A NPP estimation method integrating geographic mechanism and deep spatiotemporal network

Abstract

The present invention discloses a NPP estimation method integrating geographic mechanism and deep spatiotemporal network. The present invention uses spatiotemporal long short-term memory network combined with attention mechanism to fully extract the spatiotemporal characteristics of long time series remote sensing images, and uses geographic detectors to reveal and quantify the driving mechanism of NPP, guide deep learning modeling, and enhance the interpretability of deep learning models; fully explore the natural-human driving mechanism of NPP and capture the spatiotemporal characteristic information of driving factors, aiming to comprehensively reveal the complex spatiotemporal dependency between NPP and driving factors, and on this basis achieve accurate NPP estimation in large-scale complex environments.

Description

Technical Field

[0001] The present invention relates to the field of geographic detection technology, and in particular to a NPP estimation method integrating geographic mechanism and deep spatiotemporal network. Background Art

[0002] The earliest method for calculating net primary productivity (NPP) was field measurement. Its advantages are that the scheme is mature and highly operational. Its limitations are the restrictions on manpower and the number of sites. In addition, the net primary productivity obtained by field measurement at certain regional scales usually has errors and it is difficult to obtain the net primary productivity over a larger regional area.

[0003] With the development of technology, the main NPP estimation models currently include climate productivity models, physiological and ecological process models, light energy utilization models, and machine learning models. The climate productivity model is characterized by its simple principle, easy implementation, and wide use, but its disadvantage is that it does not take into account the physiological response of vegetation itself, the interaction between vegetation and complex ecosystems, and the prediction results are based on points, with low accuracy; the physiological and ecological process model has significant advantages in revealing ecological mechanisms, but its limitation is that the model construction is relatively complex and involves many parameters, which are difficult to obtain; CASA is the most widely used light energy utilization model, and the estimated data contains a large amount of spatial and temporal information, but under certain specific conditions or when facing different types of data, the CASA model may require more manual adjustment and optimization; the machine learning model can automatically extract effective features from the data, but its limitation is that the mining of spatiotemporal data is too shallow.

[0004] With the rapid progress of artificial intelligence technology, the application of machine learning models has become increasingly widespread. As an emerging force in this field, deep learning is rapidly rising and receiving increasing attention. Deep learning methods have the ability to automatically capture data features and distribution, and can deeply mine data information. They can automatically extract deep information from spatiotemporal data and establish a spatiotemporal mapping relationship between explanatory variables and response variables based on massive data itself, thereby accurately simulating NPP. In contrast, NPP estimation models may require more manual adjustments and optimizations under certain conditions or when faced with different types of data, such as the existing literature YU D, SHIP, SHAO H, etc. Modelling net primary productivity of terrestrial ecosystems in East Asia based on an improved CASA ecosystem model [J / OL]. International Journal of Remote Sensing, 2009, 30 (18): 4851-4866. DOI: 10.1080 / 01431160802680552; existing literature RAMMER W, SEIDL RA scalable model of vegetation transitions using deep neural networks [J / OL]. Methods in Ecology and Evolution, 2019, 10 (6): 879-890. DOI: 10.1111 / 2041-210X.13171. Deep learning models can be flexibly adjusted and optimized according to different data and application scenarios to adapt to various complex environmental conditions, and are flexible and adaptable in NPP estimation.

[0005] NPP is the comprehensive result of natural and human activities, and its pattern shows spatial stratification and heterogeneity. Constructing the driving mechanism of NPP in advance can provide clear geographical prior knowledge for deep learning methods to estimate NPP, enhance the geographical interpretability of deep learning and the rationality of the results. At present, as a statistical method, the purpose of geographic detectors is to explore the laws of spatial differentiation and their driving mechanisms. It has clear geographical significance and can objectively explain the weight distribution of the explanatory power of independent variables on NPP.

[0006] Therefore, it is a worthy research issue to provide an NPP estimation method that integrates geographic mechanisms and deep spatiotemporal networks to solve the problem that deep spatiotemporal networks for NPP estimation fail to effectively extract reliable information from spatiotemporal data. Summary of the invention

[0007] The purpose of the present invention is to provide an NPP estimation method that integrates geographical mechanisms and deep spatiotemporal networks. Based on geographical detectors, the contribution rate of various parameters to the NPP of a set geographical environment is studied, and the degree of explanation behind multiple variables is mined before modeling. The weight distribution of each driving factor in the overall driving mechanism of the set geographical environment NPP is determined, and clear geographical prior knowledge is provided for the use of deep learning for NPP spatiotemporal estimation. In terms of temporal features, a spatiotemporal feature encoding module based on spatiotemporal long-term and short-term networks is constructed to realize the deep feature fusion of temporal correlation and spatial correlation. In terms of spatial features, an attention mechanism is introduced to effectively extract global spatial dependencies.

