Soil respiration prediction method based on GIS spatial-temporal characteristics and deep learning

Abstract

The invention discloses a soil respiration prediction method based on GIS spatial-temporal characteristics and deep learning, and relates to the technical field of soil respiration prediction, and the method comprises the following steps: obtaining historical land surface temperature data, historical soil moisture data, historical altitude data and historical soil respiration data of a to-be-predicted region; a multi-source fusion data set is generated through GIS geographic coordinate alignment; extracting spatial features, time features and nonlinear features, and generating a multi-source feature data set; establishing a hybrid network architecture based on deep learning; performing adaptive training to obtain a soil respiration prediction model; and performing actual soil respiration prediction by adopting the soil respiration prediction model. By combining multi-source data fusion, spatial-temporal feature enhancement, hybrid network architecture and adaptive training, multi-source data are integrated, spatial-temporal features are captured, model expression ability and interpretation are considered, and soil respiration prediction precision is improved.

Description

Technical Field

[0001] The present application relates to the technical field of soil respiration prediction, and in particular to a soil respiration prediction method based on GIS spatiotemporal features and deep learning. Background Art

[0002] Soil respiration is the primary pathway for terrestrial ecosystems to release carbon dioxide into the atmosphere, and even small changes in it can significantly impact the climate system. Accurately predicting soil respiration helps quantify terrestrial carbon sequestration capacity, assess ecological feedbacks from extreme climate events, and provide a scientific basis for agricultural carbon management. This is crucial for understanding the dynamics of the global carbon cycle and addressing climate change. Therefore, research on soil respiration prediction methods is crucial.

[0003] In the prior art, Chinese patent CN118095536A discloses a soil respiration time series prediction method, software product, and computer equipment, comprising the following steps: using correlation analysis on soil respiration sequence data and environmental factor data having time series characteristics to obtain the main environmental factors affecting soil respiration; clustering the soil respiration sequence data to obtain a number of data clusters, marking the data of each time step with a data cluster identifier, and labeling the data cluster to which the corresponding time step belongs to obtain a data cluster identifier sequence; predicting the data cluster to which the next time step belongs based on the data cluster identifier sequence; and using an ARIMA model based on a high-order multidimensional tensor to predict the data cluster to which the next time step belongs.

[0004] However, these existing technologies rely on a single data source, resulting in incomplete feature representation; they struggle to capture spatiotemporal heterogeneity and fail to account for spatial distance decay; and their predictions lack spatial interpretability. Consequently, existing technologies suffer from poor soil respiration prediction accuracy. Summary of the Invention

[0005] This application provides a soil respiration prediction method based on GIS spatiotemporal features and deep learning to solve the problem of poor soil respiration prediction accuracy in the existing technology.

[0006] On the one hand, this application provides a soil respiration prediction method based on GIS spatiotemporal features and deep learning, comprising the following steps:

[0007] Step 1: Obtain historical surface temperature data, historical soil moisture data, historical altitude data, and historical soil respiration data for the area to be predicted, and generate a multi-source fusion dataset through GIS geographic coordinate alignment.

[0008] Step 2: extracting spatial features, temporal features, and nonlinear features of the multi-source fusion data set to generate a multi-source feature data set.

[0009] Step three: Establish a hybrid network architecture based on deep learning.

[0010] Step 4: Adaptively train the hybrid network architecture using the multi-source feature dataset to obtain a soil respiration prediction model.

[0011] Step 5: Use the soil respiration prediction model to perform actual soil respiration prediction.

[0012] In a possible implementation, in step 1, the historical surface temperature data, the historical soil moisture data, the historical altitude data, and the historical soil respiration data are time-corresponding grid data.

[0013] In a possible implementation, in step 2, the extraction of the spatial features includes: calculating the spherical distance from each point in the multi-source fusion dataset to the data center based on longitude and latitude.

[0014] The extraction of the time feature includes: performing sine / cosine transformation on the time variables in the multi-source fusion data set.

[0015] The extraction of the nonlinear features includes: generating interaction terms and polynomial terms for the multi-source fusion data set.

