A method and system for predicting tropical forest soil carbon fluxes through multi-source data fusion

Abstract

The embodiment of the application discloses a method and system for predicting tropical forest soil carbon flux by fusing multi-source data, which comprises the following steps: acquiring multi-source environmental monitoring data, including soil temperature, soil moisture, air temperature, precipitation, soil organic matter and extreme weather event markers; based on the soil temperature and the extreme weather event markers, a first soil carbon flux prediction value is generated through a predetermined exponential function in a pre-trained double-branch prediction model; based on all multi-source environmental monitoring data, a second soil carbon flux prediction value and a prediction reliability are generated through a pre-trained multilayer perception in the pre-trained double-branch prediction model; based on the prediction reliability and the extreme weather event markers, a fusion weight is calculated through a predetermined gating function; and based on the fusion weight, the second soil carbon flux prediction value and the first soil carbon flux prediction value are weighted and fused to obtain a soil carbon flux prediction value.

Description

Technical Field

[0001] This application relates to the fields of ecological monitoring and artificial intelligence technology, and relates to, but is not limited to, a method and system for predicting soil carbon flux in tropical forests by multi-source data fusion. Background Technology

[0002] Soil carbon flux, or soil respiration rate, is a key indicator characterizing the carbon cycle in forest ecosystems. Tropical coastal shelterbelts are characterized by unique alternating wet and dry seasons and frequent typhoon disturbances, and soil respiration is affected by the nonlinear coupling of multiple factors such as soil temperature, soil moisture, soil organic matter, and air temperature and humidity.

[0003] Existing prediction models can maintain a certain level of accuracy under normal weather conditions, but when extreme events such as typhoons occur, the combination of input features deviates significantly from the training set distribution, and the model's predictive ability drops sharply, which seriously affects the continuity and accuracy of regional carbon sink assessment.

[0004] Therefore, there is an urgent need for a method and system for predicting tropical forest soil carbon flux by fusing multi-source data, in order to solve the problem that the ability to quickly restore tropical forest soil carbon flux prediction is not possible after extreme weather events due to data distribution shifts and sample scarcity. Summary of the Invention

[0005] In view of this, the embodiments of this application provide a method and system for predicting soil carbon flux in tropical forests by fusing multi-source data, which at least solves the problem of not being able to quickly restore the ability to predict soil carbon flux in tropical forests after extreme weather events.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for predicting soil carbon flux in tropical forests through multi-source data fusion, the method comprising: Acquire multi-source environmental monitoring data, including soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers; Based on the soil temperature and the extreme weather event markers, a first predicted soil carbon flux value is generated using a pre-determined exponential function in a pre-trained bi-branch prediction model. Based on all the multi-source environmental monitoring data, a second soil carbon flux prediction value and prediction reliability are generated through the multilayer perceptron pre-trained in the pre-trained dual-branch prediction model. Based on the prediction reliability and the extreme weather event label, the fusion weights are calculated using a pre-determined gating function; The second predicted soil carbon flux and the first predicted soil carbon flux are weighted and fused based on the fusion weight to obtain the predicted soil carbon flux.

[0007] Secondly, embodiments of this application provide a tropical forest soil carbon flux prediction system based on multi-source data fusion. The system includes: a data acquisition module, a first prediction module, a second prediction module, a weight determination module, and a weighted fusion module. The data acquisition module is used to acquire multi-source environmental monitoring data, which includes soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers. The first prediction module is used to generate a first soil carbon flux prediction value based on the soil temperature and the extreme weather event marker, using a pre-determined exponential function in a pre-trained bi-branch prediction model. The second prediction module is used to generate a second soil carbon flux prediction value and prediction reliability based on all the multi-source environmental monitoring data and through the multilayer perceptron pre-trained in the pre-trained dual-branch prediction model. The weight determination module is used to calculate the fusion weights based on the prediction reliability and the extreme weather event label using a pre-determined gating function. The weighted fusion module is used to perform weighted fusion of the second soil carbon flux prediction value and the first soil carbon flux prediction value based on the fusion weight to obtain the soil carbon flux prediction value.