[0008] The object of the present invention is achieved in that:

[0009] A NPP estimation method integrating geographic mechanism and deep spatiotemporal network includes an IGDS model, wherein the IGDS model includes NPP driving mechanism and deep spatiotemporal feature learning, quantitatively studies the NPP driving mechanism of natural factors and human activities through geographic detectors to reveal the degree of explanation of each factor on the NPP driving mechanism, quantitatively analyzes the weight distribution of each factor in the overall pattern, constructs a spatiotemporal feature encoding module based on spatiotemporal long-term and short-term networks to extract and fuse spatiotemporal features in a targeted manner, and introduces an attention mechanism to capture dynamic spatiotemporal dependencies; uses a Conv3D layer to perform spatiotemporal estimation to achieve high-precision estimation of NPP, and the specific steps are as follows:

[0010] Step 1: Detect NPP driving factors for use; the purpose of detecting NPP driving factors is to use geographic detectors to explore the contribution rate of different factors to NPP changes, build geographical prior knowledge of NPP, and use it as the weight value input of different influencing factors in subsequent deep learning models to enhance the interpretability of deep learning models;

[0011] Step 2: Fusion of spatiotemporal features based on spatiotemporal long-term and short-term networks; use the spatiotemporal long-term and short-term unit ST-LSTM to mine the temporal and spatial information of each influencing factor; use the weight values ​​of different NPP driving factors in step 1 as input, first perform weighted processing on each influencing factor, and then use the weighted result to process the spatiotemporal long-term and short-term unit ST-LSTM to obtain the final hidden state of each node The spatiotemporal long-term short-term unit ST-LSTM is used to construct a dual-state transfer mechanism. Time status and Spatiotemporal state fusion, capturing and processing complex spatiotemporal dependencies in spatiotemporal sequence data;

[0012] Step 3: Enhancement of heterogeneous spatiotemporal features based on the attention mechanism; thus, a weighted feature representation is obtained: the feature representation generated by the self-attention mechanism emphasizes the most important features in the sequence; at the same time, an updated hidden state is obtained: in the sequence model, the output of the SAU will update the hidden state of the model, carrying all the information of the sequence so far; long-term temporal dependencies are obtained through adaptive updates, and global spatial dependencies are effectively extracted through self-attention to capture the global spatiotemporal dependencies of spatiotemporal data in the NPP estimation process; first, the weight values ​​of each influencing factor obtained in step 1 are used as the input of step 2, and the spatiotemporal long-term and short-term unit ST-LSTM is processed to mine the temporal and spatial information of each image factor. The output of the spatiotemporal long-term and short-term unit ST-LSTM, that is, the final hidden state As the input of the self-attention unit SAU; SAU further analyzes these features through the self-attention mechanism, captures the long-distance dependencies between different time steps and spatial positions in the sequence, calculates the attention weights between different features, generates weighted feature representations, and then passes through the Conv3D layer to achieve high-precision estimation of NPP;

[0013] The Self-Attention Unit (SAU) for NPP estimation is used to memorize features with long-term dependencies in the spatial and temporal domains to capture global spatiotemporal dependencies in the NPP estimation process. The Self-Attention Unit receives two inputs: the input feature H of the current dependent variable and t and independent variable feature M t-1 ;

[0014] Step 4: loss function and evaluation index; use adaptive moment estimation, Adam optimizer to train the network, and use MSLE loss function to calculate the training loss between the true value and the estimated value; the accuracy verification standard uses the determination coefficient R 2 , root mean square error RMSE and mean absolute error Mean Absolute Error, MAE are used to evaluate the model estimation accuracy.

[0015] In step 1, the NPP driving factors include natural environmental factors and human disturbance factors; the impact mechanism of NPP is modeled based on the geographic detector, and the contribution rate of each factor to the NPP of the set geographical location is quantitatively analyzed; first, the difference characteristics of the continuous variables are obtained based on the natural break point method, and the single factor detection of the driving effect of the variables on the net primary productivity is analyzed based on the geographic detector; the "natural break point" category is based on the natural grouping inherent in the data; the classification intervals will be identified, similar values ​​will be grouped, and the differences between the classes can be maximized; the elements will be divided into multiple classes, and their boundaries will be set at locations where the differences in data values ​​are relatively large; the continuous variables are divided into 10 categories using the natural break point method; the single factor detector detects to what extent a certain factor X explains the spatial differentiation of attribute Y, measured by the q value;

[0016]

[0017] Where h = 1, ..., L is the stratification, i.e., classification or partition, of variable Y or factor X; N h and N are the number of units in layer h and the whole area, respectively; and σ 2 are the variances of the Y values ​​of layer h and the entire region respectively; SSW and SST are the sum of the variances within the layer and the total variance of the entire region respectively; the range of q is [0,1], the larger the value, the stronger the explanatory power of the independent variable on NPP, and vice versa; the q value means that the independent variable explains 100×q% of the net primary productivity.

