NDVI prediction method based on space-time attention neural network model

Abstract

The invention discloses an NDVI (normalized difference vegetation index) prediction method based on a space-time attention neural network model. The method comprises the following steps: preprocessing a remote sensing image of a research area to obtain month-by-month NDVI data; according to the NDVI data set, a model STA-CNN-BiLSTM is constructed based on an attention mechanism and a bidirectional convolution long short-term memory network; based on the NDVI image, constructing a sample data set, training the STA-CNN-BiLSTM model, and obtaining an NDVI prediction model; and the trained model is used for NDVI prediction, and prediction performance is evaluated according to related evaluation indexes. And a neural network model based on time and space and attention is constructed, so that the noise influence of factors such as time and space on NDVI prediction is reduced, and the prediction precision is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of NDVI prediction and relates to an NDVI prediction method based on a spatiotemporal attention neural network model. Background Art

[0002] The Normalized Difference Vegetation Index (NDVI) is not only an important indicator for detecting vegetation growth status and vegetation coverage, but also an important parameter for carbon fixation analysis. Therefore, analyzing and predicting the distribution characteristics of NDVI is of great significance for applications such as carbon fixation analysis and natural disaster prediction. Although the NDVI data as a whole shows periodic changes, there is the disappearance and growth of vegetation in many regions due to factors such as time and space, which results in a large amount of noise in model training and leads to low prediction accuracy.

[0003] NDVI time series data are widely used in monitoring natural grassland vegetation coverage, landscape changes, land use, land classification and other fields, but the prediction research on NDVI time series is very limited. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an NDVI prediction method based on a spatiotemporal attention neural network model, to construct a spatiotemporal attention-based neural network model, to reduce the noise impact of factors such as time and space on NDVI prediction, and to improve prediction accuracy.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a NDVI prediction method based on a spatiotemporal attention neural network model, comprising the following steps: Step 1: Preprocess the remote sensing images of the study area to obtain monthly NDVI data; Step 2: Based on the NDVI dataset, a model STA-CNN-BiLSTM is constructed based on the attention mechanism and bidirectional convolutional long short-term memory network; Step 3: Based on the NDVI image, a sample data set is constructed, and the STA-CNN-BiLSTM model is trained to obtain the NDVI prediction model; Step 4: Use the trained model for NDVI prediction and evaluate the prediction performance based on relevant evaluation indicators.

[0006] In step 1, the formula for calculating NDVI is: , where NIR and R are near infrared band and red band respectively.

[0007] In step 2, a model STA-CNN-BiLSTM is constructed based on the attention mechanism and bidirectional convolutional long short-term memory network, including: A bidirectional convolutional long-term and short-term model network is constructed, remote sensing image data is deeply processed to obtain spatiotemporal information, the spatiotemporal attention module is integrated, and the weights of different months and different spaces are adjusted to improve the prediction accuracy.

[0008] In step 2, the bidirectional convolutional long-term short-term model network includes: At time t, ConvLSTM has two outputs C t and H t , that is, the output at the current moment and the state of the current unit; the gate control unit consists of three gates: input gate, forget gate and output gate; the state update process of the ConvLSTM hidden layer is: ; ; ; ; ; ; in, and Represent convolution and Hadamard product operations respectively; i t It is the input gate, which uses the sigmoid function to select the new input data and the output of the previous neuron to determine the input information of the neuron; f t It is the forget gate, and its input parameters are the same as those of the input gate. It forgets the output of the previous neuron in proportion, that is, it controls how much information of the previous neuron is retained to enter this neuron; o t is the output gate, which uses the sigmoid function to determine the proportional output unit state; X t for t The input of the neuron at time h t-1 for t -1 moment neuron output; W i , W f , W c , W o They are X in different convolution operations respectively. t The weight of R i , R f , R c , R o h t-1 The weight of i , f , c , C t-1 The weight of i , b f , b c , b o is the bias value of the convolution operation; C t for t The cell state of the neuron at that moment is a combination of the cell state of the previous neuron and t The new data input at any time forms a new cell state; finally, the state of the hidden layer h t , obtained through the output gate and the tanh function.

[0009] In step 2, the fusion spatiotemporal attention module includes: Step 2.1, continuously compress the spatial dimension of the input image sequence, that is, the picture matrix, to form a one-dimensional time data corresponding to the picture sequence, and strengthen the feature extraction of different time periods; the purpose is to convert the time information in the data into the corresponding time attention matrix and store it; this module compresses the spatial dimension of the output Z of the bidirectional ConvLSTM module, leaving the time dimension feature; then maps the convolution layer with the attention layer, trains the weights to obtain the time attention weight matrix T, and performs Hadamard product with Z to output TA. ; In step 2.2, the output TA of the temporal attention module is used as input, and the data is reduced in dimension through maximum pooling and average pooling. Then, the convolutional layer and the attention layer are used to map the data, and the weights are trained to obtain the spatial attention weight matrix SA, which is then multiplied by the Hadamard product with the input to obtain the spatiotemporal attention weight matrix STA.