[0008] The beneficial effects of the technical solutions provided in this application include at least the following: This application utilizes multi-source environmental monitoring data, including soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers. This multi-source environmental monitoring data covers the main environmental factors affecting soil carbon flux and extreme event indicators, providing comprehensive input for a bi-branch prediction model. This enables the subsequent model to leverage physical priors and data-driven capabilities, improving prediction accuracy and adaptability to extreme events. Based on soil temperature and extreme weather event markers, a first predicted soil carbon flux value is generated using a pre-defined exponential function in a pre-trained bi-branch prediction model. The pre-defined exponential function utilizes ecological exponential laws as physical priors during prediction, avoiding predictions that violate physical laws under extreme conditions from purely data-driven models. When extreme weather events occur, the pre-defined exponential function parameters are adjusted to adapt to changes in the temperature sensitivity coefficient, thereby improving the bi-branch prediction model's adaptability to extreme events without increasing the training sample size. Based on all multi-source environmental monitoring data, a second predicted soil carbon flux value and prediction reliability are generated using a pre-trained multilayer perceptron in the pre-trained bi-branch prediction model. By leveraging the complex nonlinear relationships among data such as soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers, a second predicted soil carbon flux value is output, compensating for the shortcomings of exponential functions relying solely on soil temperature in capturing the interactions of multiple factors. Simultaneously, by calculating prediction reliability, an adaptive basis for data quality is provided for subsequent gating fusion. High prediction reliability indicates that the multilayer perceptron has high confidence in predicting the current input and can therefore occupy a larger proportion in the fusion; low prediction reliability indicates that the current input may exceed the training distribution or has data quality issues, and the gating function automatically reduces the contribution of the multilayer perceptron, thus avoiding the impact of unreliable predictions on the final result. This step enables the model to possess both high-precision fitting ability and self-evaluation ability, improving the overall robustness and reliability of the prediction. Based on prediction reliability and extreme weather event markers, fusion weights are calculated using a pre-determined gating function. The contribution ratio of the knowledge-guided branch and the data-driven branch is determined by the self-evaluation information of the data-driven branch and the external environmental state information, realizing the dynamic adaptive calculation of fusion weights. When prediction reliability is high and extreme weather events are absent, the fusion weight is large, and the data-driven branch dominates in subsequent fusion, fully leveraging its multi-factor fitting capabilities. When prediction reliability is low or extreme weather events occur, the fusion weight decreases, and the contribution ratio of the knowledge-guided branch increases, ensuring that the final prediction result still has reasonable physical meaning under conditions of poor data quality or sudden environmental changes. Without the need for manually preset fixed weights or additional supervisory signals, the fusion strategy can automatically adjust based on the current input data quality and environmental state, improving adaptability to different data conditions and environmental scenarios.The second and first soil carbon flux predictions are weighted and fused based on fusion weights to obtain the final soil carbon flux prediction. This method can automatically adjust the contribution ratio of the two branches according to the current data quality and environmental conditions. When the data quality is good, it fully leverages the multilayer perceptron's ability to fit multi-factor nonlinear relationships. When the data quality is poor or extreme weather events occur, it increases the contribution of the exponential function's physical prior, thereby improving the model's adaptability to environmental changes while maintaining prediction accuracy. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating a method for predicting soil carbon flux in tropical forests using multi-source data fusion, provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a tropical forest soil carbon flux prediction system based on multi-source data fusion, provided in an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0012] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0013] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0014] This application provides a method for predicting soil carbon flux in tropical forests through multi-source data fusion. Figure 1 A flowchart illustrating a method for predicting soil carbon flux in tropical forests using multi-source data fusion, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes at least the following steps: Step S110: Obtain multi-source environmental monitoring data, which includes soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers.

[0015] Soil temperature is a measurement of the thermal state of soil at a specific depth below the forest surface, expressed in degrees Celsius. Soil temperature serves as the input to the exponential function in the knowledge-guided branch, reflecting the exponential driving effect of temperature on soil respiration rate.

[0016] Soil moisture is the volume percentage of water in soil pores, which affects soil aeration and microbial activity.

[0017] Air temperature is the atmospheric temperature near the forest canopy, measured in degrees Celsius (°C). It, along with soil temperature, influences plant physiological processes and soil microclimate.

[0018] Precipitation can be daily rainfall, measured in millimeters (mm), and indirectly regulates soil respiration by affecting soil moisture. Those skilled in the art will understand that precipitation can be measured on different time scales, such as hourly, daily, or weekly, depending on the sensor's sampling frequency and the requirements of the prediction task.

[0019] Soil organic matter is the organic carbon content in soil derived from plant and animal remains, measured in grams per kilogram (g / kg). It is the main carbon source substrate for soil respiration.

[0020] Extreme weather events are tagged with binary values, where 1 indicates an extreme weather event has occurred and 0 indicates no extreme weather event has occurred. This tag is used to adjust the parameters of the exponential function in the knowledge-guided branch during the inference phase and participates in the calculation of fusion weights in the gating function. For example, extreme weather events include, but are not limited to, typhoons, heavy rainfall, or high temperatures.

[0021] Multi-source environmental monitoring data covers the main environmental factors affecting soil carbon flux and information on extreme events, providing comprehensive input for the bi-branch prediction model. This enables the subsequent model to leverage physical priors and data-driven capabilities to improve prediction accuracy and adaptability to extreme events.

[0022] Step S120: Based on the soil temperature and the extreme weather event markers, a first predicted soil carbon flux value is generated using a pre-determined exponential function in a pre-trained bi-branch prediction model.

[0023] Soil temperature is a measurement of soil heat at a specific depth below the forest surface, measured in °C, and is a major environmental factor driving soil respiration rate.

[0024] The dual-branch prediction model comprises a knowledge-guided branch and a data-driven branch. The knowledge-guided branch calculates predicted soil carbon flux based on soil temperature using a pre-defined exponential function. When extreme weather events are indicated by markers, the parameters of the exponential function are adjusted to accommodate changes in the temperature sensitivity coefficient following the extreme event. The first predicted soil carbon flux is the result output by the knowledge-guided branch, expressed in micromoles of carbon dioxide per square meter per second (μmol CO2·m²). -2 ·s -1 ).

[0025] By using a pre-defined exponential function, ecological exponential laws can be utilized as physical priors during prediction, avoiding predictions that violate physical laws under extreme conditions from purely data-driven models. When extreme weather events occur, the pre-defined exponential function parameters are adjusted to adapt to changes in the temperature sensitivity coefficient, thereby improving the adaptability of the bi-branch prediction model to extreme events without increasing the number of training samples.