[0018] In step 2, the internal structure of the spatiotemporal long-term short-term unit ST-LSTM includes an input gate, an input control gate, a forget gate and an output gate, which work together in the processing process of the IGDS model, and follow formulas (2) and (3) to ensure the effective transmission and update of the spatiotemporal state; the spatiotemporal long-term short-term unit ST-LSTM connects the time state on the channel dimension of the spatiotemporal state three-dimensional tensor That is, the standard time unit, which is passed to the spatiotemporal state in each LSTM That is, the spatiotemporal characteristics of the independent variable; the spatiotemporal long-term and short-term unit ST-LSTM unit is Another set of gate structures is constructed while retaining the standard LSTM The original gate of the node; after that, the final hidden state of the node depends on the fused spatiotemporal state. To ensure the hidden state and space-time state and Keep the dimensions consistent, apply 1×1 convolution operation and combine with nonlinear hyperbolic activation function for dimensionality reduction, and output gate o controlled by signals from two directions tTo achieve seamless spatiotemporal feature stitching and deep feature fusion of temporal correlation and spatial correlation, it can effectively simulate the spatial characteristics and temporal trajectories in the spatiotemporal sequence of vegetation parameters.

[0019]

[0020]

[0021] Through the spatiotemporal long-term and short-term network, the deep spatiotemporal features of temporal correlation and spatial correlation in long-term series images are effectively extracted from both horizontal and vertical directions, and the simultaneous modeling of NPP spatiotemporal features is realized; the node refers to a single unit or a processing unit group in the ST-LSTM network responsible for processing and fusing spatiotemporal features; each node realizes the deep fusion and processing of spatiotemporal features through its internal gate structure and 1×1 convolution operation; the spatiotemporal long-term and short-term unit ST-LSTM finally outputs the implicit state of each ST-LSTM node t in the lth layer and the spatiotemporal state of longitudinal state transfer The two serve as input information of the "self-attention unit SAU" to realize hierarchical spatiotemporal characteristic modeling of net primary productivity.

[0022] In step 3, the self-attention unit SAU includes three parts: the first part is feature aggregation, the input feature H of the current dependent variable t and independent variable feature unit M′ t Through two self-attention modules, they are mapped to Z h and Z m , the difference is that M′ t The query matrix Q used here is calculated at the current time step, and the key matrix K is calculated at the previous time step; then, by calculating the query matrix Q h With the key matrix K m The matrix multiplication of gets the similarity score e between the input and the memory m , by using the SoftMax function to obtain the weight of the aggregated features, and then by the value matrix V for all positions m Weighted summation to obtain feature Z m The "pixel" at the ith position of ; then the aggregate feature Z is passed through Z = W Z [Z h ; Z h ]; the calculation formula is as follows:

[0023]

[0024]

[0025]

[0026]

[0027] In the formula is the j-column of memory, i, j∈{1, 2, ..., N};

[0028] The second part is the spatiotemporal state update. The self-attention unit SAU uses an adaptive gating mechanism to update the spatiotemporal unit. Enable the self-attention unit SAU to capture long-term dependencies in spatial and temporal domains; aggregate the features Z and the original input H t The input gate i′ is obtained by stacking and performing one-dimensional convolution calculation t The value and fusion feature g′ t In addition, the forget gate is replaced by 1-i′ t To reduce parameters, the update process can be expressed as follows:

[0029]

[0030]

[0031]

[0032] In order to further reduce parameters and calculations, depthwise separable convolution is used to replace standard operations; compared with the original storage unit C which is updated only by convolution operation, the update of the memory unit M of the self-attention unit SAU is not only through convolution operation, but also introduces the update of the aggregated feature Z to obtain the global spatial dependency in time, so the memory unit Contains global spatiotemporal information about the past;

[0033] The third part is the output, the output gate o′ t And the feature unit after the self-attention unit SAU update Multiply element-wise to get the output of the self-attention unit SAU It can be expressed as follows:

[0034]

[0035]

[0036] Finally, the output of the “attention mechanism” module is passed to the Conv3D layer to achieve a high-precision estimation of NPP.

[0037] In step 4, the determination coefficient R 2 , root mean square error RMSE and mean absolute error Mean AbsoluteError, MAE formula is as follows:

[0038]

[0039]

[0040]

[0041]

[0042] Among them, y i and are the true value and the estimated value on the test set respectively.

