A method, system, medium and device for estimating GPP multi-task learning with combined ground and satellite observation constraints

Abstract

The application discloses a kind of GPP multitask learning estimation method, system, medium and equipment of joint ground and satellite observation constraint, belong to remote sensing technical field, method includes obtaining OCO-2 SIF data and TROPOMI SIF data, global GPP product prediction value is exported by expert network, processing flow includes: based on OCO-2 SIF data set, TROPOMI SIF data set extracts initial feature input shared encoder module, through feature alignment and fusion mechanism learning cross-sensor space-time feature, output OCO-2 SIF time feature and TROPOMI SIF space feature and carry out separate decoding output prediction value and prediction value, and then introduce joint loss function to realize high-precision estimation of global GPP by joint optimization, the application solves the problem of insufficient space-time generalization caused by single task model estimation, causes the error mode of high value underestimation and low value overestimation.

Description

TECHNICAL FIELD

[0001] The application relates to a GPP multi-task learning estimation method, system, medium and equipment combining ground and satellite observation constraints, and belongs to the technical field of remote sensing. BACKGROUND

[0002] GPP (Gross Primary Productivity) is the total amount of carbon fixed by plants through photosynthesis, and as a key component of the terrestrial carbon cycle, accurate quantification of global GPP is essential for understanding the changes in the terrestrial carbon cycle and its response to global change. Currently, the traditional methods for quantifying GPP mainly include site and model simulation methods. Among them, the eddy correlation (EC, Eddy Covariance) technique is the most reliable method for estimating ecosystem-level GPP indirectly through different algorithms. However, flux sites are sparse in distribution, have different coverage periods, and have limited response footprints. Therefore, at the global scale, GPP is mainly simulated by models, including process-based models, light use efficiency (LUE) models, machine learning models, and models based on solar-induced chlorophyll fluorescence estimation. In recent years, solar-induced chlorophyll fluorescence (SIF) has been introduced as a new method. SIF is a fluorescence signal that re-emits energy in the form of long-wave radiation during photosynthesis, which can directly reflect the physiological state of vegetation and photosynthetic performance. Its close relationship with GPP has been confirmed by multiple studies. The estimation methods based on SIF mainly include statistical models, nonlinear models, and mechanism models. At the same time, several SIF satellite products with different spatial and temporal resolutions have been released. Among them, SCIAMACHY (Scanning Imaging Absorption spectrometer for Atmospheric Cartography), GOME-2 (Global Ozone Monitoring Experiment-2), and GOSAT (Greenhouse Gases Observing Satellite) can provide long-term coverage but have lower resolution and higher uncertainty; OCO-2 (Orbiting Carbon Observatory-2) has finer resolution but discontinuous spatial coverage; TROPOMI (Tropospheric Monitoring Instrument) has high spatial and temporal resolution but shorter time record. In addition, some studies use machine learning methods to build the relationship between driving data and EC GPP (Eddy Covariance Gross Primary Production), and use reconstructed satellite SIF images to estimate GPP, resulting in large uncertainty.

[0003] There are many limitations in GPP estimation in the prior art. The traditional method usually involves a large number of assumptions, and its accuracy is limited by the uncertainty of model parameters. Although the SIF satellite product provides a new way for GPP estimation, the single sensor data has limitations in spatial and temporal coverage: the resolutions of SCIAMACHY, GOME-2 and GOSAT are low and the uncertainty is high; the spatial coverage of OCO-2 is discontinuous; and the time record of TROPOMI is short, all of which are difficult to fully characterize the photosynthesis dynamics of complex ecological systems. Although some studies have reconstructed satellite SIF images to obtain spatial and temporal continuous data, they have not fully utilized the temporal and spatial complementary information of multi-source SIF products such as OCO-2 and TROPOMI for constraint. In addition, the GPP estimation method based on machine learning mostly relies on single SIF data, and also has the problem of not fully utilizing the multi-source temporal and spatial complementary information, and due to the lack of global-scale SIF direct observation data, the estimation results still have great uncertainty in space and time. SUMMARY

[0004] The purpose of the present application is to provide a GPP multi-task learning estimation method, system, medium and equipment constrained by joint ground and satellite observations, which realizes cross-sensor feature alignment and dynamic fusion by combining multi-source SIF data of OCO-2 and TROPOMI with ground EC GPP observations, and optimizes by combining a gating attention mechanism and a multi-task joint loss function, so as to solve the problem of insufficient spatio-temporal generalization caused by single task model estimation, and the error mode of underestimation of high values and overestimation of low values.

