Ecological process mechanism constrained multi-task learning vegetation carbon sink accounting method and system

Abstract

The application discloses an ecological process mechanism constrained multi-task learning vegetation carbon sink accounting method and system, utilizes a process model to simulate hourly scale vegetation photosynthesis, obtains various spatially layered leaf photosynthetic rates and vegetation net primary productivity, constructs a multi-task learning model, sub-tasks correspond to four photosynthetic rates and vegetation net primary productivity, inputs characteristic variables, and outputs multi-task simulated photosynthetic rates and vegetation net primary productivity, adjusts model parameters through a total loss function, the loss includes sub-task loss and process model coupling loss, verifies precision and compares a single task model, and finally accounts regional or global vegetation carbon sink capacity by using the adjusted model. The application combines ecological process mechanism and multi-task learning, realizes efficient estimation of global carbon sink, and provides a new technical means for carbon cycle research.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing and ecology, specifically to a multi-task learning method and system for calculating vegetation carbon sinks constrained by ecological process mechanisms. Background Technology

[0002] With increasing atmospheric CO2 concentrations and a worsening global climate change situation, research on the global carbon cycle is becoming increasingly urgent. Net primary productivity (NPP) is the net yield obtained by subtracting the energy consumed by plants through respiration from the amount of organic carbon fixed by green vegetation through photosynthesis per unit time. It characterizes the ecological carbon sink capacity of vegetation and is an important quantitative evaluation indicator for carbon cycle research. Therefore, accurate estimation of vegetation carbon sinks is beneficial for efficient monitoring of global carbon cycle processes and has extremely important significance for research on human life, agricultural production, the global carbon cycle, and climate change.

[0003] Ecosystem vegetation carbon sink estimation models based on remote sensing data are often used to extend site-scale carbon flux data to a wider range, such as regional scales. These models include statistical empirical models, light energy use efficiency models, and process models. Among them, process models are based on a deep understanding of ecosystem functions, can describe the physiological and ecological processes of the ecosystem in detail, and output the corresponding process parameters. Therefore, they are generally considered more reliable than other models. However, process models still face certain challenges in application, especially their computational complexity, parameter settings, and the complexity of the model itself. These factors limit their widespread application in practice.

[0004] Machine learning models are also frequently used to estimate vegetation parameters such as carbon sinks, making full use of remote sensing data to efficiently estimate vegetation carbon sinks on large scales and long time series. Currently, many studies use machine learning models (such as random forests, support vector machines, and gradient boosting regression trees) to scale up ground-observed carbon fluxes to regional scales. Furthermore, machine learning models are used to optimize key parameters in traditional ecological estimation models, thereby improving their ability to simulate vegetation carbon sinks. However, the ecological mechanisms underlying vegetation carbon sink estimation using machine learning models are still unclear, resulting in poor interpretability of simulated ecological processes and an inability to output ecological process parameters. The accuracy of these estimations largely depends on the quantity and representativeness of the training data; therefore, relying solely on machine learning methods to estimate global or large-scale vegetation carbon sinks carries significant uncertainty and error.

[0005] Therefore, by combining the advantages of process models and machine learning models, simulations using process models with strong mechanistic interpretability can provide more and more reliable training sample data for machine learning models. Employing a multi-task learning model that considers ecological process mechanism constraints can improve the estimation accuracy and efficiency of vegetation carbon sink NPP parameters, ultimately obtaining high-precision, high-spatiotemporal resolution global vegetation carbon sink data. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-task learning method and system for vegetation carbon sink accounting constrained by ecological process mechanisms. This method utilizes a multi-task learning model constrained by ecological process mechanisms to achieve high-efficiency and high-precision vegetation carbon sink accounting with ecological mechanisms.

[0007] To address the aforementioned technical problems, the present invention adopts the following technical solution: A multi-task learning method for vegetation carbon sink accounting constrained by ecological process mechanisms includes the following steps: The process of vegetation photosynthesis at the hourly scale was simulated using a pre-defined process model, and the photosynthetic rates of leaves in various spatial layers and the net primary productivity of vegetation simulated by the process model were obtained. A multi-task learning model is constructed, which includes multiple sub-tasks; the leaf photosynthetic rates of various spatial strata simulated by the acquired process model and the measured net primary productivity of vegetation are respectively used as labels for the corresponding sub-tasks. By inputting the preset feature variables into the multi-task learning model, the leaf photosynthetic rates of various spatial layers simulated by the multi-task learning model and the net primary productivity of vegetation simulated by the multi-task learning model are obtained. The parameters of the multi-task learning model are adjusted by using a preset total loss function for the multi-task learning model. The accuracy of the adjusted multi-task learning model was evaluated using a preset verification method, and the differences between the multi-task learning model and the single-task model in simulating vegetation carbon sequestration were compared. The adjusted multi-task learning model is used to calculate the regional or global vegetation carbon sink.