[0043] The beneficial effects of the present invention are as follows: the present invention constructs an NPP estimation method that integrates geographic constraint mechanisms and deep learning models, fully explores the natural-human driving mechanism of NPP and captures the spatiotemporal characteristic information of driving factors, aiming to fully reveal the complex spatiotemporal dependency between NPP and driving factors, and on this basis, achieves accurate estimation of NPP in large-scale complex environments. The present invention uses the Qinghai-Tibet Plateau as the research area, verifies the estimation accuracy of the present invention, and forms high-precision 1km resolution monthly data of NPP in the entire Qinghai-Tibet Plateau. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a structural diagram of the IGDSNet model of the present invention;

[0045] Figure 2 This is the internal structure diagram of the spatiotemporal long-term short-term unit ST-LSTM of the present invention;

[0046] Figure 3 Module diagram of the self-attention unit SAU of the present invention;

[0047] Figure 4 These are result diagrams of different space-time scenarios of the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0049] A NPP estimation method integrating geographic mechanism and deep spatiotemporal network includes an IGDS model, wherein the IGDS model includes NPP driving mechanism and deep spatiotemporal feature learning, quantitatively studies the NPP driving mechanism of natural factors and human activities through geographic detectors to reveal the degree of explanation of each factor on the NPP driving mechanism, quantitatively analyzes the weight distribution of each factor in the overall pattern, constructs a spatiotemporal feature encoding module based on spatiotemporal long-term and short-term networks to extract and fuse spatiotemporal features in a targeted manner, and introduces an attention mechanism to capture dynamic spatiotemporal dependencies; uses a Conv3D layer to perform spatiotemporal estimation to achieve high-precision estimation of NPP, and the specific steps are as follows:

[0050] Step 1: Detect NPP driving factors for use; The purpose of detecting NPP driving factors is to use geographic detectors to explore the contribution rate of different factors to NPP changes and to construct geographic prior knowledge of NPP, which can be used as the weight value input of different influencing factors in subsequent deep learning models to enhance the interpretability of deep learning models.

[0051] In step 1, the NPP driving factors include natural environmental factors and human disturbance factors; the impact mechanism of NPP is modeled based on the geographic detector, and the contribution rate of each factor to the NPP of the set geographical location is quantitatively analyzed; first, the difference characteristics of the continuous variables are obtained based on the natural break point method, and the single factor detection of the driving effect of the variables on the net primary productivity is analyzed based on the geographic detector; the "natural break point" category is based on the natural grouping inherent in the data; the classification intervals will be identified, similar values ​​will be grouped, and the differences between the classes can be maximized; the elements will be divided into multiple classes, and their boundaries will be set at locations where the differences in data values ​​are relatively large; the continuous variables will be divided into 10 categories using the natural break point method; the single factor detector detects to what extent a certain factor X explains the spatial differentiation of attribute Y, measured by the q value;

[0052]

[0053] Where h = 1, ..., L is the stratification, i.e., classification or partition, of variable Y or factor X; N h and N are the number of units in layer h and the whole area, respectively; and σ 2 are the variances of the Y values ​​of layer h and the entire region respectively; SSW and SST are the sum of the variances within the layer and the total variance of the entire region respectively; the range of q is [0,1], the larger the value, the stronger the explanatory power of the independent variable on NPP, and vice versa; the q value means that the independent variable explains 100×q% of the net primary productivity.

[0054] Finally, the weight distribution of each factor in the overall driving mechanism of NPP was obtained: precipitation (27%), land use (24%), temperature (16%), altitude (15%), solar radiation (13%), and slope (5%), which were used as the weight input of the "space-time long-term and short-term network".

[0055] Step 2: Fusion of spatiotemporal features based on spatiotemporal long-term and short-term networks; use the spatiotemporal long-term and short-term unit ST-LSTM to mine the temporal and spatial information of each influencing factor; use the weight values ​​of different NPP driving factors in step 1 as input, first perform weighted processing on each influencing factor, and then use the weighted result to process the spatiotemporal long-term and short-term unit ST-LSTM to obtain the final hidden state of each node The spatiotemporal long-term short-term unit ST-LSTM is used to construct a dual-state transfer mechanism. Time status and Spatiotemporal state fusion,captures and processes complex spatiotemporal dependencies in spatiotemporal sequence data.

[0056] Net primary productivity is closely related to the growth environment, so accurately mining the spatiotemporal information of major factors is the key to studying the NPP estimation problem. Convolutional long short-term memory networks (ConvLSTM) have shown significant effectiveness in spatiotemporal sequence tasks. However, in traditional ConvLSTM spatiotemporal networks, there are limitations in the way its memory state is updated. The memory units between different layers are independent of each other and lack cross-layer information interaction. Therefore, when processing information, the bottom layer often cannot fully utilize the memory content accumulated by the top layer in previous time steps, which easily sacrifices time correlation and limits the performance of ConvLSTM in complex spatiotemporal sequence processing tasks.