[0005] To solve the above technical problems, the present application is realized by adopting the following technical solutions:

[0006] In a first aspect, the present application provides a GPP multi-task learning estimation method constrained by joint ground and satellite observations, comprising:

[0007] obtaining OCO-2 SIF (Orbiting Carbon Observatory-2 Sun-Induced Chlorophyll Fluorescence) data and TROPOMI SIF (Tropospheric Monitoring Instrument Sun-Induced Chlorophyll Fluorescence) data;

[0008] based on the OCO-2 SIF data and the TROPOMI SIF data, processing through a pre-trained expert network to output a global GPP product prediction value;

[0009] The data processing flow of the expert network comprises:

[0010] The OCO-2 SIF dataset and the TROPOMI SIF dataset are respectively input into two layers of fully connected neural networks to extract initial features, so as to obtain OCO-2 SIF initial features and TROPOMI SIF initial features;

[0011] The OCO-2 SIF initial features and the TROPOMI SIF initial features are input into a shared encoder module, and through a feature alignment and fusion mechanism, the cross-sensor spatio-temporal features are learned, so as to output the time features of the OCO-2 SIF and the space features of the TROPOMI SIF;

[0012] The time features of the OCO-2 SIF and the space features of the TROPOMI are separately decoded, so as to output a predicted value of GPP estimated from OCO-2 SIF products (GPPestimated from OCO-2 SIF products) and a predicted value of GPP estimated from TROPOMI SIF products (GPPestimated from TROPOMI SIF products);

[0013] Through a fusion layer, the predicted value of GPPestimated from OCO-2 SIF products and the predicted value of GPPestimated from TROPOMI SIF products are dynamically weighted and fused to obtain a predicted value of a global GPP product;

[0014] During the training of the expert network, the EC GPP observation sample is taken as a training label, and the network parameters are optimized by minimizing the difference between the predicted value of the global GPP product and the true value.

[0015] Further, the expert network is a multi-task learning neural network model with a multi-gate hybrid expert architecture, comprising:

[0016] An input layer is configured to receive the OCO-2 SIF dataset and the TROPOMI SIF dataset as inputs;

[0017] A feature extraction layer comprises two layers of fully connected neural networks, each layer comprising 256 neurons and adopting a ReLU activation function, and is configured to extract OCO-2 SIF initial features and TROPOMI SIF initial features according to the OCO-2 SIF dataset and the TROPOMI SIF dataset and input the OCO-2 SIF initial features and the TROPOMI SIF initial features into a shared encoder module;

[0018] ​​The shared encoder module comprises a plurality of expert sub-networks, and is configured to learn cross-sensor spatio-temporal features according to OCO-2 SIF initial features and TROPOMI SIF initial features through a feature alignment and fusion mechanism, so as to obtain temporal features of the OCO-2 SIF and spatial features of the TROPOMI SIF;

[0019] The first gating structure is configured to construct a relationship between the temporal features of the OCO-2 SIF and the EC GPP observation samples by using an attention mechanism, and input the first decoder;

[0020] The second gating structure is configured to construct a relationship between the spatial features of the TROPOMI SIF and the EC GPP observation samples by using an attention mechanism, and input the second decoder;

[0021] The first decoder comprises two fully connected layers and a Sigmoid activation function, and is configured to decode the temporal features of the OCO-2 SIF to obtain a prediction value;

[0022] The second decoder comprises two fully connected layers and a Sigmoid activation function, and is configured to decode the spatial features of the TROPOMI SIF to obtain a prediction value;

[0023] The fusion layer is configured to dynamically weight fuse the prediction value and the prediction value by using an average loss function to obtain a global GPP product prediction value;

[0024] The output layer is configured to output the global GPP product prediction value.

[0025] Further, the expert sub-networks comprise:

[0026] The data input layer is configured to receive the OCO-2 SIF initial features or the TROPOMI SIF initial features output by the shared encoder module;

[0027] The fully connected layer comprises 128 neurons, and is configured to perform nonlinear mapping on the OCO-2 SIF initial features or the TROPOMI SIF initial features by using linear transformation, and output the mapped features to the activation function layer;

[0028] The activation function layer is configured to perform nonlinear activation on the mapped features by using a ReLU function.

[0029] Further, the training method of the expert network comprises:

[0030] The OCO-2 SIF dataset, the TROPOMI SIF dataset and the EC GPP observation samples are obtained, and a training set is constructed.​​​​

[0031] Based on the training set, the OCO-2 SIF dataset and the TROPOMI SIF dataset are respectively input into the expert network;

[0032] OCO-2 SIF initial features and TROPOMI SIF initial features are extracted by using the feature extraction layer respectively;

[0033] According to the OCO-2 SIF initial features and the TROPOMI SIF initial features, cross-sensor spatio-temporal features are learned through a feature alignment and fusion mechanism, to obtain time features of the OCO-2 SIF and spatial features of the TROPOMI SIF;

[0034] The attention mechanism is adopted to construct the relationship between the time features of the OCO-2 SIF and the EC GPP observation samples, and the first decoder is input;

[0035] The attention mechanism is adopted to construct the relationship between the spatial features of the TROPOMI SIF and the EC GPP observation samples, and the second decoder is input;

[0036] The time features of the OCO-2 SIF are decoded by using the first decoder to obtain a predicted value;

[0037] The spatial features of the TROPOMI SIF are decoded by using the second decoder to obtain a predicted value;

[0038] The average loss function is adopted to dynamically weight and fuse the predicted value and the predicted value to obtain a global GPP product predicted value;

[0039] Based on the joint loss function, the EC GPP observation samples are taken as training labels, and the network parameters are optimized by minimizing the difference between the global GPP product predicted value and the true value, to obtain the trained expert network.

