the differences in current data, ensuring that the output results are flexible and robust in response to extreme weather, seasonal changes, or rapid urban disturbances.

[0045] This invention achieves efficient calculation of photosynthetically active radiation by using a dynamic model of photosynthetically active radiation, which ensures spatial continuity while effectively suppressing local jumps in the output photosynthetically active radiation.

[0046] This invention achieves the collaborative extraction of vegetation texture features and spectral features through a dual-channel convolution model.

[0047] In summary, this method is characterized by its high computational efficiency and high accuracy. Attached Figure Description

[0048] Figure 1 The flowchart is for the dynamic monitoring method of urban green space carbon sink based on multi-source remote sensing data provided in Example 1.

[0049] Figure 2 This is a schematic diagram of data acquisition provided in Example 1.

[0050] Figure 3 A schematic diagram of the data preprocessing process provided in Example 1.

[0051] Figure 4 The diagram shows the structure of the dual-channel convolution model provided in Example 1.

[0052] Figure 5 This is a schematic diagram of the structure of the dual-channel convolution model provided in Example 1.

[0053] Figure 6 The structure diagram of the first feature fusion unit provided in Example 1.

[0054] Figure 7 This is a structural diagram of the dynamic model of photosynthetically active radiation provided in Example 1.

[0055] Figure 8 This is a schematic diagram of the dynamic model of photosynthetically active radiation provided in Example 1.

[0056] Figure 9 The structure diagram of the second feature extraction unit provided in Example 1.

[0057] Figure 10 This is a structural diagram of the attention unit provided in Example 1.

[0058] Figure 11 The structure diagram of the dual-timescale LSTM network provided in Example 1 is shown.

[0059] Figure 12 This is a schematic diagram of the dual-timescale LSTM network provided in Example 1.

[0060] Figure 13 The structure diagram of the short-term LSTM unit provided in Example 1.

[0061] Figure 14 The structure diagram of the long-term LSTM cell provided in Example 1.

[0062] Figure 15 The structure diagram of the feature fusion output unit provided in Example 1. Detailed Implementation

[0063] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0064] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions;

[0065] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0066] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0067] Example 1

[0068] like Figure 1 As shown, the method for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data includes:

[0069] S1: Acquire spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data;

[0070] S2: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector;

[0071] S3: The spectral data, the solar azimuth angle and elevation angle corresponding to the spectral data are processed using the dynamic model of photosynthetically active radiation to obtain photosynthetically active radiation;

[0072] S4: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, the carbon sink intensity is calculated using a dual-timescale LSTM network.

[0073] In one specific embodiment, satellite remote sensing utilizes high-resolution optical satellites, such as the Gaofen-2 satellite, which possesses sub-meter spatial resolution and can acquire clear images of urban green spaces. The imaging swath is approximately 45 kilometers wide, effectively covering a large urban area. The data acquisition frequency is set to once a month to meet the phased monitoring needs of the overall distribution and changes in urban green spaces. Unmanned aerial vehicle (UAV) remote sensing employs small UAVs equipped with hyperspectral imagers, such as the DJI Phantom 4 RTK with its MicaSense RedEdge-MX hyperspectral camera. This camera can acquire hyperspectral data in five bands with a spectral resolution of approximately 10 nm, enabling precise identification of the type and condition of green vegetation. The flight altitude is controlled between 500-1000 meters, and the flight path and range are flexibly adjusted according to the city size and key monitoring areas to ensure that the acquired hyperspectral data is comprehensive and targeted. The ground-based sensing unit deploys various types of sensors, including a temperature and humidity sensor (DHT11, accuracy ±2% RH, ±0.5℃), a light intensity sensor (BH1750, measurement range 1-65535 lx), and a CO2 concentration sensor (S8-CO2, accuracy ±3% FS), evenly distributed throughout the urban green space. A sensing node is placed every 100-200 meters to collect environmental parameters in real time, with a data acquisition frequency of once per minute, to promptly capture dynamic changes in environmental parameters. Spectral data includes hyperspectral and multispectral data.

[0074] like Figure 2 , Figure 3As shown in one specific embodiment, spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data can be jointly acquired by satellite remote sensing units, UAV remote sensing units, and ground sensing units. Satellite remote sensing units acquire visible and near-infrared images with sub-meter spatial resolution for large-scale green space distribution and status identification. UAV remote sensing units acquire hyperspectral data to enhance vegetation type differentiation capabilities. Ground sensing units simultaneously collect environmental parameters such as temperature, humidity, solar radiation, and leaf surface humidity, providing basic physical quantities for model input and verification. Since the data may have various structures and temporal representations, they need to be processed after acquisition using geometric correction algorithms based on feature point matching and dynamic time warping algorithms to achieve a unified spatiotemporal reference for multi-source data. The satellite-UAV data registration unit uses the SIFT feature matching algorithm to extract at least 500 feature points, and then applies the RANSAC algorithm to remove mismatched points, thereby achieving a registration accuracy better than 0.5 pixels and ensuring accurate spatial alignment of satellite and UAV data. The air-to-ground data fusion unit constructs a three-dimensional Gaussian process regression model to spatially interpolate and fuse point observation data from ground sensors with remote sensing area data, enabling different types of data to complement each other and generate a more comprehensive and accurate dataset. The time synchronization unit designs a sliding time window algorithm to dynamically adjust the observation period of data with different time resolutions and establishes a unified timestamp index system to ensure the continuity and comparability of data in the time dimension.

[0075] The specific alignment method is as follows: A geometric correction algorithm based on feature point matching and a dynamic time warping algorithm are used to unify the spatiotemporal reference of multi-source data. Specifically, after data acquisition, the data enters the spatiotemporal registration module. First, geometric correction is performed on the image. Then, a feature point matching algorithm based on scale-invariant feature transform (SIFT) is used to extract key points. Finally, the homography transformation matrix H is fitted using the minimum mean square error criterion to complete image registration. The mapping relationship is expressed as follows:

[0076]

[0077] Where (x, y) are the original image coordinates, (x′, y′) are the registered coordinates, and H is a 3×3 homography matrix. For temporal registration, a Dynamic Time Warping (DTW) algorithm is introduced to align data at different time scales. By calculating the minimum distance path between each data sequence, the remote sensing observations are synchronized with ground parameters. The cost function for the optimal matching path is defined as:

[0078]

[0079] in and These are the observations from different data sources at time i and time j, respectively. To accumulate the shortest path cost, the final output is a unified time series data combination.