[0026] Step S130: Based on all the multi-source environmental monitoring data, a second soil carbon flux prediction value and prediction reliability are generated through the multilayer perceptron pre-trained in the pre-trained dual-branch prediction model.

[0027] All multi-source environmental monitoring data include soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers.

[0028] The dual-branch prediction model includes a knowledge-guided branch and a data-driven branch. The data-driven branch corresponds to a pre-trained multilayer perceptron. A multilayer perceptron is a feedforward neural network composed of multiple fully connected layers, and the parameters of a trained multilayer perceptron are fixed.

[0029] The second predicted soil carbon flux is the result of a pre-trained multilayer perceptron. The units for the second predicted soil carbon flux are the same as those for the first predicted soil carbon flux.

[0030] Prediction reliability is a quantitative indicator of the credibility of the predicted second soil carbon flux value obtained by the data-driven branch. It is obtained by calculating the standard deviation of the output value through multiple forward propagations with discard operations. The smaller the standard deviation, the higher the prediction reliability of the second soil carbon flux value.

[0031] By leveraging the complex nonlinear relationships between data such as soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers, a second predicted soil carbon flux value is output, compensating for the shortcomings of exponential functions relying solely on soil temperature in capturing the interactions of multiple factors. Simultaneously, by calculating prediction reliability, an adaptive basis for data quality is provided for subsequent gating fusion. High prediction reliability indicates that the multilayer perceptron has high confidence in the current input and can therefore play a larger role in the fusion process; low prediction reliability indicates that the current input may deviate from the training distribution or that there are data quality issues. The gating function automatically reduces the contribution of the multilayer perceptron, thus avoiding the impact of unreliable predictions on the final result. This step enables the model to possess both high-precision fitting capabilities and self-evaluation abilities, improving the overall robustness and reliability of the predictions.

[0032] Step S140: Based on the prediction reliability and the extreme weather event label, calculate the fusion weights using a pre-determined gating function.

[0033] A pre-defined gating function maps the sum of the negative weights for prediction reliability and the positive weights for extreme weather event labels to a value between 0 and 1; the output value is the fusion weight. The fusion weight, with a value between 0 and 1, is used to allocate the contribution ratio of the data-driven branch and the knowledge-guided branch in subsequent weighted fusion steps.

[0034] The contribution ratios of the knowledge-guided branch and the data-driven branch are determined by the self-evaluation information of the data-driven branch and the external environmental state information, achieving dynamic adaptive calculation of the fusion weight. When the prediction reliability is high and there are no extreme weather events, the fusion weight is large, and the data-driven branch dominates in subsequent fusion, fully leveraging its multi-factor fitting ability. When the prediction reliability is low or extreme weather events occur, the fusion weight decreases, and the contribution ratio of the knowledge-guided branch increases, ensuring that the final prediction result still has reasonable physical meaning under conditions of poor data quality or sudden environmental changes. Without the need for manually preset fixed weights or relying on additional supervision signals, the fusion strategy can be automatically adjusted according to the current input data quality and environmental state, improving adaptability to different data conditions and environmental scenarios.

[0035] Step S150: Based on the fusion weight, the second predicted soil carbon flux value and the first predicted soil carbon flux value are weighted and fused to obtain the predicted soil carbon flux value.

[0036] It can automatically adjust the contribution ratio of the two branches according to the current data quality and environmental conditions. When the data quality is good, it can give full play to the fitting ability of the multilayer perceptron to the nonlinear relationship of multiple factors. When the data quality is poor or extreme weather events occur, it can increase the contribution of the physical prior of the exponential function, thereby improving the model's adaptability to environmental changes while maintaining prediction accuracy.

[0037] This application utilizes multi-source environmental monitoring data, including soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers. This multi-source environmental monitoring data covers the main environmental factors affecting soil carbon flux and extreme event indicators, providing comprehensive input for a bi-branch prediction model. This enables the subsequent model to leverage physical priors and data-driven capabilities, improving prediction accuracy and adaptability to extreme events. Based on soil temperature and extreme weather event markers, a first predicted soil carbon flux value is generated using a pre-defined exponential function in a pre-trained bi-branch prediction model. The pre-defined exponential function utilizes ecological exponential laws as physical priors during prediction, avoiding predictions that violate physical laws under extreme conditions from purely data-driven models. When extreme weather events occur, the pre-defined exponential function parameters are adjusted to adapt to changes in the temperature sensitivity coefficient, thereby improving the bi-branch prediction model's adaptability to extreme events without increasing the training sample size. Based on all multi-source environmental monitoring data, a second predicted soil carbon flux value and prediction reliability are generated using a pre-trained multilayer perceptron in the pre-trained bi-branch prediction model. By leveraging the complex nonlinear relationships among data such as soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers, a second predicted soil carbon flux value is output, compensating for the shortcomings of exponential functions relying solely on soil temperature in capturing the interactions of multiple factors. Simultaneously, by calculating prediction reliability, an adaptive basis for data quality is provided for subsequent gating fusion. High prediction reliability indicates that the multilayer perceptron has high confidence in predicting the current input and can therefore occupy a larger proportion in the fusion; low prediction reliability indicates that the current input may exceed the training distribution or has data quality issues, and the gating function automatically reduces the contribution of the multilayer perceptron, thus avoiding the impact of unreliable predictions on the final result. This step enables the model to possess both high-precision fitting ability and self-evaluation ability, improving the overall robustness and reliability of the prediction. Based on prediction reliability and extreme weather event markers, fusion weights are calculated using a pre-determined gating function. The contribution ratio of the knowledge-guided branch and the data-driven branch is determined by the self-evaluation information of the data-driven branch and the external environmental state information, realizing the dynamic adaptive calculation of fusion weights. When prediction reliability is high and extreme weather events are absent, the fusion weight is large, and the data-driven branch dominates in subsequent fusion, fully leveraging its multi-factor fitting capabilities. When prediction reliability is low or extreme weather events occur, the fusion weight decreases, and the contribution ratio of the knowledge-guided branch increases, ensuring that the final prediction result still has reasonable physical meaning under conditions of poor data quality or sudden environmental changes. Without the need for manually preset fixed weights or additional supervisory signals, the fusion strategy can automatically adjust based on the current input data quality and environmental state, improving adaptability to different data conditions and environmental scenarios.The second and first soil carbon flux predictions are weighted and fused based on fusion weights to obtain the final soil carbon flux prediction. This method can automatically adjust the contribution ratio of the two branches according to the current data quality and environmental conditions. When the data quality is good, it fully leverages the multilayer perceptron's ability to fit multi-factor nonlinear relationships. When the data quality is poor or extreme weather events occur, it increases the contribution of the exponential function's physical prior, thereby improving the model's adaptability to environmental changes while maintaining prediction accuracy.