[0010] In step 3, the specific process of training the STA-CNN-BiLSTM model is as follows: Step 3.1, dynamically adjust hyperparameters according to task characteristics and complexity to achieve task balance and optimize performance; Step 3.2, use the Sigmoid-Sigmoid combination method as the activation function; In step 3.3, use the Adam optimizer with adaptive learning rate as the gradient descent algorithm for training the model, minimize the loss function, and iteratively update the model.

[0011] In step 4, the original data is given spatiotemporal weight information, and the data with spatiotemporal weight information is input into the STA-CNN-BiLSTM model. The model outputs the predicted NDVI value for a period of time in the future. Five indicators are used to measure the performance of the prediction model: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), mean absolute percentage error (MAPE), determination coefficient (R 2 ); where R2 Measures the correlation between the predicted value and the actual value and the goodness of fit of the model; while MAE, MSE, RMSE and MAPE reflect the difference between the actual value and the predicted value. The calculation formula is as follows: ; ; ; ; ; ; in is the data with spatiotemporal attention weight information, is the Hadamard product, is the original NDVI data, n is the total number of input features, is the actual value, is the predicted value, is the average of the actual values.

[0012] The main beneficial effects of the present invention are: By gradually compressing the data of spatial and temporal dimensions to strengthen the feature extraction of temporal and spatial dimensions, the spatiotemporal attention weights are formed, and combined with the bidirectional convolutional long short-term memory network, the model STA-CNN-BiLSTM is established to improve the accuracy of the NDVI time series prediction method at the pixel scale.

[0013] The network model is comprehensively evaluated based on multiple indicators, and the network model is fine-tuned and improved according to the evaluation results; according to the NDVI prediction results, the data is integrated to construct the spatiotemporal distribution map of NDVI changes, which provides important information for environmental monitoring and decision-making analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0015] Figure 1 Flow chart of the present invention.

[0016] Figure 2 This is a diagram of the ConvLSTM cell unit structure of the present invention.

[0017] Figure 3 This is the framework diagram of the STA-CNN-BiLSTM model of the present invention.

[0018] Figure 4 This is a diagram of the dynamic calculation process of the temporal attention weight of the present invention.

[0019] Figure 5 This is a diagram of the dynamic calculation process of the spatial attention weight of the present invention.

[0020] Figure 6 This is a comparison chart of evaluation indicators predicted by different models of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 to Figure 6 In this paper, a NDVI prediction method based on a spatiotemporal attention neural network model includes the following steps: Step 1: Preprocess the remote sensing images of the study area to obtain monthly NDVI data; Step 2: Based on the NDVI dataset, a model STA-CNN-BiLSTM is constructed based on the attention mechanism and bidirectional convolutional long short-term memory network; Step 3: Based on the NDVI image, a sample data set is constructed, and the STA-CNN-BiLSTM model is trained to obtain the NDVI prediction model; Step 4: Use the trained model for NDVI prediction and evaluate the prediction performance based on relevant evaluation indicators.

[0022] Embodiment 1, The remote sensing images of the study area were preprocessed to obtain monthly NDVI data. The specific process is as follows: The monthly MOD13Q1 surface reflectance product data are used. These data provide a spatial resolution of 250 meters in a 16-day gridded level 3 product, as well as a 12-layer scientific data set, applying radiometric calibration and atmospheric correction, and synthesizing monthly NDVI data by combining the 16-day data.

[0023] NDVI is calculated using time series MODIS image acquisition, and the calculation formula is: ; Among them, NIR and R are the near infrared band 0.84–0.87µm and red band 0.62–0.67µm of the Modis sensor, respectively.

[0024] Embodiment 2, According to the NDVI dataset, a model STA-CNN-BiLSTM is constructed based on the attention mechanism and bidirectional convolutional long short-term memory network; Construct a CNN-BiLSTM network. At time t, ConvLSTM has two outputs C t and H t , that is, the output at the current moment and the state of the current unit. The gate control unit consists of three gates: input gate, forget gate and output gate. Figure 2 As shown in the figure, the state update process of the ConvLSTM hidden layer is: ; ; ; ; ; ; in, and Represent convolution and Hadamard product operations respectively.