[0040] Further, the OCO-2 SIF dataset and the TROPOMI SIF dataset and the EC GPP observation samples are obtained, including:

[0041] Collect ground station data, satellite observation data, meteorological data, and obtain MERRA-2 (Modern-era Retrospective Analysis for Research and Applications, Version 2) data and CarbonTracker carbon dioxide data, clean and match the spatial and temporal scales to obtain preprocessed data sources;

[0042] According to the preprocessed data sources, global OCO-2 SIF, vegetation leaf area index, vegetation near-infrared reflectivity and vegetation type data, absorbed effective radiation, air temperature, saturated water vapor pressure difference data and carbon dioxide are extracted as input samples;

[0043] According to the preprocessed data sources, global TROPOMI SIF, vegetation leaf area index, vegetation near-infrared reflectivity and vegetation type data, surface incident shortwave flux, air temperature, saturated water vapor pressure difference data and carbon dioxide are extracted as input samples;

[0044] According to the EC GPP ground station data in the FLUXNET2015 dataset with accuracy greater than a preset threshold, observation samples of a specified time scale are synthesized;

[0045] The input samples are taken as explanatory variable data, and the EC GPP observation samples are taken as label data to construct an OCO-2 SIF dataset;

[0046] The input samples are taken as explanatory variable data, and the EC GPP observation samples are taken as label data to construct a TROPOMI SIF dataset.

[0047] Further, the mapping relationship between the label data and the explanatory variable data is represented as:

[0048] ;

[0049] In the formula, represents the global GPP product predicted by the OCO-2 SIF inversion task, represents the global GPP product predicted by the TROPOMI SIF inversion task, represents the mapping relationship function between the label data and the explanatory variable data that needs to be learned by the expert network, represents an OCO-2 SIF dataset or a TROPOMI SIF dataset, represents vegetation leaf area index, represents vegetation near-infrared reflectance, represents vegetation type data, represents absorbed effective radiation, represents air temperature, represents saturated water vapor pressure difference data, represents carbon dioxide.

[0050] Further, the joint loss function is represented as:

[0051] ;

[0052] In the formula, represents a loss value of the joint loss function, represents a weight coefficient of the OCO-2 SIF inversion task loss function , represents a weight coefficient of the TROPOMI SIF inversion task loss function .

[0053] In a second aspect, the present application provides a GPP multi-task learning estimation system jointly constrained by ground and satellite observations, characterized by being used to implement the steps of the GPP multi-task learning estimation method jointly constrained by ground and satellite observations as described in the first aspect.

[0054] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the GPP multi-task learning estimation method jointly constrained by ground and satellite observations as described in the first aspect.

[0055] In a fourth aspect, the present application provides a computer device, characterized by comprising:

[0056] a memory for storing instructions;

[0057] a processor for executing the instructions, so that the device performs operations to implement the GPP multi-task learning estimation method jointly constrained by ground and satellite observations as described in the first aspect.

[0058] Compared with the prior art, the present application has the following beneficial effects:

[0059] 1. The application effectively extracts cross-sensor spatio-temporal features by constructing an expert network combined with ground and satellite observation constraints, combining the feature alignment and fusion mechanism of OCO-2 and TROPOMI multi-source SIF data, and using EC GPP observation samples as training labels for parameter optimization, realizing the collaborative constraint of multi-source satellite data and ground observation, thereby significantly improving the accuracy of global GPP product prediction value, solving the problem of insufficient spatio-temporal generalization caused by existing single task model estimation, causing the error mode of underestimation of high value and overestimation of low value.

[0060] 2. The application effectively extracts cross-sensor spatio-temporal features of OCO-2 SIF and TROPOMI SIF data by sharing the feature alignment and fusion mechanism of the encoder module, combining the attention mechanism of the multi-gate hybrid expert architecture, and screening the shared information highly related to the ECGPP observation sample, realizing efficient collaborative use of multi-source satellite data, and significantly enhancing the spatial coverage integrity and data consistency of global GPP product prediction.

[0061] 3. The application uses EC GPP ground observation samples as training labels, optimizes network parameters by minimizing the difference between global GPP product prediction value and true value, and adjusts the weight coefficients of OCO-2 SIF inversion task and TROPOMI SIF inversion task combined with joint loss function, effectively improving the prediction accuracy of the model on a global scale, especially in areas with sparse ground observations, satellite data and ground observations are used to make up for data gaps and enhance the reliability of the results.