[0038] Optionally, the method further includes: acquiring initial multi-source environmental monitoring data; performing time alignment, missing data filling, and standardization on the initial multi-source environmental monitoring data to obtain the multi-source environmental monitoring data.

[0039] Optionally, the multi-source environmental monitoring data further includes a missing mask channel; the missing mask channel is used to identify each feature data in the multi-source environmental monitoring data as a measured value or a filled value.

[0040] In actual monitoring, some data in multi-source environmental monitoring data may be missing due to sensor malfunctions, communication interruptions, or long laboratory analysis cycles. For missing data, forward imputation is used to fill in the gaps, employing the most recent valid observation to ensure the continuity of the data sequence. However, imputed data is not the actual observed value and its reliability is lower than that of the measured data. To enable the model to distinguish between measured and imputed data, a corresponding missing mask channel is generated for each data point in the multi-source environmental monitoring data. The missing mask channel is a binary identifier. When each data point is a measured value, the missing mask channel takes the first value; when the feature data is an imputed value, the missing mask channel takes the second value. For example, the first value can be 1, and the second value can be 0.

[0041] Missing mask channels are added as additional input features to the multi-source environmental monitoring data and fed into the multilayer perceptron (MLP) in the data-driven branch. During training, the MLP learns the reliability of each feature through backpropagation: when the missing mask channel corresponding to a feature takes the second value, it indicates that the feature is a filled value with low reliability, and the MLP reduces its contribution to the second soil carbon flux prediction value in its internal weight allocation; conversely, when the missing mask channel takes the first value, it indicates that the feature is a measured value, and the MLP uses the feature normally for prediction. Thus, the data-driven branch can perceive the reliability of the source of each feature data during the prediction process and automatically reduce its reliance on unreliable features when there are many missing data points, thereby improving the robustness of the prediction.

[0042] Optionally, the predetermined exponential function includes an adjustable temperature response coefficient; the step of generating a first soil carbon flux prediction value based on the soil temperature and the extreme weather event marker, using the predetermined exponential function in a pre-trained bi-branch prediction model, includes: when the extreme weather event marker indicates the existence of an extreme weather event, multiplying the adjustable temperature response coefficient by a preset adjustment coefficient to obtain an adjusted temperature response coefficient; the preset adjustment coefficient is predetermined based on the rate of change of the temperature sensitivity coefficient in historical extreme weather events; updating the exponential function based on the adjusted temperature response coefficient to obtain an adjusted exponential function; and obtaining the first soil carbon flux prediction value based on the adjusted exponential function and the soil temperature.

[0043] The exponential function is shown in formula (1): Formula (1); Where 'a' is the baseline respiratory rate coefficient. The temperature response coefficient is adjustable, with initial values ​​of a = 1 and b = 0.1. This is the first predicted value for soil carbon flux; Soil temperature.

[0044] During training, the baseline breathing rate coefficient 'a' is updated via gradient descent.

[0045] The impact of extreme weather on soil carbon flux is multifactorial, but after extreme weather occurs, adjusting the temperature response coefficient can capture the main characteristics of changes in respiration rate.

[0046] If extreme weather events are marked as 1, the adjusted temperature response coefficient is calculated as shown in formula (2): Formula (2); in, To adjust the temperature response coefficient, For adjustable temperature response coefficient, This is the preset adjustment coefficient.

[0047] The preset adjustment coefficient is determined in advance based on the rate of change of the temperature sensitivity coefficient in historical extreme weather events, for example, after a typhoon in the casuarina forest. If the value increases from 1.90 to 2.29, then β = ln(2.29) / ln(1.90) ≈ 1.29. =1.29.

[0048] The adjusted exponential function is shown in formula (3): Formula (3); When extreme weather events are marked as 0, at this time... =1, .