[0008] Furthermore, the process model is a BEPS model, and its inputs are: longitude, latitude, year, annual day, Coordinated Universal Time, leaf area index, shortwave radiation, atmospheric temperature, relative humidity, precipitation, wind speed, land cover type, soil texture, soil moisture content, soil snow depth, soil temperature, aggregation index, and carbon dioxide concentration.

[0009] Furthermore, the photosynthetic rates of leaves in various spatial layers simulated by the process model include: photosynthetic rate of upper shaded leaves, photosynthetic rate of upper sun-exposed leaves, photosynthetic rate of lower shaded leaves, and photosynthetic rate of lower sun-exposed leaves.

[0010] Furthermore, the multi-task learning model adopts a hard parameter sharing mechanism, where the bottom layer shares the network to extract the common representation of the input feature variables, and the upper layer sets multiple task-specific output layers to correspond to each sub-task.

[0011] Furthermore, the total loss function of the multi-task learning model includes sub-task loss and process model coupling loss; the leaf photosynthetic rates of various spatial strata simulated by the preset process model and the measured net primary productivity of vegetation are respectively used as corresponding sub-task labels; the leaf photosynthetic rates of various spatial strata simulated by the multi-task learning model and the net primary productivity of vegetation simulated by the multi-task learning model are used to construct the sub-task loss; the net primary productivity of vegetation simulated by the process model and the measured net primary productivity of vegetation are used to construct the loss of the multi-task learning model coupling process model.

[0012] Furthermore, the total loss function of the multi-task learning model is:

[0013] in, It is the total loss function of the multi-task learning model; when hour, For each sub-task loss function, These are the various sub-tasks Related learnable uncertainty parameters; It is an additional loss term that couples the multi-task learning model with the process model. These are the corresponding uncertainty weights. N This represents the total number of subtasks. Time Sub-task Loss Item The mean square error of the photosynthetic rates of leaves in different spatial strata simulated using a multi-task learning model and the photosynthetic rates of leaves in different spatial strata simulated using a process model is measured, and the formula is as follows:

[0014] in The process model simulations are respectively the first The photosynthetic rate of leaves in spatially stratified regions and the first simulation of the multi-task learning model Photosynthetic rates of leaves in spatially stratified species Indicates mean square error; Time Sub-task Loss Item Net primary productivity of vegetation simulated using a multi-task learning model Compared with measured net primary productivity of vegetation The mean squared error is measured by the following formula:

[0015] in These are the simulated net primary productivity of vegetation by the multi-task learning model and the measured net primary productivity of vegetation, respectively. Process model coupling loss term The mean square error between the simulated net primary productivity of vegetation and the measured net primary productivity of vegetation is used as a measure, and the formula is as follows:

[0016] in These are the net primary productivity of vegetation simulated by the process model and the measured net primary productivity of vegetation, respectively.

[0017] Furthermore, the preset verification method includes at least one of the hold-out method, site cross-validation, or regional cross-validation; the accuracy evaluation includes calculating the coefficient of determination and the root mean square error.

[0018] Furthermore, the characteristic variables are: longitude, latitude, year, annual day, Coordinated Universal Time, leaf area index, shortwave radiation, atmospheric temperature, relative humidity, wind speed, land cover type, soil texture, soil temperature, and aggregation index.

[0019] Furthermore, the single-task model is a deep learning model or machine learning model that uses only the measured net primary productivity of vegetation as a label and ignores the subtask of photosynthetic rate of stratified leaves.