[0057] like Figure 2 As shown in the step 2, the internal structure of the spatiotemporal long-term short-term unit ST-LSTM includes an input gate, an input control gate, a forget gate and an output gate, which work together in the processing process of the IGDS model. It follows formulas (2) and (3) to ensure the effective transmission and update of the spatiotemporal state. The spatiotemporal long-term short-term unit ST-LSTM connects the time state in the channel dimension of the spatiotemporal state three-dimensional tensor. That is, the standard time unit, which is passed to the spatiotemporal state in each LSTM unit That is, the spatiotemporal characteristics of the independent variable; the spatiotemporal long-term and short-term unit ST-LSTM is Another set of gate structures is constructed while retaining the standard LSTM The original gate of the node; after that, the final hidden state of the node depends on the fused spatiotemporal state. To ensure the hidden state and space-time state and Keeping the same dimension, we apply 1×1 convolution operation and combine it with nonlinear hyperbolic activation function to reduce the dimension. The output gate O is controlled by the signal from two directions. t To achieve seamless spatiotemporal feature stitching and deep feature fusion of temporal correlation and spatial correlation, it can effectively simulate the spatial characteristics and temporal trajectories in the spatiotemporal sequence of vegetation parameters.

[0058]

[0059]

[0060] Through the spatiotemporal long-term and short-term network, the deep spatiotemporal features of temporal correlation and spatial correlation in long-term series images are effectively extracted from both horizontal and vertical directions, and the simultaneous modeling of NPP spatiotemporal features is realized; the node refers to a single unit or a processing unit group in the ST-LSTM network responsible for processing and fusing spatiotemporal features; each node realizes the deep fusion and processing of spatiotemporal features through its internal gate structure and 1×1 convolution operation; the spatiotemporal long-term and short-term unit ST-LSTM finally outputs the implicit state of each ST-LSTM node t in the lth layer and the spatiotemporal state of longitudinal state transfer The two serve as input information of the "self-attention unit SAU" to realize hierarchical spatiotemporal characteristic modeling of net primary productivity.

[0061] Step 3: Enhancement of heterogeneous spatiotemporal features based on the attention mechanism; thus, a weighted feature representation is obtained: the feature representation generated by the self-attention mechanism emphasizes the most important features in the sequence; at the same time, an updated hidden state is obtained: in the sequence model, the output of the self-attention unit SAU will update the hidden state of the model, carrying all the information of the sequence so far; long-term temporal dependencies are obtained through adaptive updates, and global spatial dependencies are effectively extracted through self-attention to capture the global spatiotemporal dependencies of spatiotemporal data in the NPP estimation process; first, the weight values ​​of each influencing factor obtained in step 1 are used as the input of step 2, and the spatiotemporal long-term and short-term unit ST-LSTM is processed to mine the temporal and spatial information of each image factor. The output of the spatiotemporal long-term and short-term unit ST-LSTM, that is, the final hidden state As the input of the self-attention unit SAU; the self-attention unit SAU further analyzes these features through the self-attention mechanism, captures the long-distance dependencies between different time steps and spatial positions in the sequence, calculates the attention weights between different features, generates weighted feature representations, and then passes through the Conv3D layer to achieve high-precision estimation of NPP;

[0062] The Self-Attention Unit (SAU) for NPP estimation is used to memorize features with long-term dependencies in the spatial and temporal domains to capture global spatiotemporal dependencies in the NPP estimation process; Figure 3 As shown, the self-attention unit receives two inputs, the input features H of the current dependent variable t and independent variable feature M t-1 .

[0063] There is a complex dynamic spatiotemporal dependency between net primary productivity and driving factors. However, existing methods rely on stacked convolutional layers to capture local spatial dependencies. The effective receptive field is much smaller than the theoretical receptive field. When extracting features far away from a specific location, the process of feedforward and backpropagation goes through many layers, making it difficult to optimize during training. The attention mechanism can capture position dependencies in space and focus on the key information of the input data, which can effectively solve this problem.