[0049] Optionally, the step of generating a second soil carbon flux prediction value and prediction reliability based on all the multi-source environmental monitoring data and through a pre-trained multilayer perceptron in the pre-trained dual-branch prediction model includes: inputting all the multi-source environmental monitoring data into the multilayer perceptron, randomly deactivating some neurons during each forward propagation due to a dropout operation, thereby obtaining multiple candidate prediction values ​​for soil carbon flux; calculating the arithmetic mean of the multiple candidate prediction values ​​for soil carbon flux to obtain the second soil carbon flux prediction value; calculating the standard deviation of the multiple candidate prediction values ​​for soil carbon flux to obtain a dispersion index; and obtaining the prediction reliability based on the dispersion index and the second soil carbon flux prediction value.

[0050] The calculation of the predicted reliability is shown in formula (4): Formula (4); in, To predict reliability, This is the second predicted value for soil carbon flux. The dispersion index is used to characterize the fluctuation range or dispersion of multiple prediction results. This is a very small positive constant. Introducing this positive constant can avoid the situation where the denominator is zero or meaningless when the predicted value of the second soil carbon flux is close to zero, thus ensuring the stability of the calculation.

[0051] A smaller prediction reliability value indicates a more concentrated distribution of candidate soil carbon flux predictions obtained from multiple forward propagations, resulting in a more stable prediction by the multilayer perceptron for that input and higher prediction reliability. Conversely, a larger prediction reliability value indicates a more dispersed distribution of candidate soil carbon flux predictions obtained from multiple forward propagations, resulting in greater fluctuations in the multilayer perceptron's prediction for that input and lower prediction reliability. Through this mechanism, this invention simultaneously outputs the prediction reliability as a quantitative confidence reference for the second soil carbon flux prediction value, enabling subsequent gating fusion to adaptively adjust the contribution ratio of the data-driven branch based on the prediction reliability, thereby improving the model's robustness under conditions of fluctuating data quality.

[0052] Optionally, the gating function is a logistic function; the step of calculating the fusion weight based on the prediction reliability and the extreme weather event label using a predetermined gating function includes: summing the negative weighted value of the prediction reliability and the positive weighted value of the extreme weather event label to obtain a gating input value; inputting the gating input value into the logistic function to obtain the fusion weight; wherein, the lower the prediction reliability, the smaller the fusion weight.

[0053] The calculation of the fusion weight is shown in formula (5): Formula (5); in, To integrate weights, For logical stearic functions; Weighting coefficients for predicting reliability, Weighting coefficients for extreme weather events. For bias terms, To predict reliability, Mark extreme weather events. To predict the negative weighting value for reliability, Positive weighted values ​​for labeling extreme weather events.

[0054] The range of values ​​for the weighting coefficients of predictive reliability is as follows: ,For example, Initialized to 1.0 and maintained as a positive value during training through parameter constraints. Weight coefficients for extreme weather event labels. Initialize to negative values, for example, It is initialized to -0.5 and kept negative or zero during training via projected gradient descent. Initialize to a positive value, for example, Initialize to 1.5 to ensure that the data-driven branch dominates under non-extreme weather events and low uncertainty conditions; simultaneously To ensure that g decreases as uncertainty increases, the weights of the physical prior branches increase smoothly. The greater the uncertainty u, the better. The smaller the value of g, the smaller the value of g, and the more the model depends on the physical prior branches.

[0055] Prediction reliability is negatively correlated with fusion weights; that is, the lower the prediction reliability, the smaller the fusion weight. When prediction reliability is low, it indicates that the multilayer perceptron's multiple forward propagations of the current input result in a high degree of dispersion among the various candidate soil carbon flux predictions, leading to a lower confidence level in the second soil carbon flux prediction. In this case, the gating mechanism automatically reduces the fusion weights, allowing the first soil carbon flux prediction generated by the knowledge-guided branch to have a larger weight in the final weighted fusion. When extreme weather event markers indicate the presence of extreme weather events, the positive weighting term increases the gating input value, further reducing the fusion weights, and making the model more reliant on the physical prior information provided by the knowledge-guided branch. This adaptive adjustment mechanism enables the model to automatically reduce the contribution of the data-driven branch and increase the contribution of the knowledge-guided branch under conditions of poor data quality or sudden environmental changes, thereby enhancing the model's prediction stability in complex and variable environments.

[0056] Optionally, the pre-trained multilayer perceptron is trained through the following steps: acquiring historical multi-source environmental monitoring data and corresponding historical measured soil carbon flux values; inputting all the historical multi-source environmental monitoring data into the multilayer perceptron to be trained, performing multiple forward propagation operations with drop-off operations on the multilayer perceptron to be trained, obtaining multiple candidate predicted values ​​of historical soil carbon flux; calculating the mean of the multiple candidate predicted values ​​of historical soil carbon flux to obtain a second predicted value of historical soil carbon flux; calculating a loss value based on the difference between the second predicted value of historical soil carbon flux and the measured value of historical soil carbon flux; and updating the parameters of the multilayer perceptron to be trained based on the loss value using a backpropagation algorithm to obtain the pre-trained multilayer perceptron.