[0080] Taking the spatiotemporal registration of satellite remote sensing data and ground sensor data as an example:

[0081] 1) Data Acquisition

[0082] Satellite remote sensing data: Gaofen-2 satellite imagery (sub-meter resolution), containing visible light band data of green area distribution.

[0083] Ground sensor data: The temperature and humidity sensor at a green space monitoring point collects data once per minute. The temperature sequence from 9:50 to 10:10 on a certain day is: [25.3℃, 25.5℃, 25.6℃, 25.8℃, 26.1℃, 26.3℃, 26.2℃, 26.0℃, 25.9℃, 25.7℃, 25.6℃, 25.5℃] (a total of 12 time points, denoted as sequence T).

[0084] 2) Geometric correction (spatial registration) steps:

[0085] ①SIFT feature point extraction: SIFT feature points are extracted from green areas in satellite imagery and aerial reference images near ground sensors.

[0086] 800 feature points (tree canopy edges, road corners, etc.) were extracted from satellite imagery, and 600 feature points were extracted from UAV reference images.

[0087] Feature point matching: The KD tree algorithm is used for fast matching, and 300 pairs of matching points are initially obtained, such as the tree canopy vertex A (100, 200) in the satellite image and the corresponding point A' (150, 250) in the UAV image.

[0088] ② RANSAC algorithm removes mismatches

[0089] The homography matrix H is calculated by randomly sampling 4 pairs of matching points. The coordinates of the matching points obtained in a certain sampling are:

[0090] Point 1: (x=50,y=100)→(x'=80,y'=120)

[0091] Point 2: (x=200,y=50)→(x'=230,y'=70)

[0092] Point 3: (x=150,y=150)→(x'=180,y'=170)

[0093] Point 4: (x=300,y=200)→(x'=330,y'=220)

[0094] Calculate the H matrix:

[0095]

[0096] All matching points were verified using H, and points with errors exceeding 0.5 pixels were removed, ultimately retaining 250 pairs of valid matching points, with the registration accuracy meeting the requirements.

[0097] ③ Application of coordinate transformation

[0098] The coordinates of a pixel (x=250, y=180) in a satellite image, after H transformation:

[0099]

[0100] That is, the registered coordinates are (298, 225), which are consistent with the coordinates of the UAV image.

[0101] 3) Time registration derivation using the Dynamic Time Warping (DTW) algorithm

[0102] ① Definition of data sequence:

[0103] The satellite imagery was acquired at 10:00, denoted as time t=0. The timestamps for the ground temperature sequence T are from 9:50 to 10:10, with each window lasting 10 minutes, simplified to three time points:

[0104] Sequence S, satellite corresponding time window: [10:00], only 1 point, satellite inverted temperature 26.0℃.

[0105] Sequence T (Ground sensor): [9:50 (25.6℃), 10:00 (26.1℃), 10:10 (25.9℃)]

[0106] ② Cost function calculation:

[0107] Construct a cost matrix D, with rows and columns representing time points S and T respectively:

[0108] ③ Calculate the cost of a single point:

[0109] D(1,1) = |26.0 - 25.6| + min{∞, ∞, ∞} = 0.4 + 0 = 0.4

[0110] D(1,2) = |26.0 - 26.1| + min{D(1,1), ∞, D(0,1)} = 0.1 + 0.4 = 0.5

[0111] D(1,3) = |26.0 - 25.9| + min{ D(1,2), D(0,3), D(0,2)} = 0.1 + 0.5 =0.6

[0112] DTW calculates the cumulative cost recursively, ensuring that each choice is based on the global optimum, not a local single-point optimum. The essence of time series matching requires not only small numerical differences but also consistency in temporal order. In the case study, although D(1,2) has a slightly higher single-point cost, it matches times with the same timestamp (10:00 vs 10:00); the temperature difference (0.1℃) is less than the 0.4℃ difference of D(1,1), and the path is more consistent with temporal logic (no jumps). The optimal path is D(1,2), matching the satellite time 10:00 with the ground time 10:00, with the minimum cost of 0.5.

[0113] ④ Time series alignment results

[0114] The temperature of 26.1℃ at 10:00 in the ground sensor data is aligned with the temperature of 26.0℃ in the satellite imagery to form a dataset with a unified timestamp (10:00) for subsequent analysis.

[0115] 4) Summary of spatiotemporal registration results

[0116] Spatial dimension: Satellite imagery is registered with ground reference data using SIFT+RANSAC, with a pixel error of <0.5, such as coordinates (250,180)→(298,225), to ensure spatial consistency.

[0117] Time dimension: The DTW algorithm matches the temperature data at satellite time 10:00 with the temperature data at ground time 10:00, eliminating time scale differences, and finally outputs a dataset with a unified spatiotemporal reference (correlation between satellite temperature of 26.0℃ and ground temperature of 26.1℃ at 10:00).