[0062] 4. The fusion layer adopts average loss function to dynamically weight the two-way decoding prediction value, combines the nonlinear mapping capability of the expert subnetwork, and adaptively adjusts the contribution proportion of different sensor data in global prediction, effectively solving the problem of local deviation caused by single sensor spatial resolution, observation time or instrument error, improving the spatio-temporal continuity and accuracy of global GPP product. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 is a flowchart of a GPP multi-task learning estimation method combined with ground and satellite observation constraints provided by an embodiment of the application;

[0064] Figure 2 is a structure diagram of an expert network provided by an embodiment of the application;

[0065] Figure 3 is a comparison diagram of prediction results of different time scale TROPOMI SIF inversion task, OCO-2 SIF inversion task and expert network inversion task every 8 days and every month provided by an embodiment of the application. DETAILED DESCRIPTION

[0066] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0067] Example 1

[0068] like Figure 1 As shown, this embodiment introduces a multi-task learning estimation method for GPP with joint ground and satellite observation constraints, including:

[0069] Step 1: Obtain OCO-2 SIF data and TROPOMI SIF data.

[0070] The SIF data from OCO-2 and TROPOMI satellites differ in spatial resolution, observation frequency, and coverage. For example, OCO-2 has a higher spatial resolution but a narrower coverage, while TROPOMI has a wider coverage but a lower resolution. By acquiring both OCO-2 and TROPOMI SIF data simultaneously, high-resolution and wide-area observations can complement each other, providing a more comprehensive spatial information foundation for global GPP estimation.

[0071] Sun-induced chlorophyll fluorescence (SIF) is a direct product of vegetation photosynthesis and has a high linear correlation with total primary productivity (GPP). Obtaining two types of OCO-2 SIF data and TROPOMI SIF data can double-verify vegetation photosynthetic activity, reduce observation bias caused by instrument errors or environmental interference such as cloud cover and atmospheric interference caused by a single sensor, and improve the reliability of input data.

[0072] Step 2: Based on the OCO-2 SIF data and TROPOMI SIF data, the data is processed through a pre-trained expert network to output the global GPP product prediction value.

[0073] In this invention, the data processing flow of the expert network includes:

[0074] The OCO-2 SIF dataset and the TROPOMI SIF dataset are respectively input into a two-layer fully connected neural network to extract initial features, resulting in the initial features of OCO-2 SIF and TROPOMI SIF.

[0075] The application captures the complex high-order nonlinear relationship in the OCO-2 SIF and TROPOMI SIF data by the combination of linear transformation and nonlinear activation of the two-layer fully connected neural network, converts the OCO-2 SIF and TROPOMI SIF data into more discriminative initial features, can retain the high-resolution detail features of OCO-2 and the large-range trend features of TROPOMI, avoid the loss of sensor characteristics caused by mixed input, and provide differentiated input for subsequent cross-sensor feature alignment.

[0076] The OCO-2 SIF initial features and TROPOMI SIF initial features are input into a shared encoder module, the cross-sensor spatio-temporal features are learned through feature alignment and fusion mechanism, and the time features of OCO-2 SIF and the space features of TROPOMI SIF are output.

[0077] The application eliminates the feature distribution inconsistency problem of OCO-2 SIF and TROPOMI SIF data caused by sensor parameter difference through feature alignment, so that OCO-2 SIF and TROPOMI SIF data have comparability in the shared encoder; common patterns related to GPP in the two types of SIF data are mined through fusion to generate cross-sensor spatio-temporal features, which reduces redundant information and improves the compactness of feature expression; the learning of spatio-temporal features enhances the adaptability of the model to different regions around the world, avoiding overfitting to single sensor data.

[0078] The time features of OCO-2 SIF and the space features of TROPOMI SIF are decoded separately, and predicted values and predicted values are output.

[0079] The application uses the gating structure to automatically learn the correlation weight of the EC GPP observation sample and the OCO-2 SIF and TROPOMI SIF spatio-temporal features, filters out the shared information highly related to GPP prediction, suppresses irrelevant noise, and improves the input quality of the decoder; the combination of two-layer fully connected layer and Sigmoid activation function maps the filtered shared information to the physical dimension of GPP, captures the complex mapping relationship between features and GPP through nonlinear transformation, and can output predicted values closer to the true values.

[0080] The global GPP product predicted value is obtained by dynamically weighting and fusing predicted values and predicted values through the fusion layer.

[0081] The global GPP product predicted value is obtained by dynamically weighting and fusing predicted values and The predicted values are dynamically weighted and fused to obtain a global GPP product predicted value, which balances local accuracy and global consistency, and reduces local deviation caused by coverage blank or error of a single sensor; an average loss function quantifies the difference between the two types of predicted values and the true value, and dynamically adjusts the fusion weight, so that the model automatically corrects the system error of the two types of sensors in the training, and improves the spatiotemporal continuity of the global GPP product.

[0082] During the training of the expert network, the EC GPP observation sample is used as a training label, and the network parameters are optimized by minimizing the difference between the global GPP product predicted value and the true value.

[0083] The present application uses the EC GPP ground observation sample as a training label, minimizes the difference between the predicted value and the true value, forces the expert network to learn the true physical relationship between the SIF data and the GPP, avoids the deviation of the model caused by the error of the satellite data itself, balances the contribution of the two types of sensors to the global GPP through the joint loss function, avoids the dominance of a certain sensor in the training, ensures that the expert network can adapt to the two types of data at the same time, and improves the balance and accuracy of the global prediction.