[0057] The training process of the pre-trained multilayer perceptron is achieved by minimizing the mean squared error (MSE) and combining it with L2 regularization. This optimization objective updates the network parameters through the backpropagation algorithm, reducing the risk of model overfitting while ensuring prediction accuracy. The specific loss function is defined as shown in formula (6): Formula (6); in, This is the loss value. B is the sample sequence number index, and B is the batch size. For the predicted value output for the j-th sample, This corresponds to the actual observed soil carbon flux. The loss function consists of two parts: the first term is the empirical risk, used to measure prediction bias; the latter two terms are regularization terms, used to constrain the model parameters. For all trainable weights of the data-driven branch (multilayer perceptron); These are the trainable parameters of the gating unit. By setting the regularization coefficient and Applying different degrees of L2 penalty to the parameters of the data-driven branch and the gating unit effectively prevents overfitting of the data-driven branch, while avoiding the impact on the flexibility of the fusion mechanism due to excessive constraints on the gating parameters. The training process uses the Adam optimizer to iteratively update all trainable parameters. Regularization coefficients... and The possible values ​​are: , .

[0058] To prevent gradient explosion of the exponential function, gradient clipping (with a gradient norm threshold set to 1.0) can be used during training to ensure physical meaning. To prevent the gating weights from prematurely converging to extreme values ​​and causing the gradient of the physical prior branch to vanish, the two branches can be trained separately until the loss on the validation set decreases by less than 0.1% for five consecutive rounds, before unfreezing the gating parameters for joint training. Since extreme weather events account for a relatively small proportion of the samples in the training set, when calculating the mean squared error term, samples marked as 1 for extreme weather events can be assigned higher loss weights, for example, a weight coefficient of 5, while the regularization term remains unchanged, to enhance the gating's generalization ability to extreme events.

[0059] Optionally, the step of weighting and fusing the second predicted soil carbon flux value and the first predicted soil carbon flux value based on the fusion weight to obtain the predicted soil carbon flux value includes: weighting the second predicted soil carbon flux value using the fusion weight to obtain a first weighted component; determining the complementary weight corresponding to the first predicted soil carbon flux value using the fusion weight; weighting the first predicted soil carbon flux value using the complementary weight to obtain a second weighted component; and adding the first weighted component and the second weighted component to obtain the predicted soil carbon flux value.

[0060] The predicted value of soil carbon flux is calculated as shown in formula (7): Formula (7); in, This is a predicted value for soil carbon flux. To integrate weights, This is the second predicted value for soil carbon flux. This is the first predicted value for soil carbon flux. These are complementary weights.

[0061] Soil carbon flux predictions can be used in the following scenarios: In real-time monitoring, soil carbon flux predictions are used to plot dynamic curves of soil carbon flux and trigger anomaly alarms when the predicted values ​​exceed a preset range. In carbon sink assessment, daily, monthly, and annual carbon emissions are cumulatively calculated based on soil carbon flux predictions. In ecological early warning, during non-extreme weather events, if the prediction reliability consistently exceeds a preset threshold and the predicted soil carbon flux deviates from the historical average for the same period by more than 20%, it indicates potential unrecorded interference, including pests, diseases, or human disturbance. Increased prediction reliability during extreme weather events is a normal adaptive behavior of the model design and does not trigger anomaly alarms.

[0062] The following describes a method for predicting tropical forest soil carbon flux using multi-source data fusion, based on a specific embodiment. This specific embodiment is merely for illustrating the invention and does not constitute an undue limitation thereof.

[0063] The target study area was a national-level forest ecosystem observation and research station in Hainan Province. Forest types included Casuarina equisetifolia (CA), coconut palm (CO), broadleaf mixed forest (MF), and secondary forest (SF). The data period was from September 2023 to August 2024, covering a full year, with the dry season from November to April of the following year and the wet season from May to October. One sample was collected daily, with approximately 365 samples per forest type. The dataset was divided chronologically: the first 70% (approximately 255 days) was used for training, the middle 15% (approximately 55 days) for validation, and the last 15% (approximately 55 days) for testing. The test set covered the period following typhoons. Evaluation metrics included root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ).

[0064] This application designs four representative comparative models to verify the effectiveness of the proposed method for soil carbon flux prediction. The first is a purely data-driven multilayer perceptron (MLP) model, which uses multi-source environmental monitoring features for regression without introducing any physical mechanism constraints. The second is a physical model based on a fixed-parameter exponential function (Q-model). 10 The first method uses soil temperature as the sole input variable, constructing a linear response relationship using a fixed temperature sensitivity coefficient, with parameters entirely dependent on historical data for fitting. The second method is the Fixed Gate fusion model, which employs a dual-branch architecture but uses statically set fusion weights to fuse the outputs of the data-driven and knowledge-guided branches, with a fixed fusion weight ratio of 7:3 between the data-driven and knowledge-guided branches. The third method, the one described in this application (Ours), constructs a dual-branch fusion architecture, introduces learnable physical prior constraints, integrates prediction reliability estimation based on multiple forward propagations including dropout operations, and possesses an adaptive switching mechanism for extreme weather events, thereby improving the model's physical interpretability and robustness while ensuring prediction accuracy.

[0065] Table 1. Performance comparison of different methods on the test set

[0066] As shown in Table 1, the experimental results indicate that MLP has a certain fitting ability, but because it relies entirely on the statistical correlation of data and does not introduce any physical prior constraints, the prediction error is still quite significant. 10Because the temperature sensitivity coefficient is fixed, this model cannot adapt to different forest types and environmental changes, resulting in the lowest prediction accuracy among the four methods. Fixed Gate improves prediction accuracy compared to single-branch models by fusing data-driven and knowledge-guided branches, but because the fusion weights are statically set, they cannot be dynamically adjusted according to data quality and environmental conditions, failing to fully leverage the synergistic effect of the two branches. Compared to MLP, the method in this application reduces the root mean square error by 31.5%; compared to Fixed Gate, the root mean square error is reduced by 21.3%. This result fully demonstrates the effectiveness of introducing learnable physical priors and adaptive gating mechanisms in this application, not only improving the accuracy of soil carbon flux prediction but also ensuring the reliability of the model output.