[0118] The air-to-ground data fusion unit employs a 3D Gaussian process regression model to achieve spatial scale alignment between remote sensing imagery and ground sensor observation data. The main task of this unit is to map point-level environmental parameters such as temperature, humidity, and carbon dioxide concentration acquired by ground sensors to the same spatial resolution and raster structure as the remote sensing imagery through probabilistic inference, thereby constructing a composite dataset with strong spatial continuity and spatiotemporal consistency. During the fusion process, the ground sensor data is first temporally aligned and spatially standardized, embedded into the coordinate grid covering the remote sensing imagery area, and training samples for the Gaussian process model are constructed based on the geographic coordinates and observation values ​​of known points. This regression model predicts the environmental parameters of any target raster point in 3D space, and its predicted mean is expressed as:

[0119]

[0120] in Location to be predicted The estimated value, Let y be the kernel function vector between this location and the known sensor locations, K be the kernel matrix between all known sensors, and y be the corresponding observation vector. To observe the noise variance, I is the identity matrix. The kernel function used is constructed on the Euclidean distance in three-dimensional space, and is often in the form of an isotropic Gaussian kernel function:

[0121]

[0122] in and Let be the position coordinates of any two sensors or prediction points, and l be the scale factor controlling the decay range of the kernel function. In this way, the model not only achieves seamless spatial fusion of remote sensing images and ground sensors, but also dynamically captures the changing trends of the local microenvironment. This ensures that the fused composite dataset possesses the same resolution and spatial distribution characteristics as the remote sensing image in variables such as temperature, humidity, and CO2 concentration, providing high-quality auxiliary input for the subsequent carbon sequestration dynamic modeling module.

[0123] It should be noted that the geometric correction algorithm based on feature point matching employs the SIFT (Scale Invariant Feature Transform) algorithm to extract feature points from satellite and UAV remote sensing images. First, a scale space is constructed for the images, and a Difference of Gaussian Pyramid (DoG) is used to detect stable feature points. Then, a 128-dimensional descriptor for each feature point is calculated to describe the gradient information of the surrounding region. The KD-tree algorithm is used for fast feature point matching to obtain preliminary matching point pairs. Next, the RANSAC (Random Sample Consensus) algorithm is applied, randomly sampling a subset of samples from the preliminary matching point pairs to calculate the homography matrix. The homography matrix is ​​then used to verify other matching points, eliminating mismatched points, ultimately achieving a registration accuracy better than 0.5 pixels.

[0124] Dynamic time warping algorithms are used in time synchronization units to handle time alignment issues for data with different time resolutions. For satellite remote sensing data (low time resolution, acquired monthly), UAV remote sensing data (moderate time resolution, acquired per flight), and ground-based sensor data (high time resolution, acquired every minute), a common time reference is first determined, such as the hour. For ground-based sensor data, a sliding time window algorithm is used to aggregate the data collected every minute according to time windows, such as one hour. The average value or other statistical values ​​of the data within the time window are calculated to match the time resolution of the satellite and UAV remote sensing data. Then, based on the timestamps, data from different sources are aligned on the time axis, establishing a unified timestamp index system to ensure the consistency and comparability of data in the time dimension.

[0125] In the satellite-UAV data registration unit, after achieving accurate registration using SIFT and RANSAC algorithms, the registration results still need to be evaluated for accuracy. A certain number of ground control points (GCPs) can be selected, and their precise coordinates obtained through field measurements. Then, the coordinates of corresponding points on the registered image are compared with the actual coordinates, and the root mean square error (RMSE) is calculated to verify whether the registration accuracy meets the requirement of being better than 0.5 pixels. In the air-to-ground data fusion unit, when constructing a 3D Gaussian process regression model, point-like temperature, humidity, and CO2 concentration data from ground sensors are used as training data, and pixel values ​​from remote sensing images are used as prediction targets. First, the kernel function of the Gaussian process is determined, using a radial basis function (RBF) kernel to measure the similarity between data points. Then, the hyperparameters of the kernel function, such as length scale and variance, are estimated using the maximum likelihood estimation method. The trained model is then used to perform spatial interpolation fusion on the remote sensing image, expanding the point data from the ground sensors into area data with the same resolution as the remote sensing image, generating a composite dataset with consistent resolution, providing more comprehensive data support for subsequent analysis.

[0126] Methods for dynamic monitoring of carbon sequestration in urban green spaces under extreme weather conditions include:

[0127] S01: Acquire spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data;

[0128] S02: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector;

[0129] S03: Using the Monte Carlo model, process the spectral data and the corresponding solar azimuth and elevation angles to obtain the first photosynthetically active radiation.

[0130] S04: The spectral data, the solar azimuth angle and elevation angle corresponding to the spectral data are processed using the dynamic model of photosynthetically active radiation to obtain the second photosynthetically active radiation;

[0131] S05: Integrate the first photosynthetically active radiation and the second photosynthetically active radiation to obtain photosynthetically active radiation;

[0132] S06: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, the carbon sink intensity is calculated using a dual-timescale LSTM network.

[0133] It should be noted that the meteorological monitoring unit continuously receives real-time meteorological data collected by ground-based micro-meteorological sensors deployed in urban green spaces, focusing on rainfall intensity and visibility. When rainfall intensity exceeds 50 mm per hour or visibility falls below 1 km, the system automatically determines that there is severe atmospheric disturbance in the current area, triggering a verification mechanism to enter local correction mode. At this time, the dynamic model of photosynthetically active radiation (PARF) may have systematic biases in its predictions under abnormal meteorological conditions. Therefore, the local correction unit activates the Monte Carlo model for the determined area, significantly reducing the number of photon tracking attempts to compress computational costs. The number of photon tracking attempts in the original model, which was in the millions, is reduced to one-tenth, retaining only the simulation of the main scattering paths, thus quickly completing local correction. The Monte Carlo model uses statistical sampling to simulate the scattering and absorption process of photons between the atmosphere and the ground surface. Its PARF value is expressed as:

[0134]

[0135] Where M is the number of photons under simplified sampling. The energy represented by each photon, Its transmittance during transmission, This represents the percentage of PAR absorbed by vegetation on the land surface. The corrected PAR value forms a high-confidence cover map in the local area, and then a weighted fusion method is used to embed it into the global PAR distribution map generated by the original proxy model. The fusion expression is:

[0136]

[0137] in The value is the predicted value of photosynthetically active radiation (PAR) from the dynamic model. α is the fusion weight, dynamically set according to the degree of meteorological disturbance, typically between 0.6 and 0.9. This fusion strategy maximizes the preservation of the global efficiency of the surrogate model while ensuring local prediction accuracy, achieving dynamic adaptive correction of the PAR distribution map under extreme weather conditions. This significantly improves the robustness of the system in abnormal scenarios and the reliability of carbon sink monitoring data.