[0016] In a possible implementation, in step 2, before calculating the spherical distance from each point in the multi-source fusion dataset to the data center based on the longitude and latitude, the longitude and latitude are normalized.

[0017] In one possible implementation, in step three, the hybrid network architecture includes: a residual connection module and an attention mechanism.

[0018] The residual connection module is used to add the input features after being mapped by the fully connected layer and the skip connection.

[0019] The attention mechanism is used to calculate feature weights.

[0020] In one possible implementation, in step 4, the adaptive training adopts PowerTransformer feature normalization, QuantileTransformer target distribution normalization, dynamic learning rate adjustment and stratified sampling strategy.

[0021] In a possible implementation, in step 4, during the adaptive training of the hybrid network architecture, the loss function adopts the Huber loss function.

[0022] In a possible implementation, in step 4, the adaptive training results of the hybrid network architecture are visualized and verified.

[0023] The soil respiration prediction method based on GIS spatiotemporal features and deep learning in this application has the following advantages:

[0024] By combining multi-source data fusion, spatiotemporal feature enhancement, hybrid network architecture and adaptive training, multi-source data fusion integrates multi-source data, spatiotemporal feature enhancement captures spatiotemporal characteristics, hybrid network architecture takes into account both model expressiveness and interpretability, and adaptive training improves model robustness. Together, these factors improve the accuracy of soil respiration prediction.

[0025] The proposed hybrid network architecture includes: residual connection module and attention mechanism. The residual connection module maps the input features through the fully connected layer and then adds them to the jump connection. The feature weights are calculated through the attention mechanism, taking into account both the model's expressiveness and interpretability.

[0026] The proposed adaptive training adopts PowerTransformer feature normalization, QuantileTransformer target distribution normalization, dynamic learning rate adjustment and stratified sampling strategy to balance the model convergence speed and generalization ability and improve the model robustness.

[0027] During the adaptive training process of the proposed hybrid network architecture, the Huber loss function is used as the loss function to enhance the robustness to outliers. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0029] Figure 1 A schematic diagram of a process for predicting soil respiration based on GIS spatiotemporal features and deep learning, provided in an embodiment of the present application;

[0030] Figure 2 The corresponding relationship between the predicted value and the measured value of soil respiration in the training set provided in the embodiment of the present application;

[0031] Figure 3 The corresponding relationship between the predicted value and the measured value of soil respiration in the validation set provided in the embodiment of this application. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0033] like Figure 1 As shown, the embodiment of the present application provides a soil respiration prediction method based on GIS spatiotemporal features and deep learning, comprising the following steps:

[0034] Step 1: Obtain historical surface temperature data, historical soil moisture data, historical altitude data, and historical soil respiration data for the area to be predicted, and generate a multi-source fusion dataset through GIS geographic coordinate alignment.

[0035] Step 2: extracting spatial features, temporal features, and nonlinear features of the multi-source fusion data set to generate a multi-source feature data set.

[0036] Step three: Establish a hybrid network architecture based on deep learning.

[0037] Step 4: Adaptively train the hybrid network architecture using the multi-source feature dataset to obtain a soil respiration prediction model.

[0038] Step 5: Use the soil respiration prediction model to perform actual soil respiration prediction.

[0039] Illustratively, in step 1, the historical surface temperature data, the historical soil moisture data, the historical altitude data, and the historical soil respiration data are raster data corresponding to time.

[0040] Specifically, in this example, historical surface temperature, soil moisture, altitude, and soil respiration data were obtained from raster data corresponding to a specific region from 2010 to 2018, all with a resolution of 5 km. The historical soil respiration data served as the target variable (i.e., the actual value) and was compared with the predicted value ultimately output by the model.

[0041] Exemplarily, in step 2, the extraction of the spatial features includes: calculating the spherical distance from each point in the multi-source fusion dataset to the data center based on longitude and latitude.

[0042] The extraction of the time feature includes: performing sine / cosine transformation on the time variables in the multi-source fusion data set.