[0084] Embodiment 2

[0085] Step 1: Obtain OCO-2 SIF data and TROPOMI SIF data.

[0086] Step 2: Based on the OCO-2 SIF data and the TROPOMI SIF data, the expert network is processed through pre-training, and a global GPP product predicted value is output.

[0087] Step 2.1: Train the expert network:

[0088] Step 2.1.1: Obtain OCO-2 SIF data set and TROPOMI SIF data set and EC GPP observation sample, and construct a training set.

[0089] Collect ground station data, satellite observation data and meteorological data, and obtain MERRA-2 data and CarbonTracker carbon dioxide data, and perform cleaning and spatiotemporal scale matching to obtain preprocessed data sources;

[0090] According to the preprocessed data sources, global OCO-2 SIF, vegetation leaf area index, vegetation near-infrared reflectivity and vegetation type data, absorbed effective radiation, air temperature, saturated water vapor pressure difference data and carbon dioxide are extracted as input samples;

[0091] According to the preprocessed data source after preprocessing, global TROPOMI SIF, vegetation leaf area index, vegetation near-infrared reflectivity and vegetation type data, surface incident shortwave flux, air temperature, saturated water vapor pressure difference data and carbon dioxide are extracted as Input samples;

[0092] According to the EC GPP ground station data in the FLUXNET2015 dataset with accuracy greater than the preset threshold, observation samples of a specified time scale are synthesized;

[0093] The Input samples are used as explanatory variable data, and the EC GPP observation samples are used as Label data to construct an OCO-2 SIF dataset;

[0094] The Input samples are used as explanatory variable data, and the EC GPP observation samples are used as Label data to construct a TROPOMI SIF dataset.

[0095] In this embodiment, the mapping relationship between the label data and the explanatory variable data is represented as:

[0096] ;

[0097] In the formula, represents a global GPP product predicted by an OCO-2 SIF inversion task, represents a global GPP product predicted by a TROPOMI SIF inversion task, represents a mapping relationship function between label data and explanatory variable data that needs to be learned by an expert network, represents an OCO-2 SIF dataset or a TROPOMI SIF dataset, represents a vegetation leaf area index, represents a vegetation near-infrared reflectivity, represents vegetation type data, represents an absorption effective radiation, represents an air temperature, represents saturated water vapor pressure difference data, represents carbon dioxide.

[0098] Step 2.1.2: Based on the training set, the OCO-2 SIF dataset and the TROPOMI SIF dataset are respectively input into the expert network.

[0099] Step 2.1.3: The OCO-2 SIF initial features and the TROPOMI SIF initial features are respectively extracted by using the feature extraction layer.

[0100] Step 2.1.4: Learning the cross-sensor spatio-temporal features by feature alignment and fusion mechanism according to the OCO-2 SIF initial features and the TROPOMI SIF initial features, to obtain the temporal features of the OCO-2 SIF and the spatial features of the TROPOMI SIF.

[0101] Step 2.1.5: Building the relationship between the temporal features of the OCO-2 SIF and the EC GPP observation samples by using the attention mechanism, and inputting the first decoder.

[0102] Step 2.1.6: Building the relationship between the spatial features of the TROPOMI SIF and the EC GPP observation samples by using the attention mechanism, and inputting the second decoder.

[0103] Step 2.1.7: Decoding the temporal features of the OCO-2 SIF by using the first decoder to obtain the predicted value.

[0104] Step 2.1.8: Decoding the spatial features of the TROPOMI SIF by using the second decoder to obtain the predicted value.

[0105] Step 2.1.9: Dynamically weighting and fusing the predicted value and the predicted value by using the average loss function to obtain the global GPP product predicted value.

[0106] Step 2.1.10: Based on the joint loss function, taking the EC GPP observation samples as the training labels, optimizing the network parameters by minimizing the difference between the global GPP product predicted value and the true value, to obtain the trained expert network.

[0107] In this embodiment, the joint loss function is represented as:

[0108] ;

[0109] In the formula, represents the loss value of the joint loss function, represents the weight coefficient of the OCO-2 SIF inversion task loss function , and represents the weight coefficient of the TROPOMI SIF inversion task loss function .

[0110] Step 2.2: Processing the OCO-2 SIF data and the TROPOMI SIF data by using the expert network:

[0111] Step 2.2.1: The OCO-2 SIF dataset and the TROPOMI SIF dataset are respectively input into two layers of fully connected neural networks to extract initial features, obtaining OCO-2 SIF initial features and TROPOMI SIF initial features.

[0112] Step 2.2.2: The OCO-2 SIF initial features and the TROPOMI SIF initial features are input into a shared encoder module to learn cross-sensor spatio-temporal features through feature alignment and fusion mechanism, outputting OCO-2 SIF temporal features and TROPOMI SIF spatial features.

[0113] Step 2.2.3: The OCO-2 SIF temporal features and the TROPOMI SIF spatial features are separately decoded, outputting predicted values and predicted values.