[0138] Furthermore, such as Figure 4 , Figure 5 As shown, the spectral data includes visible light spectral data and near-infrared spectral data;

[0139] The dual-channel convolutional model includes: a first feature extraction channel, a second feature extraction channel, a first feature fusion unit, and a first output unit;

[0140] Visible light band spectral data is input to the input terminal of the first feature extraction channel, and near-infrared band spectral data is input to the input terminal of the second feature extraction channel. The output terminals of the first feature extraction channel and the second feature extraction channel are connected to the input terminal of the first feature fusion unit. The output terminal of the first feature fusion unit is connected to the input terminal of the first output unit. The output terminal of the first output unit outputs a vegetation parameter vector.

[0141] It should be noted that, through the first feature fusion unit, this module dynamically adjusts the contribution of the two branches through a learnable weight matrix, and finally outputs a quantitative result containing 12 vegetation parameters, including vegetation coverage, leaf area index, etc., which provides key vegetation information for subsequent carbon sink calculation.

[0142] In one specific embodiment, the two channels (feature extraction channel and spatiotemporal feature channel) each contain multiple convolutional and pooling layers. The outputs of the two branches are multi-dimensional feature maps, which then enter a feature fusion unit. The extracted high-dimensional features are concatenated and normalized in the first feature fusion unit to jointly represent the vegetation spatial information of the input image. Dynamic weighted integration is achieved through a learnable weight matrix, and the fusion operation is defined as follows:

[0143]

[0144] in and These are the feature maps for the visible light and near-infrared branches, respectively. W1 and W2 are adaptively updated fusion weight matrices during training, used to adjust the contribution of the two branches' features in different spatial regions. The fused features are then... The joint feature map, which serves as the final representation of vegetation status, is input into the fully connected layer for quantitative regression output of 12 types of vegetation parameters. These parameters include biophysical variables such as vegetation cover, leaf area index, chlorophyll content, specific leaf mass, and canopy height.

[0145] The calculation of vegetation cover (FVC) is based on the inversion of the normalized vegetation index (NDVI) and an empirical model. Its expression is as follows:

[0146]

[0147] NDVI is the calculated pixel value. and These represent the typical NDVI values ​​for full vegetation cover and bare soil, respectively. The Leaf Area Index (LAI) is directly regressed by a CNN and trained by comparing it with ground sample data to ultimately form a vegetation feature extraction model with end-to-end predictive capabilities.

[0148] The output uses linear regression units to map features to physical quantities, and the prediction process can be expressed as:

[0149]

[0150] in This is the predicted vegetation parameter vector. is the output weight matrix, and b is the bias term. This dual-channel structure achieves high-dimensional and refined vegetation state extraction by fusing spatial texture and spectral response, providing a stable and reliable input parameter foundation for subsequent carbon sink modeling.

[0151] It should be noted that in the first feature fusion unit, the contribution of the two branches is dynamically adjusted through a learnable weight matrix. This weight matrix is ​​initialized with random values ​​and updated during training using the backpropagation algorithm. The training dataset collects a large number of remote sensing images of urban green spaces, including images from different seasons and weather conditions, and labels parameters such as the actual vegetation cover and leaf area index in the images. The number of training iterations is set to 50-100, using a stochastic gradient descent (SGD) optimizer with a learning rate initialized to 0.001. An exponential decay strategy is used during training, multiplying the learning rate by 0.9 every 10 training iterations to ensure more stable convergence of the model in the later stages of training. During training, data augmentation techniques can be used to prevent overfitting, such as rotating, scaling, and flipping the images to increase the diversity of the training data. Simultaneously, L2 regularization is used to constrain the model's weights, reducing model complexity and improving its generalization ability.

[0152] Furthermore, such as Figure 6 As shown, the first feature fusion unit includes: a first feature concatenation layer, a first convolutional layer, a batch normalization layer, a first activation layer, and a first global average pooling layer;

[0153] The outputs of the first feature extraction channel and the second feature extraction channel are connected to the input of the first feature splicing layer. The output of the first feature splicing layer is connected to the input of the first convolutional layer. The output of the first convolutional layer is connected to the input of the batch normalization layer. The output of the batch normalization layer is connected to the input of the first activation layer. The output of the first activation layer is connected to the input of the first global average pooling layer. The output of the first global average pooling layer is connected to the input of the first output unit.

[0154] Furthermore, such as Figure 7 , Figure 8 As shown, the dynamic model of photosynthetically active radiation includes: an input unit, a second feature extraction unit, an attention mechanism unit, a second feature fusion unit, and a second output unit;

[0155] Spectral data, the corresponding solar azimuth angle and elevation angle are input to the input unit. The output of the input unit is connected to the input of the second feature extraction unit. The output of the second feature extraction unit is connected to the input of the attention mechanism unit. The output of the attention mechanism unit is connected to the input of the second feature fusion unit. The output of the second feature fusion unit is connected to the input of the second output unit. The output of the second output unit outputs photosynthetically active radiation.

[0156] In one specific embodiment, a PAR simulation dataset of urban scenes is generated based on the Monte Carlo radiative transfer model. Multispectral remote sensing data and solar geometric parameters are used as inputs, and a photosynthetically active radiation dynamic model is trained under supervision.

[0157] First, a high-quality training sample is constructed using a dataset generation unit. The Monte Carlo radiative transfer model is then used to simulate the light radiative transfer of typical urban land cover types under different combinations of solar altitude angles, atmospheric conditions, and surface reflectivity parameters, generating spatial distribution maps of photosynthetically active radiation values. These simulation results, combined with multispectral data (including visible and near-infrared bands) and solar geometric parameters from remote sensing imagery at the corresponding time points, constitute supervised training samples for neural network modeling.