[0020] On the other hand, the present invention provides a multi-task learning vegetation carbon sink accounting system constrained by ecological process mechanisms, comprising: The process simulation module is used to simulate the hourly-scale vegetation photosynthesis process using a preset process model, and to obtain the leaf photosynthetic rate and vegetation net primary productivity simulated by the process model in various spatial layers. A multi-task model building module is used to build a multi-task learning model, which includes multiple sub-tasks; the leaf photosynthetic rates of various spatial strata simulated by the acquired process model and the measured net primary productivity of vegetation are respectively used as labels for the corresponding sub-tasks. The vegetation carbon sink NPP simulation module is used to input preset feature variables into the multi-task learning model to obtain leaf photosynthetic rates and vegetation net primary productivity simulated by the multi-task learning model in various spatial layers. The model training module adjusts the parameters of the multi-task learning model using a preset total loss function for the multi-task learning model. The accuracy evaluation module is used to evaluate the accuracy of the adjusted multi-task learning model using a preset verification method and compare the difference between the multi-task learning model and the single-task model in simulating vegetation carbon sequestration. The carbon sequestration module is used to calculate regional or global vegetation carbon sequestration using an adjusted multi-task learning model.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) Traditional process models are widely used for estimating carbon flux and have strong mechanistic explanatory power. This invention uses process parameters simulated by process models as target values ​​for machine learning, so that the final vegetation carbon sink accounting has strong ecological mechanism.

[0022] (2) This invention uses a machine learning multi-task model and introduces ecological process mechanism constraints and flux site data to modify the total loss function of the multi-task learning model, making the estimation of vegetation carbon sink more accurate and efficient.

[0023] The multi-task learning vegetation carbon sink accounting method with ecological process mechanism constraints proposed in this invention can effectively combine the advantages of process models and machine learning models, improve the overall accuracy of vegetation carbon sink accounting, and has strong stability and versatility, making it easy to promote and use. Attached Figure Description

[0024] Figure 1 This is a flowchart of the multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanism in an embodiment of the present invention; Figure 2 This is a schematic diagram comparing the vegetation carbon sink estimation accuracy of the machine learning multi-task learning model with that of the single-task model in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] Example 1 like Figure 1 As shown, this embodiment provides a multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanisms, including the following steps: Step 1: Use a preset process model to simulate the hourly-scale vegetation photosynthesis process, and obtain the leaf photosynthetic rate and vegetation net primary productivity simulated by the process model for various spatial layers. Step 2: Construct a multi-task learning model, which includes multiple sub-tasks; use the leaf photosynthetic rates of different spatial layers simulated by the acquired process model and the measured net primary productivity of vegetation as labels for the corresponding sub-tasks. Step 3: Input the preset feature variables into the multi-task learning model to obtain the leaf photosynthetic rate of various spatial layers simulated by the multi-task learning model and the net primary productivity of vegetation simulated by the multi-task learning model. Step 4: Adjust the parameters of the multi-task learning model using the preset total loss function of the multi-task learning model; Step 5: Use a preset verification method to evaluate the accuracy of the adjusted multi-task learning model and compare the differences between the multi-task learning model and the single-task model in simulating vegetation carbon sequestration. Step 6: Calculate the regional or global vegetation carbon sink using the adjusted multi-task learning model.

[0027] In step 1 of this embodiment, an observation station—AR-SLu—from the internationally recognized FLUXNET2015 dataset is selected for demonstration. This station provides diurnal observation data for three consecutive years, from January 1, 2009 to December 31, 2011, totaling 1095 valid records.

[0028] In step 1 of this embodiment, the process model selected is BEPS (Biosphere-atmosphere Exchange Process Simulator). The inputs of the BEPS model are shown in the table below (the input data not provided by the flux station comes from the remote sensing dataset).

[0029] Table 1

[0030] The photosynthetic rate of the upper shaded leaves, simulated by the process model, can be obtained from the data in Table 1 using the BEPS model. Photosynthetic rate of upper sun-exposed leaves Photosynthetic rate of lower shaded leaves Photosynthetic rate of lower sun-exposed leaves Four different spatially stratified leaf photosynthetic rates Net primary productivity of vegetation simulated by process model .

[0031] The BEPS model considers the vertical structure and radiation distribution of the canopy, and divides the leaves into four categories: upper shaded leaves, upper sun-loving leaves, lower shaded leaves, and lower sun-loving leaves. The photosynthetic rate of the upper and lower shaded leaves is calculated using the same formula, but the parameters of the four types of leaves are different.