[0114] Step 2.2.4: The predicted values and predicted values are dynamically weighted and fused by a fusion layer to obtain global GPP product predicted values.

[0115] Wherein, during the training of the expert network, the EC GPP observation samples are used as training labels, and the network parameters are optimized by minimizing the difference between the global GPP product predicted values and the true values.

[0116] In this embodiment, the expert network is a multi-task learning neural network model with a multi-gated hybrid expert architecture, as shown in Figure 2 , which includes an input layer, a feature extraction layer, a shared encoder module, a first gating structure, a second gating structure, a first decoder, a second decoder, a fusion layer, and an output layer.

[0117] In the embodiment, the input layer is configured to receive the OCO-2 SIF dataset and the TROPOMI SIF dataset as inputs; the feature extraction layer includes two fully connected neural networks, each of which contains 256 neurons and adopts a ReLU activation function, and is configured to extract OCO-2 SIF initial features and TROPOMI SIF initial features according to the OCO-2 SIF dataset and the TROPOMI SIF dataset and input the OCO-2 SIF initial features and the TROPOMI SIF initial features into a shared encoder module; the shared encoder module includes a plurality of expert sub-networks, and is configured to learn cross-sensor spatio-temporal features according to the OCO-2 SIF initial features and the TROPOMI SIF initial features through a feature alignment and fusion mechanism, and obtain time features of the OCO-2 SIF and spatial features of the TROPOMI SIF. The first gating structure is configured to adopt an attention mechanism to construct a relationship between the time features of the OCO-2 SIF and the ECGPP observation samples, and input the time features of the OCO-2 SIF into a first decoder; the second gating structure is configured to adopt an attention mechanism to construct a relationship between the spatial features of the TROPOMI SIF and the ECGPP observation samples, and input the spatial features of the TROPOMI SIF into a second decoder; the first decoder includes two fully connected layers and a Sigmoid activation function, and is configured to decode the time features of the OCO-2 SIF to obtain a predicted value; the second decoder includes two fully connected layers and a Sigmoid activation function, and is configured to decode the spatial feature information of the TROPOMI SIF to obtain a predicted value; the fusion layer is configured to dynamically weight and fuse the predicted values and the predicted values by using an average loss function to obtain a global GPP product predicted value; and the output layer is configured to output the global GPP product predicted value.

[0118] In the embodiment, the expert sub-network includes a data input layer, a fully connected layer and an activation function layer; the data input layer is configured to receive the OCO-2 SIF initial features or the TROPOMI SIF initial features output by the shared encoder module. The fully connected layer contains 128 neurons, adopts linear transformation to perform nonlinear mapping on the OCO-2 SIF initial features or the TROPOMI SIF initial features, and outputs the mapped features to the activation function layer. The activation function layer adopts a ReLU function to perform nonlinear activation on the mapped features.

[0119] Step 3: Model performance evaluation.

[0120] The embodiment verifies the prediction performance of the expert network through a multi-dimensional evaluation method, and ensures the accuracy and reliability of the global GPP product predicted value. The specific evaluation method is as follows:

[0121] Quantitative evaluation: Based on the EC GPP ground observation sample, the determination coefficient R2, root mean square error RMSE and mean absolute error MAE are used as core indicators to quantify the matching degree of the global GPP product prediction value output by the expert network and the true value. Among them, the larger the R2 indicates that the model explanation ability is stronger, and the smaller the RMSE and MAE indicates that the prediction deviation is lower.

[0122] Qualitative evaluation: Combined with global-scale typical GPP products such as TL-LUE, BESS, FluxCom and GOSIFGPP, the spatial distribution pattern output by the expert network is compared and analyzed to verify the spatial consistency with mainstream products and the adaptability to complex vegetation-climate regions.

[0123] Example verification results:

[0124] Figure 3 Based on the station leave-one-out method to verify the accuracy and spatial generalization of the model, the results show that the prediction results of the expert network inversion task are better than the TROPOMI inversion task and the OCO-2 inversion task at different time scales of every 8 days and every month. Not only is the estimation accuracy higher, but also the spatial generalization ability is stronger.

[0125] In summary, through the dual verification of quantitative indicators and qualitative spatial patterns, the global GPP product output by the expert network provided in this embodiment has high precision, strong generalization and robustness, and meets the needs of global-scale carbon cycle research.

[0126] Embodiment 3

[0127] Based on the same inventive concept as embodiment 1, this embodiment introduces a GPP multi-task learning estimation system combined with ground and satellite observation constraints, which is used to implement the steps of the above-mentioned embodiments 1 or 2.

[0128] For specific function implementation, refer to the related content in the method of embodiment 1, which will not be repeated here.

[0129] Embodiment 4

[0130] Based on the same inventive concept as other embodiments, this embodiment introduces a computer readable storage medium, which stores computer instructions, and the computer instructions are executed by a processor to implement the steps of the above-mentioned embodiments 1 or 2.

[0131] Embodiment 5

[0132] Based on the same inventive concept as other embodiments, this embodiment introduces a computer device, which includes:

[0133] a memory for storing instructions;

[0134] A processor configured to execute the instructions to cause the device to perform operations implementing the method of any of embodiments 1 or 2.