[0067] To further verify the generalization ability of the method in this application under different forest types, comparative experiments were conducted for different forest types, as shown in Table 2: Table 2 Comparison of RMSE between this application and MLP in different forest types

[0068] As shown in Table 2, in casuarina forests, the root mean square error (RMSE) of the MLP was 1.18, while the RMSE of the method in this application was reduced to 0.82, representing a relative improvement of 30.5% in prediction accuracy. In coconut forests, the RMSE of the MLP was 1.09, while the RMSE of the method in this application was reduced to 0.74, representing a relative improvement of 32.1%. In broadleaf mixed forests, the RMSE of the MLP was 1.31, while the RMSE of the method in this application was reduced to 0.91, representing a relative improvement of 30.5%. In secondary forests, the RMSE of the MLP was the highest, reaching 1.38, while the method in this application reduced it to 0.93, representing a relative improvement of 32.6%. The above data fully demonstrate that the method in this application can adapt to the soil carbon flux variation patterns of different types of forest ecosystems and maintain excellent prediction performance and stability under various complex ecological scenarios.

[0069] To evaluate the stability and robustness of the method in this application when dealing with missing data in real-world scenarios, this application conducted a data missing robustness verification. As shown in Table 3: Table 3 Robustness Validation for Missing Data

[0070] As shown in Table 3, under the condition of 30% missing data, the root mean square error (RMSE) of the MLP deteriorated sharply from 1.24 to 1.98, a performance decrease of 59.7%, demonstrating extremely low fault tolerance. The performance of the Fixed Gate also showed a significant decline, with the RMS error increasing from 1.08 to 1.45, a performance decrease of 34.3%. In contrast, the method in this application exhibited excellent robustness, with the RMS error only slightly increasing from 0.85 to 0.98, and the performance decrease controlled at 15.3%. This result fully demonstrates that by introducing learnable physical prior constraints, this application can effectively compensate for the information loss caused by missing data, and can still maintain high accuracy and strong stability in prediction capabilities in complex and ever-changing real-world observation environments.

[0071] To verify the adaptability of the method in this application to extreme weather, a comparative experiment was conducted using a typhoon as an example. As shown in Table 4: Table 4 Comparison of forecast performance in the first week after a typhoon

[0072] As shown in Table 4, without a specific adaptation mechanism, the root mean square error (RMSE) of the MLP after a typhoon is as high as 2.83, indicating that the model is almost completely ineffective; the RMSE of Q10 is 2.41, also losing its predictive ability; and the RMSE of FixedGate is 2.17, showing a severe performance degradation. The method proposed in this application, however, can quickly adapt to extreme environments with an RMSE of 1.07. This result demonstrates that the method proposed in this application can quickly activate and dominate predictions under extreme weather conditions, effectively correcting data distortion and thus ensuring the robustness and reliability of the model under harsh conditions.

[0073] Figure 2 This is a schematic diagram of the structure of a tropical forest soil carbon flux prediction system based on multi-source data fusion, as provided in an embodiment of this application. Figure 2 As shown, this application proposes a multi-source data fusion system for predicting soil carbon flux in tropical forests. The system 200 includes: a data acquisition module 210, a first prediction module 220, a second prediction module 230, a weight determination module 240, and a weighted fusion module 250. The data acquisition module 210 is used to acquire multi-source environmental monitoring data, which includes soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers. The first prediction module 220 is used to generate a first soil carbon flux prediction value based on the soil temperature and the extreme weather event marker, using a pre-determined exponential function in a pre-trained bi-branch prediction model. The second prediction module 230 is used to generate a second soil carbon flux prediction value and prediction reliability based on all the multi-source environmental monitoring data and through the multilayer perceptron pre-trained in the pre-trained dual-branch prediction model. The weight determination module 240 is used to calculate the fusion weights based on the prediction reliability and the extreme weather event label using a pre-determined gating function. The weighted fusion module 250 is used to perform weighted fusion of the second soil carbon flux prediction value and the first soil carbon flux prediction value based on the fusion weight to obtain the soil carbon flux prediction value.

[0074] It should be noted that the description of the above system embodiments is similar to the description of the above method embodiments, and has similar beneficial effects. For technical details not disclosed in the system embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0075] It should be noted that, in the embodiments of this application, if the above-mentioned method for predicting tropical forest soil carbon flux through multi-source data fusion is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0076] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements the steps in the multi-source data fusion method for predicting tropical forest soil carbon flux described in any of the above embodiments. Correspondingly, embodiments of this application also provide a computer program product, which, when executed by a processor of an electronic device, is used to implement the steps in the multi-source data fusion method for predicting tropical forest soil carbon flux described in any of the above embodiments.

[0077] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0078] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0079] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0080] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. In addition, each functional unit in the embodiments of this application may be fully integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in the form of hardware plus software functional units.

[0081] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0082] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.