[0158] To improve the model's response under complex atmospheric conditions, a channel attention mechanism is embedded in the network structure. Feature recalibration is used to enhance the channel weights of aerosol optical thickness and cloud occlusion factor, thereby increasing the contribution of key features to PAR prediction. This attention mechanism calculates the response weight vectors of each channel to suppress redundant channels and strengthen effective channels. The specific calculation process is as follows:

[0159]

[0160] Where z is the global average pooling result of the channel features, W1 and W2 are the weight matrices of the two fully connected layers, δ is the ReLU activation function, σ is the Sigmoid function, and the output weight w c The channels are recalibrated by multiplying point-by-point with the original channel features. Model optimization uses mean squared error as the loss function, defined as:

[0161]

[0162] in, Let be the PAR value predicted by the model for the i-th pixel. Where N is the total number of training samples and the corresponding true value is used, the weights are iteratively updated using the Adam optimizer, and the learning rate and gradient direction are adaptively adjusted to accelerate convergence. Training terminates and the surrogate model is solidified for subsequent online inference unit deployment when the model's prediction error on the validation set is continuously less than 5% during training. The entire process fully integrates physical modeling and data-driven methods, ensuring both inference efficiency and prediction accuracy and generalization ability.

[0163] In the workflow of the dynamic model inference of photosynthetically active radiation, the input preprocessing unit first standardizes the real-time acquired multispectral remote sensing data to ensure it conforms to the input format and physical meaning of the surrogate model. This step includes atmospheric correction and radiometric calibration, used to eliminate interference from atmospheric scattering and absorption on the remote sensing signal and to convert digital values ​​into ground reflectance, respectively. Atmospheric correction uses the 6S radiative transfer model and derives the surface reflectance based on the observation conditions at the time, where the formula for calculating reflectance R is:

[0164]

[0165] Where L is the radiance received by the sensor, L p For path radiation, E s θ represents the solar incident irradiance in the upper atmosphere. s T is the solar zenith angle. s With T v Atmospheric transmittance from the sun to the Earth's surface and from the Earth's surface to the sensor, respectively, and reflectance R are used for accurate input to the subsequent model. After preprocessing, the surrogate model inference unit inputs the corrected multispectral image data and corresponding solar azimuth parameters into the trained convolutional neural network. This network structure has been optimized into a lightweight model during the training phase, capable of generating a spatial distribution map of photosynthetically active radiation (PAR) over the entire city within 10 minutes, while maintaining a prediction error within 5%. The output PAR image maintains the same spatial resolution as the original image. To meet the requirement of a uniform spatial scale for urban carbon sink modeling, the post-processing unit performs spatial interpolation and resampling on the predicted image, using a bilinear interpolation algorithm to smoothly map the PAR values ​​between irregular pixels to a uniform 100-meter grid cell. The interpolated PAR value P is calculated at the center of each target grid cell as follows:

[0166]

[0167] in The PAR values ​​are the four neighboring pixels in the original image. The spatial weighting coefficients relative to the center of the target grid satisfy... This method effectively suppresses local jumps in the predicted image while ensuring spatial continuity, ensuring that the data can be directly used in the downstream carbon flux dynamic modeling module. The entire process design realizes efficient conversion from real-time remote sensing input to standardized PAR output, providing spatiotemporally consistent and high-precision radiation input guarantee for urban green space carbon sink estimation;

[0168] In one specific embodiment, the Monte Carlo radiative transfer model is used to simulate the PAR values ​​of urban scenes under different combinations of solar altitude angles (0°-90°, with a simulation point set every 5°), atmospheric conditions (including different aerosol optical thicknesses, ranging from 0.1 to 1.0, with a simulation point set every 0.1; and different cloud cover conditions, divided into no clouds, thin clouds, and thick clouds), and surface reflectance (ranging from 0.1 to 0.5, with a simulation point set every 0.05). A large-scale training dataset is constructed, containing at least 10,000 samples. The neural network architecture of the surrogate model adopts a lightweight CNN. The input layer receives multispectral data (8 bands in total) in the visible and near-infrared bands, as well as solar azimuth and altitude angle parameters (2 parameters in total). These input data are concatenated into a feature vector and input into the network. The network comprises multiple convolutional and pooling layers. The convolutional layers use 3×3 kernels with a stride of 1 and padding of 1 to extract data features. The pooling layers employ max pooling with a 2×2 kernel and a stride of 2 to reduce the resolution of the feature maps. The output layer is mapped to PAR values ​​via fully connected layers. An attention mechanism is applied in the embedded channel attention module (the Squeeze-Excitation module in SE-Net). This module first performs global average pooling on the feature map in the spatial dimension, resulting in a 1×1×C vector (where C is the number of channels). Then, two fully connected layers perform a non-linear transformation on this vector to obtain the weight coefficients for each channel. Finally, the weight coefficients are multiplied by the original feature map to achieve dynamic weighting of features from different channels, enabling the network to adaptively focus on key influencing factors such as aerosol optical thickness and cloud occlusion factors. The model optimization step uses the mean squared error loss function (MSE) to measure the error between the PAR value predicted by the surrogate model and the actual PAR value simulated by the Monte Carlo radiative transfer model. The adaptive momentum optimizer (Adam) is used for training, with an initial learning rate of 0.001, β1=0.9, and β2=0.999. The loss function value is monitored during training. When the loss function value decreases by less than 0.001 for 5 consecutive training iterations, the model is considered to have converged and training is stopped. At this point, the prediction error of the photosynthetically active radiation dynamic model is less than 5%.