[0043] The extraction of the nonlinear features includes: generating interaction terms and polynomial terms for the multi-source fusion data set.

[0044] Specifically, in this embodiment, the calculation formula of the spherical distance adopts the Haversine formula, as shown below:

[0045] D=2R·arcsin(sin2(Δφ / 2)+cosφ1cosφ2sin2(Δλ / 2)).

[0046] Where D represents the spherical distance, R represents the radius of the earth, Δφ and Δλ represent the latitude difference and longitude difference, respectively, and φ1 and φ2 represent the latitude values ​​of the first and second points, respectively.

[0047] In this embodiment, the time variable is the year, and the formula for performing sine / cosine transformation on the time variable is as follows:

[0048] tsin=sin(2π(T-Tbase) / ΔT).

[0049] tcos=cos(2π(T-Tbase) / ΔT).

[0050] Where tsin represents sine transform, tcos represents cosine transform, T represents the current year, Tbase represents the base year, and ΔT represents the time span of one cycle.

[0051] In this embodiment, the interaction terms in the nonlinear feature include: historical surface temperature data×historical soil moisture data, historical surface temperature data×historical altitude data, and the polynomial terms include: the square of historical altitude data, and the square of historical surface temperature data.

[0052] Illustratively, in step 2, before calculating the spherical distance from each point in the multi-source fusion dataset to the data center based on the longitude and latitude, the longitude and latitude are normalized.

[0053] Exemplarily, in step three, the hybrid network architecture includes: a residual connection module and an attention mechanism.

[0054] The residual connection module is used to add the input features after being mapped by the fully connected layer and the skip connection.

[0055] The attention mechanism is used to calculate feature weights.

[0056] Specifically, in this embodiment, the residual connection module maps the input features through the fully connected layer and adds them to the skip connection as follows:

[0057] xout=F(xin,{Wi})+G(xin).

[0058] Among them, xin represents the input feature, F represents the residual function, Wi represents the set of learnable weight parameters in the residual function, G represents the skip connection function, which is used to adjust the dimension of the input feature xin to match the dimension of F(xin,{Wi}), and xout represents the output feature of the residual connection module.

[0059] In this embodiment, the formula for calculating feature weights by the attention mechanism is as follows:

[0060] αc=σ(W2δ(W1xc)).

[0061] Among them, αc represents the attention weight of the c-th feature dimension, σ represents the Sigmoid activation function, δ represents the ReLU activation function, xc represents the original value of the c-th dimension in the feature, W1 represents the first layer learnable weight matrix, and W2 represents the second layer learnable weight matrix.

[0062] Specifically, in this embodiment, the hybrid network architecture is provided with three residual connection modules (256-128-64 units) with a dropout rate of 0.3.

[0063] Exemplarily, in step 4, the adaptive training adopts PowerTransformer feature normalization, QuantileTransformer target distribution normalization, dynamic learning rate adjustment and stratified sampling strategy.

[0064] Specifically, in this embodiment, the stratified sampling strategy includes: dividing the target variable of the multi-source feature dataset into 5 subsets according to quantiles, retaining 20% ​​of the samples in each subset as a validation set, and the remaining 80% as a training set.

[0065] In this embodiment, the formula for dynamic learning rate adjustment is as follows:

[0066] Bt+1=Bt·(1+η·sign(Rval2-Rtrain2)).

[0067] Among them, Bt+1 represents the learning rate of the t+1th iteration cycle (next period), Bt represents the learning rate of the tth iteration cycle (current period), η represents the growth factor, the default value is 0.2, sign represents the sign function, Rval2 represents the determination coefficient of the validation set, which reflects the goodness of fit of the model in the validation set, and Rtrain2 represents the determination coefficient of the training set, which reflects the goodness of fit of the model in the training set.

[0068] Specifically, in this embodiment, the initial learning rate of the hybrid network architecture is set to 0.0003, the batch size is 512, and the early stopping threshold is 25 rounds.

[0069] Exemplarily, in step 4, during the adaptive training process of the hybrid network architecture, the loss function adopts the Huber loss function.