[0064] In step 3, the self-attention unit SAU includes three parts: the first part is feature aggregation, the input feature H of the current dependent variable t and independent variable feature unit M′ t Through two self-attention modules, they are mapped to Z h and Z m , the difference is that M′ t The query matrix Q used here is calculated at the current time step, and the key matrix K is calculated at the previous time step; then, by calculating the query matrix Q h With the key matrix K m The matrix multiplication of gets the similarity score e between the input and the memory m , by using the SoftMax function to obtain the weight of the aggregated features, and then by the value matrix V for all positions m Weighted summation to obtain feature Z m The "pixel" at the ith position of ; then the aggregate feature Z is passed through Z = W Z [Z h ; Z h ]; the calculation formula is as follows:

[0065]

[0066]

[0067]

[0068]

[0069] In the formula is the j-column of memory, i, j∈{1, 2, ..., N};

[0070] The second part is the spatiotemporal state update. The self-attention unit SAU uses an adaptive gating mechanism to update the spatiotemporal unit. Enable the self-attention unit SAU to capture long-term dependencies in spatial and temporal domains; aggregate the features Z and the original input H t The input gate i′ is obtained by stacking and performing one-dimensional convolution calculation t The value and fusion feature g′ tIn addition, the forget gate is replaced by 1-i′ t To reduce parameters, the update process can be expressed as follows:

[0071]

[0072]

[0073]

[0074] In order to further reduce parameters and calculations, depthwise separable convolution is used to replace standard operations; compared with the original storage unit C which is updated only by convolution operation, the update of the memory unit M of the self-attention unit SAU is not only through convolution operation, but also introduces the update of the aggregated feature Z to obtain the global spatial dependency in time, so the memory unit Contains global spatiotemporal information about the past;

[0075] The third part is the output, the output gate o′ t And the feature unit after the self-attention unit SAU update Multiply element-wise to get the output of the self-attention unit SAU It can be expressed as follows:

[0076]

[0077]

[0078] Finally, the output of the “attention mechanism” module is passed to the Conv3D layer to achieve a high-precision estimation of NPP.

[0079] Step 4: loss function and evaluation index; use adaptive moment estimation, Adam optimizer to train the network, and use MSLE loss function to calculate the training loss between the true value and the estimated value; the accuracy verification standard uses the determination coefficient R 2 , root mean square error RMSE and mean absolute error Mean Absolute Error, MAE are used to evaluate the model estimation accuracy.

[0080] In step 4, the determination coefficient R 2 , root mean square error RMSE and mean absolute error Mean AbsoluteError, MAE formula is as follows:

[0081]

[0082]

[0083]

[0084]

[0085] Among them, y i and are the true value and the estimated value on the test set respectively.

[0086] Comparative analysis of model structures

[0087] In order to verify the superiority of the present invention, it is compared with several of the most widely used deep learning models, and all models use the same NPP dataset. The dataset contains NPP data of three different spatiotemporal scenes from 2001 to 2020, with a temporal resolution of months and a spatial resolution of 1000 meters × 1000 meters, divided into a training set (80%) and a validation set (20%). The experiment in this paper is based on the CentOS-Linux system environment. A 24-core HygonC86 processor with 128GB of memory and 4 ASPEED Graphics series GPU graphics cards with 4GB of video memory are used on the platform. On this basis, the Tensorflow (v2.6.0) and Keras (2.3.1) deep learning frameworks are used for model design and construction. The Adam optimizer is used to learn the model weights during training. The initial learning rate is set to 0.001, the batch size during training is set to 1, and the model parameters with the best effect during training are saved.

[0088] The selected models include ConvLSTM, PredRNN, and SAUConvLSTM. The output layer of each model is a Conv3D layer;

[0089] Table 1 Comparative experimental results

[0090]

[0091] To ensure a fair comparison, all estimation models are trained and tested in the same hardware and software environment. Figure 4It can be seen that: 1) The proposed IGDS spatiotemporal estimation model R2 performs well in different spatiotemporal scenarios, and R2 can be improved by 3.7%, 14.4%, and 4.6% compared with other models. MAE can be reduced by 0.0058, 0.0110, and 0.0155 compared with other models, verifying that the IGDS model can accurately estimate the spatiotemporal distribution of net primary productivity in different spatiotemporal scenarios. However, for some months (April) with large changes in NPP, the estimation error of the IGDS model is relatively large. The reason for this error is that the growth environment of vegetation in this month changes complexly, and uncertain factors lead to poor model accuracy and high estimation errors. 2) The performance of the SAUConvLSTM model is comparable to that of the IGDS model in areas with less NPP distribution, but not as good as the IGDS model in other areas; the performance of the ConvLSTM and IGDS models is comparable in areas with medium NPP distribution, and the performance of the IGDS model is far better than the other three models in areas with the most complex NPP distribution, indicating that the IGDS model is robust, and its accuracy does not decrease with changes in input data, and it can effectively estimate NPP in different spatiotemporal scenarios. 3) The accuracy and robustness of the NPP spatiotemporal modeling method that integrates geographic mechanisms and deep spatiotemporal feature learning models are verified, and a comparative experiment is conducted on the improved deep learning model network structure. The results show that the IGDS model can more effectively mine the rich spatiotemporal dependencies in remote sensing images in the study of NPP spatiotemporal estimation problems, and has geographical interpretability.