[0135] In summary, the present application effectively extracts cross-sensor spatio-temporal features by constructing an expert network jointly constrained by ground and satellite observations, combining the feature alignment and fusion mechanism of OCO-2 and TROPOMI multi-source SIF data, and using ECGPP observation samples as training labels for parameter optimization, realizing the collaborative constraint of multi-source satellite data and ground observations, thereby significantly improving the accuracy and spatial coverage integrity of global GPP product prediction values, solving the problem of insufficient precision and globality caused by the inability of existing GPP estimation methods to effectively combine ground and satellite observation data.

[0136] The present application effectively extracts cross-sensor spatio-temporal features of OCO-2 SIF and TROPOMI SIF data by sharing the feature alignment and fusion mechanism of the encoder module, combining the attention mechanism of the multi-gate hybrid expert architecture, and filtering the shared information highly correlated with ECGPP observation samples, realizing efficient collaborative use of multi-source satellite data, and significantly enhancing the spatial coverage integrity and data consistency of global GPP product prediction.

[0137] The present application uses ECGPP ground observation samples as training labels to optimize network parameters by minimizing the difference between global GPP product prediction values and true values, and adjusts the weight coefficients of the OCO-2 SIF inversion task and the TROPOMI SIF inversion task by combining the joint loss function, effectively improving the prediction accuracy of the model on a global scale, especially in areas with sparse ground observations, satellite data and ground observations are collaboratively constrained to make up for data gaps and enhance the reliability of the results.

[0138] The present application uses the average loss function to dynamically weight the two-way decoding prediction values, combines the nonlinear mapping capability of the expert subnetwork, and adaptively adjusts the contribution proportion of different sensor data in global prediction, effectively solving the problem of local deviation caused by single sensor spatial resolution, observation time or instrument error, and improving the spatio-temporal continuity and accuracy of global GPP products.

[0139] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer usable program code.

[0140] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0141] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0142] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0143] The embodiments of the present application described above are merely intended to illustrate the present application, but not to limit the present application. The skilled in the art can make many modifications and improvements without departing from the spirit and scope of the present application, which should be protected as long as they fall into the scope of the appended claims.

Claims

1. A multi-task learning method for estimating total primary productivity under combined ground and satellite observation constraints, characterized in that, The method comprises the following steps: obtaining OCO-2 SIF data and TROPOMI SIF data; processing the OCO-2 SIF data and the TROPOMI SIF data through a pre-trained expert network to output a global GPP product prediction value; the data processing process of the expert network comprises: inputting the OCO-2 SIF data set and the TROPOMI SIF data set into two layers of fully connected neural networks respectively to extract initial features, thereby obtaining OCO-2 SIF initial features and TROPOMI SIF initial features; inputting the OCO-2 SIF initial features and the TROPOMI SIF initial features into a shared encoder module to learn cross-sensor spatio-temporal features through a feature alignment and fusion mechanism, thereby outputting OCO-2 SIF temporal features and TROPOMI SIF spatial features; The temporal characteristics of OCO-2 SIF and the spatial characteristics of TROPOMI SIF are separately decoded, outputting a predicted value and a predicted value; Through the fusion layer Predicted values ​​and The predicted values ​​are dynamically weighted and fused to obtain the global GPP product forecast. wherein, during the training of the expert network, the EC GPP observation sample is taken as a training label, and the network parameters are optimized by minimizing the difference between the global GPP product prediction value and the true value; the expert network is a multi-task learning neural network model with a multi-gate hybrid expert architecture, comprising: an input layer for receiving the OCO-2 SIF data set and the TROPOMI SIF data set as input; a feature extraction layer comprising two layers of fully connected neural networks, each layer comprising 256 neurons and adopting a ReLU activation function, for extracting OCO-2 SIF initial features and TROPOMI SIF initial features from the OCO-2 SIF data set and the TROPOMI SIF data set and inputting them into a shared encoder module; a shared encoder module comprising a plurality of expert sub-networks, for learning cross-sensor spatio-temporal features from the OCO-2 SIF initial features and the TROPOMI SIF initial features through a feature alignment and fusion mechanism, thereby obtaining OCO-2 SIF temporal features and TROPOMI SIF spatial features; a first gating structure for constructing the relationship between the OCO-2 SIF temporal features and the EC GPP observation sample by adopting an attention mechanism, and inputting a first decoder; a second gating structure for constructing the relationship between the TROPOMI SIF spatial features and the EC GPP observation sample by adopting an attention mechanism, and inputting a second decoder; The first decoder includes two fully connected layers and a sigmoid activation function, which is used to decode the temporal features of the OCO-2 SIF to obtain Predicted values; The second decoder comprises two full connection layers and a Sigmoid activation function, and is used for decoding spatial features of the TROPOMI SIF to obtain a predicted value; The fusion layer is used to apply the average loss function to... Predicted values ​​and The predicted values ​​are dynamically weighted and fused to obtain the global GPP product forecast. an output layer for outputting a global GPP product prediction value.