[0083] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting soil carbon flux in tropical forests using multi-source data fusion, characterized in that, The method includes: Acquire multi-source environmental monitoring data, including soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers; Based on the soil temperature and the extreme weather event markers, a first predicted soil carbon flux value is generated using a pre-determined exponential function in a pre-trained bi-branch prediction model. Based on all the multi-source environmental monitoring data, a second soil carbon flux prediction value and prediction reliability are generated through the multilayer perceptron pre-trained in the pre-trained dual-branch prediction model. Based on the prediction reliability and the extreme weather event label, the fusion weights are calculated using a pre-determined gating function; The second predicted soil carbon flux and the first predicted soil carbon flux are weighted and fused based on the fusion weight to obtain the predicted soil carbon flux.

2. The method according to claim 1, characterized in that, The predetermined exponential function includes an adjustable temperature response coefficient; The process of generating a first predicted soil carbon flux value based on the soil temperature and the extreme weather event markers, using a pre-determined exponential function in a pre-trained bi-branch prediction model, includes: When the extreme weather event marker indicates the presence of an extreme weather event, the adjustable temperature response coefficient is multiplied by a preset adjustment coefficient to obtain the adjusted temperature response coefficient; the preset adjustment coefficient is predetermined based on the rate of change of the temperature sensitivity coefficient in historical extreme weather events. The exponential function is updated based on the adjusted temperature response coefficient to obtain the adjusted exponential function; Based on the adjusted exponential function and the soil temperature, the first predicted value of soil carbon flux is obtained.

3. The method according to claim 1, characterized in that, The process of generating a second predicted soil carbon flux value and its reliability based on all the multi-source environmental monitoring data, using a pre-trained multilayer perceptron in the pre-trained dual-branch prediction model, includes: All the multi-source environmental monitoring data are input into the multilayer sensor. During each forward propagation, some neurons are randomly deactivated due to the discard operation, thereby obtaining multiple candidate predicted values ​​for soil carbon flux. The second predicted soil carbon flux value is obtained by taking the arithmetic mean of the multiple candidate predicted soil carbon flux values. Calculate the standard deviation of the multiple candidate predicted soil carbon flux values ​​to obtain the dispersion index; The prediction reliability is obtained based on the dispersion index and the second predicted soil carbon flux value.

4. The method according to claim 1, characterized in that, The gating function is a logistic function; the calculation of fusion weights based on the prediction reliability and the extreme weather event label using a pre-determined gating function includes: The gated input value is obtained by summing the negative weighted value of the prediction reliability and the positive weighted value of the extreme weather event label. The gated input value is input to the logistic function to obtain the fusion weight; wherein, the lower the prediction reliability, the smaller the fusion weight.

5. The method according to claim 1, characterized in that, The step of weighting and fusing the second predicted soil carbon flux value and the first predicted soil carbon flux value based on the fusion weight to obtain the predicted soil carbon flux value includes: The second predicted soil carbon flux value is weighted using the fusion weights to obtain the first weighted component; Using the fusion weights, the complementary weights corresponding to the first predicted soil carbon flux value are determined; The first predicted soil carbon flux is weighted using the complementary weights to obtain a second weighted component; The first weighted component is added to the second weighted component to obtain the predicted value of soil carbon flux.

6. The method according to claim 1, characterized in that, The method further includes: Acquire initial multi-source environmental monitoring data; The initial multi-source environmental monitoring data is time-aligned, missing data is filled, and standardized to obtain the multi-source environmental monitoring data.

7. The method according to claim 1, characterized in that, The multi-source environmental monitoring data also includes a missing mask channel; the missing mask channel is used to identify each feature data in the multi-source environmental monitoring data as a measured value or a filled value.

8. The method according to claim 1, characterized in that, The pre-trained multilayer perceptron is trained through the following steps: Acquire historical multi-source environmental monitoring data and corresponding historical measured values ​​of soil carbon flux; All the historical multi-source environmental monitoring data are input into the multilayer perceptron to be trained, and multiple forward propagations with drop-off operations are performed on the multilayer perceptron to be trained to obtain multiple candidate predicted values ​​of historical soil carbon flux. The mean of the multiple candidate predicted historical soil carbon flux values ​​is calculated to obtain the second predicted historical soil carbon flux value. The loss value is calculated based on the difference between the predicted value of the second historical soil carbon flux and the measured value of the historical soil carbon flux. Based on the loss value, the parameters of the multilayer perceptron to be trained are updated through the backpropagation algorithm to obtain the pre-trained multilayer perceptron.

9. A multi-source data fusion system for predicting soil carbon flux in tropical forests, characterized in that, The system includes: a data acquisition module, a first prediction module, a second prediction module, a weight determination module, and a weighted fusion module; The data acquisition module is used to acquire multi-source environmental monitoring data, which includes soil temperature, soil moisture, air temperature, precipitation, soil organic matter, and extreme weather event markers. The first prediction module is used to generate a first soil carbon flux prediction value based on the soil temperature and the extreme weather event marker, using a pre-determined exponential function in a pre-trained bi-branch prediction model. The second prediction module is used to generate a second soil carbon flux prediction value and prediction reliability based on all the multi-source environmental monitoring data and through the multilayer perceptron pre-trained in the pre-trained dual-branch prediction model. The weight determination module is used to calculate the fusion weights based on the prediction reliability and the extreme weather event label using a pre-determined gating function. The weighted fusion module is used to perform weighted fusion of the second soil carbon flux prediction value and the first soil carbon flux prediction value based on the fusion weight to obtain the soil carbon flux prediction value.

Images