Claims

1. A NPP estimation method integrating geographic mechanism and deep spatiotemporal network, characterized by: The IGDS model includes the NPP driving mechanism and deep spatiotemporal feature learning. The geographical detector is used to quantitatively study the NPP driving mechanism of natural factors and human activities to reveal the degree of explanation of each factor on the NPP driving mechanism, and the weight distribution of each factor in the overall pattern is quantitatively analyzed. A spatiotemporal feature encoding module based on the spatiotemporal long-term and short-term network is constructed to extract and fuse spatiotemporal features in a targeted manner, and an attention mechanism is introduced to capture dynamic spatiotemporal dependencies. The Conv3D layer is used for spatiotemporal estimation to achieve high-precision estimation of NPP. The specific steps are as follows: Step 1: Detect NPP driving factors for use; the purpose of detecting NPP driving factors is to use geographic detectors to explore the contribution rate of different factors to NPP changes, build geographical prior knowledge of NPP, and use it as the weight value input of different influencing factors in subsequent deep learning models to enhance the interpretability of deep learning models; Step 2: Fusion of spatiotemporal features based on spatiotemporal long-term and short-term networks; use the spatiotemporal long-term and short-term unit ST-LSTM to mine the temporal and spatial information of each influencing factor; use the weight values ​​of different NPP driving factors in step 1 as input, first perform weighted processing on each influencing factor, and then use the weighted result to process the spatiotemporal long-term and short-term unit ST-LSTM to obtain the final hidden state of each node The spatiotemporal long-term short-term unit ST-LSTM is used to construct a dual-state transfer mechanism. Time status and Spatiotemporal state fusion, capturing and processing complex spatiotemporal dependencies in spatiotemporal sequence data; Step 3: Enhancement of heterogeneous spatiotemporal features based on the attention mechanism; thus, a weighted feature representation is obtained: the feature representation generated by the self-attention mechanism emphasizes the most important features in the sequence; at the same time, an updated hidden state is obtained: in the sequence model, the output of the SAU will update the hidden state of the model, carrying all the information of the sequence so far; long-term temporal dependencies are obtained through adaptive updates, and global spatial dependencies are effectively extracted through self-attention to capture the global spatiotemporal dependencies of spatiotemporal data in the NPP estimation process; first, the weight values ​​of each influencing factor obtained in step 1 are used as the input of step 2, and the spatiotemporal long-term and short-term unit ST-LSTM is processed to mine the temporal and spatial information of each image factor. The output of the spatiotemporal long-term and short-term unit ST-LSTM, that is, the final hidden state As the input of the self-attention unit SAU; the self-attention unit SAU further analyzes these features through the self-attention mechanism, captures the long-distance dependencies between different time steps and spatial positions in the sequence, calculates the attention weights between different features, generates weighted feature representations, and then passes through the Conv3D layer to achieve high-precision estimation of NPP; The Self-Attention Unit (SAU) for NPP estimation is used to memorize features with long-term dependencies in the spatial and temporal domains to capture global spatiotemporal dependencies in the NPP estimation process. The Self-Attention Unit receives two inputs: the input feature H of the current dependent variable and t and independent variable feature M t-1 ; Step 4: loss function and evaluation index; use adaptive moment estimation, Adam optimizer to train the network, and use MSLE loss function to calculate the training loss between the true value and the estimated value; The accuracy verification standard uses the determination coefficient R 2 , root mean square error RMSE and mean absolute error Mean Absolute Error, MAE are used to evaluate the model estimation accuracy.

2. The NPP estimation method integrating geographic mechanism and deep spatiotemporal network according to claim 1 is characterized by: In step 1, NPP driving factors include natural environmental factors and human disturbance factors; Based on the geographic detector modeling of the impact mechanism of NPP, the contribution rate of each factor to the NPP of the set geographical location is quantitatively analyzed; first, the difference characteristics of the continuous variables are obtained based on the natural break point method, and the single factor detection of the driving effect of the variables on the net primary productivity is analyzed based on the geographic detector; the "natural break point" category is based on the natural grouping inherent in the data; the classification intervals will be identified, similar values ​​will be grouped, and the differences between each class can be maximized; the features will be divided into multiple classes, and their boundaries will be set at locations where the differences in data values ​​are relatively large; The continuous variables were divided into 10 categories using the natural break method; the single factor detector detected the extent to which a factor X explained the spatial differentiation of attribute Y, measured by the q value; Where h = 1, ..., L is the stratification, i.e., classification or partition, of variable Y or factor X; N h and N are the number of units in layer h and the whole area, respectively; and σ 2 are the variances of the Y values ​​of layer h and the entire region respectively; SSW and SST are the sum of the variances within the layer and the total variance of the entire region respectively; the value range of q is [0, 1], the larger the value, the stronger the explanatory power of the independent variable on NPP, and vice versa; the q value means that the independent variable explains 100×q% of the net primary productivity.