[0169] In the inference phase, the dynamic model for photosynthetically active radiation (PAR) first performs atmospheric correction on the real-time acquired multispectral remote sensing data. Employing the 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) model, it corrects the radiance values ​​of the images based on sensor parameters, imaging time, and geographical location information, considering the scattering and absorption of solar radiation by atmospheric gas molecules and aerosols. Then, radiometric calibration is performed, converting the image's digital quantization (DN) values ​​into physical radiance values ​​for subsequent analysis. The preprocessed data is input into the trained surrogate model for inference. Through forward propagation calculations, a full-scene PAR distribution map is directly output, significantly reducing the computation time per calculation and meeting the system's real-time requirements. Post-processing optimizes the PAR distribution map through spatial interpolation using a bilinear interpolation algorithm, interpolating the PAR values ​​to a 100m × 100m raster resolution to match the spatial resolution of other data in the system, facilitating subsequent carbon flux calculation and analysis.

[0170] Furthermore, such as Figure 9 As shown, the second feature extraction unit includes: a second convolutional layer, a first max pooling layer, a third convolutional layer, a second max pooling layer, a fourth convolutional layer, and a third max pooling layer.

[0171] The output of the input unit is connected to the input of the second convolutional layer; the output of the second convolutional layer is connected to the input of the first max pooling layer; the output of the first max pooling layer is connected to the input of the third convolutional layer; the output of the third convolutional layer is connected to the input of the second max pooling layer; the output of the second max pooling layer is connected to the input of the fourth convolutional layer; the output of the fourth convolutional layer is connected to the input of the third max pooling layer; and the output of the third max pooling layer is connected to the input of the attention mechanism unit.

[0172] like Figure 10 As shown, the attention mechanism unit includes: a second global average pooling layer, a first fully connected layer, a second activation layer, a third activation layer, a weight generation layer, and a feature recalibration layer.

[0173] The output of the second feature extraction unit is connected to the input of the second global average pooling layer. The output of the second global average pooling layer is connected to the input of the first fully connected layer. The output of the first fully connected layer is connected to the input of the second activation layer. The output of the second activation layer is connected to the input of the second fully connected layer. The output of the second fully connected layer is connected to the input of the third activation layer. The output of the third activation layer is connected to the input of the weight generation layer. The output of the weight generation layer is connected to the input of the feature recalibration layer. The output of the feature recalibration layer is connected to the input of the second feature fusion unit.

[0174] Furthermore, such as Figure 11 , Figure 12 As shown, the dual-timescale LSTM network includes: a short-term LSTM unit, a long-term LSTM unit, and a feature fusion output unit.

[0175] Meteorological data and photosynthetically active radiation are input to the input terminal of the short-term LSTM unit, vegetation status data and vegetation parameter vectors are input to the input terminal of the long-term LSTM unit, and the output terminals of the short-term LSTM unit and the long-term LSTM unit are connected to the input terminal of the feature fusion output unit; the output terminal of the feature fusion output unit outputs carbon sink intensity.

[0176] It should be noted that a dual-timescale LSTM neural network structure is employed to achieve high-frequency, dynamic prediction of urban green space carbon sequestration intensity. This structure consists of two parallel LSTM sub-networks. One sub-network specifically processes short-term meteorological data, with inputs consisting of hourly sequences of meteorological parameters such as photosynthetically active radiation (PAR), air temperature and humidity, wind speed, and precipitation intensity collected over the past 24 hours. This is used to capture the real-time regulatory effects of rapidly changing environmental factors on carbon absorption capacity. The other sub-network processes long-term vegetation data, with inputs consisting of key physiological parameters such as daily average leaf area index, chlorophyll content, and vegetation cover over 30 consecutive days. This reflects the evolutionary trend of vegetation growth status and the underlying regulatory effects on carbon fixation potential. After deep temporal features are extracted from the data at each timescale using their respective LSTM units, feature fusion is performed using an attention mechanism. The system automatically assigns weights to short-term and long-term features based on the differences in the current time-series data, ensuring the output results are flexible and robust in response to extreme weather, seasonal changes, or rapid urban disturbances. Taking a city's green space as an example, during the growing season with several consecutive days of ample sunshine and moderate rainfall, the LSTM output of long-term vegetation parameters plays a dominant role. However, when a sudden heavy rainfall or strong wind occurs, the influence weight of the short-term meteorological LSTM output increases rapidly, and the carbon sink intensity (carbon flux) predicted by the model also decreases immediately, thus achieving an accurate response to changes in actual carbon flux. This process updates the predicted values ​​hourly and can be dynamically verified with measured or remote sensing inversion results, forming a core of carbon sink modeling that is timely and adaptable.

[0177] In one specific embodiment, the meteorological monitoring unit continuously receives real-time meteorological data collected by ground-based micro-meteorological sensors deployed in urban green spaces, focusing on rainfall intensity and visibility. When rainfall intensity exceeds 50 mm per hour or visibility is less than 1 km, the system automatically determines that there is severe atmospheric disturbance in the current area, triggering a verification mechanism to enter a local correction mode. At this time, the surrogate model's predictions under abnormal meteorological conditions may have systematic biases. Therefore, the local correction unit uses a simplified Monte Carlo radiative transfer model for the determined area, significantly reducing the number of photon tracking operations to compress computational costs. The number of photon tracking operations in the original model, which was in the millions, is reduced to one-tenth, retaining only the simulation of the main scattering paths, thereby quickly completing local correction. This model uses a statistical sampling method to simulate the scattering and absorption process of photons between the atmosphere and the ground surface, and its photosynthetically active radiation value is expressed as:

[0178]

[0179] Where M is the number of photons under simplified sampling. The energy represented by each photon, Its transmittance during transmission, This represents the percentage of PAR absorbed by vegetation on the land surface. The corrected PAR value forms a high-confidence cover map in the local area, and then a weighted fusion method is used to embed it into the global PAR distribution map generated by the original proxy model. The fusion expression is:

[0180]

[0181] in The value is the predicted value of photosynthetically active radiation (PAR) from the dynamic model. α is the fusion weight, dynamically set according to the degree of meteorological disturbance, typically between 0.6 and 0.9. This fusion strategy maximizes the preservation of the global efficiency of the surrogate model while ensuring local prediction accuracy, achieving dynamic adaptive correction of the PAR distribution map under extreme weather conditions. This significantly improves the robustness of the system in abnormal scenarios and the reliability of carbon sink monitoring data.