[0032] The formula for calculating the photosynthetic rate is as follows:

[0033]

[0034]

[0035]

[0036] In the formula, A Net photosynthetic rate, W c and W j The total photosynthetic rates were determined by Rubisco (ribulose-1,5-bisphosphate carboxylase / oxygenase) restriction and light restriction, respectively. R d This represents the dark respiration of the leaves, and the calculation formula mainly refers to the Farquhar model. V m The maximum carboxylation rate is at 25°C. C i and O 2 represents the intercellular carbon dioxide and oxygen concentrations, respectively; Γ The carbon dioxide compensation point for the absence of dark respiration; K c and K o These are the Michaelis constants for carbon dioxide and oxygen, respectively. J This refers to the electron transport rate.

[0037] In step 2 of this embodiment, the photosynthetic rates of leaves in four different spatial strata are simulated based on the process model obtained in step 1. Net primary productivity of vegetation simulated by process model Combined with the measured net primary productivity of vegetation at flux stations The feature variables of the multi-task learning model (see Table 2) are used to construct the multi-task learning model.

[0038] Table 2

[0039] This embodiment of the multi-task learning model includes five sub-tasks, namely: photosynthetic rate of upper shaded leaves (… ), photosynthetic rate of upper sun-loving leaves ( ), photosynthetic rate of lower shaded leaves ( ), photosynthetic rate of lower sun-loving leaves ( There are four different spatially stratified leaf photosynthetic rates. Net primary productivity of vegetation measured at flux stations .

[0040] In step 3 of this embodiment, the preset feature variables (see Table 2) are input into the multi-task learning model to obtain the leaf photosynthetic rates of four different spatial strata simulated by the multi-task learning model. ) and vegetation carbon sinks simulated by multi-task learning models ( ).

[0041] In step 4 of this embodiment, the parameters of the multi-task learning model in step 3 are optimized.

[0042] The measured net primary productivity of vegetation at the AR-SLu flux station Net primary productivity of vegetation simulated by a multi-task learning model ( The truth label of ).

[0043] The photosynthetic rate of the upper shaded leaves simulated by the process model BEPS ( ), photosynthetic rate of upper sun-loving leaves ( ), photosynthetic rate of lower shaded leaves ( ), photosynthetic rate of lower sun-loving leaves ( ) as the leaf photosynthetic rate simulated by the corresponding multi-task learning model ( The truth label of ).

[0044] The total loss function of the multi-task learning model is as follows:

[0045] In the formula It is the total loss function of the multi-task learning model, calculated by combining different loss terms and a weight function that considers the uncertainty of each loss term. When hour, For each subtask in a multi-task learning model loss function, These are the various sub-tasks The relevant learnable uncertainty parameters are used to calculate the weights of each subtask when calculating the total loss. It is an additional loss term that couples the multi-task learning model with the process model. This is the corresponding uncertainty weight.

[0046] Subtask loss item The mean square error of the photosynthetic rates of leaves in four different spatial strata simulated using a multi-task learning model and the photosynthetic rates of leaves in four different spatial strata simulated using a process model is measured, and the formula is as follows:

[0047] in The process model simulations are respectively the first The photosynthetic rate of leaves in spatial stratification, simulated by a multi-task learning model. Photosynthetic rates of leaves in spatially stratified species This represents the mean square error.

[0048] Subtask loss item The mean squared error between the NPP simulated by the multi-task learning model and the NPP measured at the flux site is used as the measurement, and the formula is as follows:

[0049] in These represent the net primary productivity of vegetation simulated by a multi-task learning model and the net primary productivity of vegetation measured at flux stations, respectively. This represents the mean square error.

[0050] During backpropagation, the multi-task learning model learns not only the expert module's network but also the uncertainty parameters of the sub-tasks. The reason for adding these uncertainty parameters is that when multi-task learning utilizes a single representation framework to accomplish multiple target tasks, these tasks often imply different units and scales. Differences in unit scale can lead to significant differences in the loss values ​​between sub-tasks; furthermore, noise can also affect the magnitude of the loss. To reduce the loss differences caused by units and noise and to unify the scale of the loss, instead of simply adding the losses of each sub-task to obtain the total loss, a different approach is used. To eliminate dimensions, use To measure noise.

[0051] Differences in scale and noise exist between subtasks, and the multi-task learning model incurs additional loss when coupled with the multi-task learning process model. Therefore, an additional loss term is proposed to characterize the loss introduced by the multi-task learning coupling process model. Loss Term The mean square error between the NPP simulated by the process model and the NPP measured at flux sites is used as a measure, and the formula is as follows:

[0052] in These are the net primary productivity of vegetation simulated by the process model and the net primary productivity of vegetation measured at flux stations, respectively. This represents the mean square error.