[0070] Specifically, the Huber loss function can enhance outlier robustness.

[0071] Exemplarily, in step 4, the adaptive training results of the hybrid network architecture are visualized and verified.

[0072] Specifically, if Figure 2The following shows the corresponding relationship between the predicted value and the measured value of soil respiration in the training set. Figure 3 The corresponding relationship between the predicted and measured soil respiration values ​​of the validation set is shown. Figure 2 and Figure 3 It can be seen that the fitting degree R of the training set 2 =0.71, RMSD

[0073] RMSE = 130.7 g·C·m -2 ·year -1 , the fitness R of the validation set 2 =0.71, RMSE = 130.8 g·C·m -2 ·year -1 . It has better model performance.

[0074] The embodiments of the present application combine multi-source data fusion, spatiotemporal feature enhancement, a hybrid network architecture, and adaptive training. Multi-source data fusion integrates multi-source data, spatiotemporal feature enhancement captures spatiotemporal characteristics, the hybrid network architecture takes into account both model expressiveness and interpretability, and adaptive training improves model robustness. Together, these factors improve the accuracy of soil respiration prediction.

[0075] The proposed hybrid network architecture includes: residual connection module and attention mechanism. The residual connection module maps the input features through the fully connected layer and then adds them to the jump connection. The feature weights are calculated through the attention mechanism, taking into account both the model's expressiveness and interpretability.

[0076] The proposed adaptive training adopts PowerTransformer feature normalization, QuantileTransformer target distribution normalization, dynamic learning rate adjustment and stratified sampling strategy to balance the model convergence speed and generalization ability and improve the model robustness.

[0077] During the adaptive training process of the proposed hybrid network architecture, the Huber loss function is used as the loss function to enhance the robustness to outliers.

[0078] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0079] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning, characterized by: The following steps are involved: Step 1: Obtain historical surface temperature data, historical soil moisture data, historical altitude data, and historical soil respiration data for the area to be predicted, and generate a multi-source fusion dataset through GIS geographic coordinate alignment; Step 2: extracting spatial features, temporal features, and nonlinear features of the multi-source fusion data set to generate a multi-source feature data set; Step 3: Establish a hybrid network architecture based on deep learning; Step 4: Adaptively train the hybrid network architecture using the multi-source feature dataset to obtain a soil respiration prediction model; Step 5: Use the soil respiration prediction model to perform actual soil respiration prediction.

2. The soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning according to claim 1 is characterized in that: In step 1, the historical surface temperature data, the historical soil moisture data, the historical altitude data, and the historical soil respiration data are raster data corresponding to time.

3. The soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning according to claim 1 is characterized in that: In step 2, the extraction of the spatial features includes: calculating the spherical distance from each point in the multi-source fusion data set to the data center based on longitude and latitude; The extraction of the time feature includes: performing sine / cosine transformation on the time variable in the multi-source fusion data set; The extraction of the nonlinear features includes: generating interaction terms and polynomial terms for the multi-source fusion data set.

4. The soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning according to claim 3 is characterized in that: In step 2, before calculating the spherical distance from each point in the multi-source fusion dataset to the data center based on the longitude and latitude, the longitude and latitude are normalized.

5. The soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning according to claim 1 is characterized in that, In step 3, the hybrid network architecture includes: a residual connection module and an attention mechanism; The residual connection module is used to add the input features after being mapped by the fully connected layer and the skip connection; The attention mechanism is used to calculate feature weights.

6. The soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning according to claim 1 is characterized in that: In step 4, the adaptive training adopts PowerTransformer feature normalization, QuantileTransformer target distribution normalization, dynamic learning rate adjustment and stratified sampling strategy.

7. The soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning according to claim 1 is characterized in that: In step 4, during the adaptive training of the hybrid network architecture, the loss function adopts the Huber loss function.

8. The soil respiration prediction method based on GIS spatiotemporal characteristics and deep learning according to claim 1 is characterized in that: In step 4, the adaptive training results of the hybrid network architecture are visualized and verified.

Images