[0182] It should be noted that the short-term LSTM unit (24-hour scale) processes rapidly changing meteorological factors, such as hourly photosynthetically active radiation (PAR), air temperature and humidity, wind speed, and precipitation intensity. The input layer feeds this meteorological data into the network in chronological order. The LSTM unit uses forget gates, input gates, and output gates to control the transmission and updating of information, capturing the short-term changing characteristics of the meteorological data. The long-term LSTM unit (30-day scale) tracks vegetation growth trends. The input layer feeds parameters such as leaf area index and vegetation cover into the network in chronological order, which are also processed by the LSTM unit to learn the long-term trend of vegetation growth. An attention mechanism is used to achieve the fusion of the outputs of the two networks. First, the attention weights of the short-term and long-term network outputs are calculated. Then, the outputs of the two networks are weighted and summed according to the weights to obtain the final carbon flux prediction value. The output is a carbon sink intensity prediction value updated hourly, to more accurately reflect the dynamic changes in carbon flux.

[0183] Furthermore, such as Figure 13 As shown, the short-term LSTM unit includes: a first preprocessing layer, a first LSTM layer, a first dropout layer, a second LSTM layer, and a first attention layer;

[0184] Meteorological data and photosynthetically active radiation are input to the input end of the first preprocessing layer. The output end of the first preprocessing layer is connected to the input end of the first LSTM layer. The output end of the first LSTM layer is connected to the input end of the first dropout layer. The output end of the first dropout layer is connected to the input end of the second LSTM layer. The output end of the second LSTM layer is connected to the input end of the first attention layer. The output end of the first attention layer is connected to the input end of the feature fusion output unit.

[0185] like Figure 4 As shown, the long-term LSTM unit includes: a second preprocessing layer, a third LSTM layer, a second dropout layer, a fourth LSTM layer, and a second attention layer;

[0186] Vegetation state data and vegetation parameter vectors are input to the input of the second preprocessing layer. The output of the second preprocessing layer is connected to the input of the third LSTM layer. The output of the third LSTM layer is connected to the input of the second dropout layer. The output of the second dropout layer is connected to the input of the fourth LSTM layer. The output of the fourth LSTM layer is connected to the input of the second attention layer. The output of the second attention layer is connected to the input of the feature fusion unit.

[0187] Furthermore, such as Figure 15 As shown, the feature fusion output unit includes: a second feature splicing layer, a third attention layer, a third fully connected layer, a fourth activation layer, and a fourth fully connected layer;

[0188] The outputs of the short-term LSTM unit and the long-term LSTM unit are connected to the input of the second feature splicing layer. The output of the second feature splicing layer is connected to the input of the third attention layer. The output of the third attention layer is connected to the input of the third fully connected layer. The output of the third fully connected layer is connected to the input of the fourth activation layer. The output of the fourth activation layer is connected to the input of the fourth fully connected layer. The output of the fourth fully connected layer outputs the carbon sink intensity.

[0189] A dynamic monitoring system for urban green space carbon sequestration based on multi-source remote sensing data includes:

[0190] Data acquisition module: Acquires spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data;

[0191] Vegetation parameter vector module: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector;

[0192] Photosynthetically Active Radiation Module: The spectral data and the corresponding solar azimuth and elevation angles are processed using a dynamic model of photosynthetically active radiation to obtain photosynthetically active radiation;

[0193] Carbon sink intensity calculation module: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, the carbon sink intensity is calculated using a dual-timescale LSTM network.

[0194] The same or similar labels correspond to the same or similar parts;

[0195] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0196] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data, characterized in that, include: S1: Acquire spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data; S2: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector; S3: The spectral data, the solar azimuth angle and elevation angle corresponding to the spectral data are processed using the dynamic model of photosynthetically active radiation to obtain photosynthetically active radiation; S4: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, carbon sink intensity is calculated using a dual-timescale LSTM network. The spectral data includes visible light spectral data and near-infrared spectral data; The dual-channel convolutional model includes: a first feature extraction channel, a second feature extraction channel, a first feature fusion unit, and a first output unit; Visible light band spectral data is input to the input end of the first feature extraction channel, near-infrared band spectral data is input to the input end of the second feature extraction channel, the output ends of the first feature extraction channel and the second feature extraction channel are connected to the input end of the first feature fusion unit, the output end of the first feature fusion unit is connected to the input end of the first output unit, and the output end of the first output unit outputs a vegetation parameter vector. The dual-timescale LSTM network includes: a short-term LSTM unit, a long-term LSTM unit, and a feature fusion output unit. Meteorological data and photosynthetically active radiation are input to the input terminal of the short-term LSTM unit, vegetation status data and vegetation parameter vectors are input to the input terminal of the long-term LSTM unit, and the output terminals of the short-term LSTM unit and the long-term LSTM unit are connected to the input terminal of the feature fusion output unit; the output terminal of the feature fusion output unit outputs carbon sink intensity.

2. The method for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data according to claim 1, characterized in that, The first feature fusion unit includes: a first feature concatenation layer, a first convolutional layer, a batch normalization layer, a first activation layer, and a first global average pooling layer; The outputs of the first feature extraction channel and the second feature extraction channel are connected to the input of the first feature splicing layer. The output of the first feature splicing layer is connected to the input of the first convolutional layer. The output of the first convolutional layer is connected to the input of the batch normalization layer. The output of the batch normalization layer is connected to the input of the first activation layer. The output of the first activation layer is connected to the input of the first global average pooling layer. The output of the first global average pooling layer is connected to the input of the first output unit.

3. The method for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data according to claim 1, characterized in that, The dynamic model of photosynthetically active radiation includes: an input unit, a second feature extraction unit, an attention mechanism unit, a second feature fusion unit, and a second output unit; Spectral data, the corresponding solar azimuth angle and elevation angle are input to the input unit. The output of the input unit is connected to the input of the second feature extraction unit. The output of the second feature extraction unit is connected to the input of the attention mechanism unit. The output of the attention mechanism unit is connected to the input of the second feature fusion unit. The output of the second feature fusion unit is connected to the input of the second output unit. The output of the second output unit outputs photosynthetically active radiation.