3. The NPP estimation method integrating geographic mechanism and deep spatiotemporal network according to claim 1 is characterized by: In step 2, the internal structure of the spatiotemporal long-term short-term unit ST-LSTM includes an input gate, an input control gate, a forget gate and an output gate, which work together in the processing process of the IGDS model, and follow formulas (2) and (3) to ensure the effective transmission and update of the spatiotemporal state; the spatiotemporal long-term short-term unit ST-LSTM connects the time state on the channel dimension of the spatiotemporal state three-dimensional tensor That is, the standard time unit, which is passed to the spatiotemporal state in each LSTM unit That is, the spatiotemporal characteristics of the independent variable; the spatiotemporal long-term and short-term unit ST-LSTM is Another set of gate structures is constructed while retaining the standard LSTM The original gate of the node; after that, the final hidden state of the node depends on the fused spatiotemporal state. To ensure the hidden state and space-time state and Keeping the same dimension, we apply 1×1 convolution operation and combine it with nonlinear hyperbolic activation function to reduce the dimension. The output gate O is controlled by the signal from two directions. t To achieve seamless spatiotemporal feature stitching, deep feature fusion of temporal correlation and spatial correlation, and effectively simulate the spatial characteristics and temporal trajectories in the spatiotemporal sequence of vegetation parameters Through the spatiotemporal long-term and short-term network, the deep spatiotemporal features of temporal correlation and spatial correlation in long-term series images are effectively extracted from both horizontal and vertical directions, and the simultaneous modeling of NPP spatiotemporal features is realized; the node refers to a single unit or a processing unit group in the ST-LSTM network responsible for processing and fusing spatiotemporal features; each node realizes the deep fusion and processing of spatiotemporal features through its internal gate structure and 1×1 convolution operation; the spatiotemporal long-term and short-term unit ST-LSTM finally outputs the implicit state of each ST-LSTM node t in the lth layer and the spatiotemporal state of longitudinal state transfer The two serve as input information of the "self-attention unit SAU" to realize hierarchical spatiotemporal characteristic modeling of net primary productivity.

4. The NPP estimation method integrating geographic mechanism and deep spatiotemporal network according to claim 1, characterized in that: In step 3, the self-attention unit SAU includes three parts: the first part is feature aggregation, the input feature H of the current dependent variable t and independent variable feature unit M′ t Through two self-attention modules, they are mapped to Z h and Z m , the difference is that M′ t The query matrix Q used here is calculated at the current time step, and the key matrix K is calculated at the previous time step; then, by calculating the query matrix Q h With the key matrix K m The matrix multiplication of gets the similarity score e between the input and the memory m ., by using the SoftMax function to obtain the weight of the aggregated features, and then by the value matrix V of all positions m Weighted summation to obtain feature Z m The "pixel" at the ith position of ; then the aggregate feature Z is passed through Z = W Z [Z h ; Z h ]; the calculation formula is as follows: In the formula is the j-column of memory, i, j∈{1, 2, ..., N}; The second part is the spatiotemporal state update. The self-attention unit SAU uses an adaptive gating mechanism to update the spatiotemporal unit. Enable the self-attention unit (SAU) to capture long-term dependencies in both spatial and temporal domains; Aggregate features Z and original input H t The input gate i′ is obtained by stacking and performing one-dimensional convolution calculation t The value and fusion feature g′ t In addition, the forget gate is replaced by 1-i′ t To reduce parameters, the update process can be expressed as follows: In order to further reduce parameters and calculations, depthwise separable convolution is used to replace standard operations; compared with the original storage unit C which is updated only by convolution operation, the update of the memory unit M of the self-attention unit SAU is not only through convolution operation, but also introduces the update of the aggregated feature Z to obtain the global spatial dependency in time, so the memory unit Contains global spatiotemporal information about the past; The third part is the output, the output gate o′ t And the feature unit after the self-attention unit SAU update Multiply element-wise to get the output of the self-attention unit SAU The statement is as follows: Finally, the output of the "attention mechanism" module is passed to the Conv3D layer to achieve a high-precision estimate of NPP.

5. The NPP estimation method integrating geographic mechanism and deep spatiotemporal network according to claim 1, characterized in that: In step 4, the determination coefficient R 2 , RMSE and Mean Absolute Error, the MAE formula is as follows: Among them, y′i and are the true value and the estimated value on the test set respectively.

Images