[0053] The core of this embodiment lies in introducing the mechanistic constraints of the ecological process model. The BEPS mechanistic model is introduced to simulate more refined processes within vegetation. The output of the model carries the mechanistic information of the ecological process. The output of the mechanistic model is used to construct labels and loss functions for multi-task learning, thus constructing a mechanistic-constrained multi-task learning model.

[0054] In step 5 of this embodiment, the "hold-out method" is used to divide the data into a training set and a validation set. Specifically, all 1095 data records are randomly shuffled, and 2 / 3 (approximately 730 records) are selected as the training set for training. Figure 1 The model shown is a "multi-task learning model constrained by ecological process mechanisms"; the remaining 1 / 3 (approximately 365 data points) serves as a validation set to evaluate the prediction accuracy and generalization ability of the trained model on unknown data. The model learning rate is set to 0.001, ultimately forming a model similar to... Figure 2 The result is that by inputting the trained model into global grid data, a global NPP spatial distribution map can be generated, thus enabling the multi-task learning model to be used for global carbon sink accounting.

[0055] like Figure 2 As shown, by using the trained model to validate at the site scale, it can be seen that the multi-task learning model has a higher coefficient of determination and a lower root mean square error compared to the single-task model, indicating that the multi-task learning model has higher accuracy.

[0056] Using a trained model and global-scale satellite grid data, the spatial distribution of global net primary productivity of vegetation as of January 1, 2010, can be plotted, demonstrating that multi-task learning models can be used for global carbon sink accounting.

[0057] Example 2 This embodiment provides a multi-task learning vegetation carbon sink accounting system constrained by ecological process mechanisms, including: The process simulation module is used to simulate the hourly-scale vegetation photosynthesis process using a preset process model, and to obtain the leaf photosynthetic rate and vegetation net primary productivity simulated by the process model in various spatial layers. A multi-task model building module is used to build a multi-task learning model, which includes multiple sub-tasks; the leaf photosynthetic rates of various spatial strata simulated by the acquired process model and the measured net primary productivity of vegetation are respectively used as labels for the corresponding sub-tasks. The vegetation carbon sink NPP simulation module is used to input preset feature variables into the multi-task learning model to obtain leaf photosynthetic rates and vegetation net primary productivity simulated by the multi-task learning model in various spatial layers. The model training module adjusts the parameters of the multi-task learning model using a preset total loss function for the multi-task learning model. The accuracy evaluation module is used to evaluate the accuracy of the adjusted multi-task learning model using a preset verification method and compare the difference between the multi-task learning model and the single-task model in simulating vegetation carbon sequestration. The carbon sequestration module is used to calculate regional or global vegetation carbon sequestration using an adjusted multi-task learning model.

[0058] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0059] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations to the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.

[0060] All other parts not described in detail are existing technologies.

Claims

1. A multi-task learning method for vegetation carbon sequestration constrained by ecological process mechanisms, characterized in that, Includes the following steps: The process of vegetation photosynthesis at the hourly scale was simulated using a pre-defined process model, and the photosynthetic rates of leaves in various spatial layers and the net primary productivity of vegetation simulated by the process model were obtained. The process model simulates the photosynthetic rates of leaves in various spatial layers, including: photosynthetic rate of upper shaded leaves, photosynthetic rate of upper sun-exposed leaves, photosynthetic rate of lower shaded leaves, and photosynthetic rate of lower sun-exposed leaves. A multi-task learning model is constructed, which includes multiple sub-tasks; the leaf photosynthetic rates of various spatial strata simulated by the acquired process model and the measured net primary productivity of vegetation are respectively used as labels for the corresponding sub-tasks. By inputting the preset feature variables into the multi-task learning model, the leaf photosynthetic rates of various spatial layers simulated by the multi-task learning model and the net primary productivity of vegetation simulated by the multi-task learning model are obtained. The parameters of the multi-task learning model are adjusted by a preset total loss function. The total loss function includes sub-task loss and process model coupling loss. The leaf photosynthetic rates of various spatial strata simulated by the preset process model and the measured net primary productivity of vegetation are used as corresponding sub-task labels. The leaf photosynthetic rates of various spatial strata simulated by the multi-task learning model and the net primary productivity of vegetation simulated by the multi-task learning model are used to construct the sub-task loss. The net primary productivity of vegetation simulated by the process model and the measured net primary productivity of vegetation are used to construct the loss of the multi-task learning model coupling process model. The accuracy of the adjusted multi-task learning model was evaluated using a preset verification method, and the differences between the multi-task learning model and the single-task model in simulating vegetation carbon sequestration were compared. The adjusted multi-task learning model is used to calculate the regional or global vegetation carbon sink.