4. The method for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data according to claim 3, characterized in that, The second feature extraction unit includes: a second convolutional layer, a first max pooling layer, a third convolutional layer, a second max pooling layer, a fourth convolutional layer, and a third max pooling layer; The output of the input unit is connected to the input of the second convolutional layer; the output of the second convolutional layer is connected to the input of the first max pooling layer; the output of the first max pooling layer is connected to the input of the third convolutional layer; the output of the third convolutional layer is connected to the input of the second max pooling layer; the output of the second max pooling layer is connected to the input of the fourth convolutional layer; the output of the fourth convolutional layer is connected to the input of the third max pooling layer; and the output of the third max pooling layer is connected to the input of the attention mechanism unit. The attention mechanism unit includes: a second global average pooling layer, a first fully connected layer, a second activation layer, a third activation layer, a weight generation layer, and a feature recalibration layer. The output of the second feature extraction unit is connected to the input of the second global average pooling layer. The output of the second global average pooling layer is connected to the input of the first fully connected layer. The output of the first fully connected layer is connected to the input of the second activation layer. The output of the second activation layer is connected to the input of the second fully connected layer. The output of the second fully connected layer is connected to the input of the third activation layer. The output of the third activation layer is connected to the input of the weight generation layer. The output of the weight generation layer is connected to the input of the feature recalibration layer. The output of the feature recalibration layer is connected to the input of the second feature fusion unit.

5. The method for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data according to claim 1, characterized in that, The short-term LSTM unit includes: a first preprocessing layer, a first LSTM layer, a first dropout layer, a second LSTM layer, and a first attention layer; Meteorological data and photosynthetically active radiation are input to the input end of the first preprocessing layer. The output end of the first preprocessing layer is connected to the input end of the first LSTM layer. The output end of the first LSTM layer is connected to the input end of the first dropout layer. The output end of the first dropout layer is connected to the input end of the second LSTM layer. The output end of the second LSTM layer is connected to the input end of the first attention layer. The output end of the first attention layer is connected to the input end of the feature fusion output unit. The long-term LSTM unit includes: a second preprocessing layer, a third LSTM layer, a second dropout layer, a fourth LSTM layer, and a second attention layer; Vegetation state data and vegetation parameter vectors are input to the input of the second preprocessing layer. The output of the second preprocessing layer is connected to the input of the third LSTM layer. The output of the third LSTM layer is connected to the input of the second dropout layer. The output of the second dropout layer is connected to the input of the fourth LSTM layer. The output of the fourth LSTM layer is connected to the input of the second attention layer. The output of the second attention layer is connected to the input of the feature fusion unit.

6. The method for dynamic monitoring of urban green space carbon sequestration based on multi-source remote sensing data according to claim 1, characterized in that, The feature fusion output unit includes: a second feature splicing layer, a third attention layer, a third fully connected layer, a fourth activation layer, and a fourth fully connected layer; The outputs of the short-term LSTM unit and the long-term LSTM unit are connected to the input of the second feature splicing layer. The output of the second feature splicing layer is connected to the input of the third attention layer. The output of the third attention layer is connected to the input of the third fully connected layer. The output of the third fully connected layer is connected to the input of the fourth activation layer. The output of the fourth activation layer is connected to the input of the fourth fully connected layer. The output of the fourth fully connected layer outputs the carbon sink intensity.

7. A method for dynamic monitoring of carbon sequestration in urban green spaces under extreme weather conditions, characterized in that... include: S01: Acquire spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data; S02: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector; S03: Using the Monte Carlo model, process the spectral data and the corresponding solar azimuth and elevation angles to obtain the first photosynthetically active radiation. S04: The spectral data, the solar azimuth angle and elevation angle corresponding to the spectral data are processed using the dynamic model of photosynthetically active radiation to obtain the second photosynthetically active radiation; S05: Integrate the first photosynthetically active radiation and the second photosynthetically active radiation to obtain photosynthetically active radiation; S06: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, carbon sink intensity is calculated using a dual-timescale LSTM network. The dual-channel convolutional model includes: a first feature extraction channel, a second feature extraction channel, a first feature fusion unit, and a first output unit; Visible light band spectral data is input to the input end of the first feature extraction channel, near-infrared band spectral data is input to the input end of the second feature extraction channel, the output ends of the first feature extraction channel and the second feature extraction channel are connected to the input end of the first feature fusion unit, the output end of the first feature fusion unit is connected to the input end of the first output unit, and the output end of the first output unit outputs a vegetation parameter vector. The dual-timescale LSTM network includes: a short-term LSTM unit, a long-term LSTM unit, and a feature fusion output unit. Meteorological data and photosynthetically active radiation are input to the input terminal of the short-term LSTM unit, vegetation status data and vegetation parameter vectors are input to the input terminal of the long-term LSTM unit, and the output terminals of the short-term LSTM unit and the long-term LSTM unit are connected to the input terminal of the feature fusion output unit; the output terminal of the feature fusion output unit outputs carbon sink intensity.

8. A dynamic monitoring system for urban green space carbon sequestration based on multi-source remote sensing data, applied to the monitoring method described in any one of claims 1 to 6, characterized in that, include: Data acquisition module: Acquires spectral data, the corresponding solar azimuth and elevation angles, meteorological data, and vegetation status data; Vegetation parameter vector module: The spectral data is processed using a dual-channel convolution model to obtain a vegetation parameter vector; Photosynthetically Active Radiation Module: The spectral data and the corresponding solar azimuth and elevation angles are processed using a dynamic model of photosynthetically active radiation to obtain photosynthetically active radiation; Carbon sink intensity calculation module: Based on the meteorological data, photosynthetically active radiation, vegetation status data, and vegetation parameter vector, the carbon sink intensity is calculated using a dual-timescale LSTM network.

Images