2. The joint ground and satellite observation constrainted total primary productivity multi-task learning estimation method according to claim 1, characterized in that, the expert sub-network comprises: a data input layer for receiving OCO-2 SIF initial features or TROPOMI SIF initial features output by the shared encoder module; a fully connected layer comprising 128 neurons, which adopts linear transformation to perform nonlinear mapping on the OCO-2 SIF initial features or the TROPOMI SIF initial features, and outputs the mapped features to an activation function layer; an activation function layer adopting a ReLU function to perform nonlinear activation on the mapped features.

3. The joint ground and satellite observation constrainted total primary productivity multi-task learning estimation method according to claim 1, characterized in that, the training method of the expert network comprises: The OCO-2 SIF dataset and the TROPOMI SIF dataset and the ECGPP observation sample are acquired to construct a training set; Based on the training set, the OCO-2 SIF dataset and the TROPOMI SIF dataset are respectively input into the expert network; OCO-2 SIF initial features and TROPOMI SIF initial features are extracted by using a feature extraction layer; OCO-2 SIF time features and TROPOMI SIF space features are obtained by learning cross-sensor spatio-temporal features through a feature alignment and fusion mechanism according to the OCO-2 SIF initial features and the TROPOMI SIF initial features; The relationship between the OCO-2 SIF time features and the ECGPP observation sample is constructed by using an attention mechanism, and a first decoder is input; The relationship between the TROPOMI SIF space features and the ECGPP observation sample is constructed by using an attention mechanism, and a second decoder is input; Decoding the temporal features of the OCO-2 SIF with the first decoder yields predicted values; Decoding the spatial features of TROPOMI SIF with the second decoder yields predicted values; Using the average loss function Predicted values ​​and The predicted values ​​are dynamically weighted and fused to obtain the global GPP product forecast. Based on a joint loss function, network parameters are optimized by minimizing the difference between global GPP product predicted values and global GPP product true values with the ECGPP observation sample as a training label, and a trained expert network is obtained.

4. The joint ground and satellite observation constrainted total primary productivity multi-task learning estimation method according to claim 3, characterized in that, The OCO-2 SIF dataset and the TROPOMI SIF dataset and the ECGPP observation sample are acquired, including: Ground station data, satellite observation data and meteorological data are collected, and MERRA-2 data and CarbonTracker carbon dioxide data are acquired, cleaned and matched in time and space scales to obtain preprocessed data sources; According to the pre-processed data source, global OCO-2 SIF, vegetation leaf area index, vegetation near-infrared reflectivity and vegetation type data, absorption effective radiation, air temperature, saturated water vapor pressure difference data and carbon dioxide are extracted as input samples; According to the pre-processed data source, global TROPOMI SIF, vegetation leaf area index, vegetation near-infrared reflectivity and vegetation type data, surface incident shortwave flux, air temperature, saturated water vapor pressure difference data and carbon dioxide are extracted as input samples; ECGPP ground station data with accuracy greater than a preset threshold in the FLUXNET2015 dataset are collected to synthesize observation samples of a specified time scale; Will Input samples were used as explanatory variable data, and EC GPP observation samples were used as... The OCO-2 SIF dataset was constructed using the labeled data; Will Input samples were used as explanatory variable data, and EC GPP observation samples were used as... The labeled data is used to construct the TROPOMI SIF dataset.

5. The joint ground and satellite observation constrainted total primary productivity multi-task learning estimation method according to claim 4, characterized in that, The mapping relationship between the label data and the explanatory variable data is represented as: ； In the formula, represents label data, represents label data, represents a mapping relationship function between label data and interpretation variable data that needs to be learned by the expert network, represents an OCO-2 SIF data set or a TROPOMI SIF data set, represents a vegetation leaf area index, represents a vegetation near-infrared reflectance, represents vegetation type data, represents an absorption effective radiation, represents air temperature and saturation, represents water vapor pressure difference data, represents carbon dioxide.

6. The joint ground and satellite observation constrainted total primary productivity multi-task learning estimation method according to claim 3, characterized in that, The joint loss function is represented as: ； In the formula, denotes the loss value of the joint loss function, denotes the weight coefficient of the OCO-2 SIF inversion task loss function denotes the weight coefficient of the TROPOMI SIF inversion task loss function denotes the weight coefficient of the TROPOMI SIF inversion task loss function denotes the weight coefficient of the TROPOMI SIF inversion task loss function 7. A system for joint ground and satellite observation constrained gross primary productivity multi-task learning estimation, the system comprising: A computer program for implementing the steps of the joint ground and satellite observation constrained total primary productivity multi-task learning estimation method according to any one of claims 1-6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the joint ground and satellite observation constrained total primary productivity multi-task learning estimation method according to any one of claims 1-6.

9. A computer device, comprising: Including: A memory for storing instructions; A processor for executing the instructions to enable the device to perform operations for implementing the joint ground and satellite observation constrained total primary productivity multi-task learning estimation method according to any one of claims 1-6.

Images