2. The multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanisms according to claim 1, characterized in that, The process model is the BEPS model, and its inputs are: longitude, latitude, year, annual day, Coordinated Universal Time, leaf area index, shortwave radiation, atmospheric temperature, relative humidity, precipitation, wind speed, land cover type, soil texture, soil moisture content, soil snow depth, soil temperature, aggregation index, and carbon dioxide concentration.

3. The multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanisms according to claim 1, characterized in that, The multi-task learning model adopts a hard parameter sharing mechanism. The bottom layer shares a network to extract the common representation of the input feature variables, and the upper layer sets multiple task-specific output layers to correspond to each sub-task.

4. The multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanisms according to claim 1, characterized in that, The total loss function of the multi-task learning model is: in, It is the total loss function of the multi-task learning model; when hour, For each sub-task loss function, These are the various sub-tasks Related learnable uncertainty parameters; It is an additional loss term that couples the multi-task learning model with the process model. These are the corresponding uncertainty weights. N This represents the total number of subtasks. Time Sub-task Loss Item The mean square error of the photosynthetic rates of leaves in different spatial strata simulated using a multi-task learning model and the photosynthetic rates of leaves in different spatial strata simulated using a process model is measured, and the formula is as follows: in The process model simulations are respectively the first The photosynthetic rate of leaves in spatially stratified regions and the first simulation of the multi-task learning model Photosynthetic rates of leaves in spatially stratified species Indicates mean square error; Time Sub-task Loss Item Net primary productivity of vegetation simulated using a multi-task learning model Compared with measured net primary productivity of vegetation The mean squared error is measured by the following formula: in These are the simulated net primary productivity of vegetation by the multi-task learning model and the measured net primary productivity of vegetation, respectively. Process model coupling loss term The mean square error between the simulated net primary productivity of vegetation and the measured net primary productivity of vegetation is used as a measure, and the formula is as follows: in These are the net primary productivity of vegetation simulated by the process model and the measured net primary productivity of vegetation, respectively.

5. The multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanisms according to claim 1, characterized in that, The preset verification method includes at least one of the hold-out method, site cross-validation, or regional cross-validation; the accuracy evaluation includes calculating the coefficient of determination and the root mean square error.

6. The multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanisms according to claim 1, characterized in that, The characteristic variables are: longitude, latitude, year, annual day, Coordinated Universal Time, leaf area index, shortwave radiation, atmospheric temperature, relative humidity, wind speed, land cover type, soil texture, soil temperature, and aggregation index.

7. The multi-task learning vegetation carbon sink accounting method constrained by ecological process mechanisms according to claim 1, characterized in that, The single-task model is a deep learning or machine learning model that uses only the measured net primary productivity of vegetation as a label and ignores the subtask of photosynthetic rate of stratified leaves.

8. A multi-task learning vegetation carbon sink accounting system constrained by ecological process mechanisms, characterized in that, include: The process simulation module is used to simulate the hourly-scale vegetation photosynthesis process using a preset process model, and to obtain the leaf photosynthetic rate and vegetation net primary productivity simulated by the process model in various spatial layers. A multi-task model building module is used to build a multi-task learning model, which includes multiple sub-tasks; the leaf photosynthetic rates of various spatial strata simulated by the acquired process model and the measured net primary productivity of vegetation are respectively used as labels for the corresponding sub-tasks. The vegetation carbon sink NPP simulation module is used to input preset feature variables into the multi-task learning model to obtain leaf photosynthetic rates and vegetation net primary productivity simulated by the multi-task learning model in various spatial layers. The model training module adjusts the parameters of the multi-task learning model using a preset total loss function for the multi-task learning model. The accuracy evaluation module is used to evaluate the accuracy of the adjusted multi-task learning model using a preset verification method and compare the difference between the multi-task learning model and the single-task model in simulating vegetation carbon sequestration. The carbon sequestration module is used to calculate regional or global vegetation carbon sequestration using an adjusted multi-task learning model. The ecological process mechanism-constrained multi-task learning vegetation carbon sink accounting system is used to execute the steps in the ecological process mechanism-constrained multi-task learning vegetation carbon sink accounting method according to any one of claims 1-7.

Images