[0025] i t It is the input gate, which uses the sigmoid function to select the new input data and the output of the previous neuron to determine the input information of the neuron; f t It is the forget gate, and its input parameters are the same as those of the input gate. It forgets the output of the previous neuron in proportion, that is, it controls how much information of the previous neuron is retained to enter this neuron; o t It is the output gate, which uses the sigmoid function to determine the proportional output unit state.

[0026] X t for t The input of the neuron at time h t-1 for t -1 moment neuron output; W i , W f , W c , W o They are X in different convolution operations respectively. t The weight of R i , R f , R c , R o h t-1 The weight of i , f , c , C t-1 The weight of i , b f , b c , b o is the bias value for the convolution operation.

[0027] C t for t The cell state of the neuron at that moment is a combination of the cell state of the previous neuron and tNew data is input at all times, forming a new cell state.

[0028] Finally, the state of the hidden layer is h t , obtained through the output gate and the tanh function.

[0029] Embodiment 3, Design a temporal attention module such as Figure 4 As shown in Figure 1, the dynamic calculation process of the specific time series attention weight is as follows: Suppose that after passing through the bidirectional mapping module, the two-dimensional tensor data X of the tth image n×m .X n×m First, according to the average pooling process, X n×m Each row is calculated by average pooling to obtain one-dimensional data X' n Next, X'n is pooled again to obtain the temporal information T representing the tth image t , and finally perform Hadamard product with Z to output TA. The calculation formula is: ; .

[0030] Embodiment 4, like Figure 5 As shown in Figure 2, the dynamic calculation process of the specific spatial sequence attention weight is as follows: Assume that after passing through the generalization module 1, the time series data of each pixel of the image is obtained through the maximum pooling operation, representing the first ij The spatial information of pixels SA ij , and finally perform Hadamard product with the input to output STA. The calculation formula is: ; ; Among them, STA is the spatiotemporal attention weight matrix, SA is the spatial attention weight matrix, TA is the temporal attention weight matrix, It is the Hadamard product, which means the multiplication of elements in corresponding positions.

[0031] Embodiment 5, Based on NDVI images, a sample data set was constructed, and the STA-CNN-BiLSTM model was trained to obtain the NDVI prediction model. The normalized NDVI data is calculated with the spatiotemporal attention weight matrix to give the NDVI data spatiotemporal information. The calculation formula is: ,in, is the data with spatiotemporal attention weight information, is the Hadamard product, and X is the original NDVI data.

[0032] Use the Sigmoid-Sigmoid combination method as the activation function. Use the Adam optimizer with adaptive learning rate as the gradient descent algorithm for training the model, minimize the loss function, and iteratively update the model. Determine the optimal parameters.

[0033] Then, the data with spatiotemporal weight information is input into the long short-term memory network for training.

[0034] Embodiment 6, The trained model was used for NDVI prediction, and the prediction performance of this method was evaluated based on relevant evaluation indicators.

[0035] Prediction is performed using CNN, CNN-LSTM, CNN-BILSTM, and STA-CNN-BiLSTM.

[0036] The data with spatiotemporal weight information are input into the model, and the model outputs the predicted NDVI value for a period of time in the future. Five indicators are used to measure the performance of the prediction model: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), mean absolute percentage error (MAPE), determination coefficient (R 2 ). Where R 2 Measures the correlation between the predicted value and the actual value and the goodness of fit of the model. MAE, MSE, RMSE and MAPE reflect the difference between the actual value and the predicted value. The calculation formula is as follows: ; ; ; ; ; in is the data with spatiotemporal attention weight information, is the Hadamard product, is the original NDVI data, n is the total number of input features, is the actual value, is the predicted value, is the average of the actual values.

[0037] like Figure 6 In the figure, there is a comparison of indicators of various models.

[0038] The implementation results show that spatial and temporal characteristics play an important role in NDVI time series prediction. By comparing the evaluation indicators, adding the spatiotemporal attention module can effectively improve the prediction effect of the model.

[0039] The above-mentioned embodiments are only preferred technical solutions of the present invention and should not be regarded as limitations of the present invention. The embodiments and features in the embodiments of the present application can be arbitrarily combined with each other without conflict. The protection scope of the present invention shall be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A NDVI prediction method based on a spatiotemporal attention neural network model, characterized in that: The steps include: Step 1: Preprocess the remote sensing images of the study area to obtain monthly NDVI data; Step 2: Based on the NDVI dataset, a model STA-CNN-BiLSTM is constructed based on the attention mechanism and bidirectional convolutional long short-term memory network; Step 3: Based on the NDVI image, a sample data set is constructed, and the STA-CNN-BiLSTM model is trained to obtain the NDVI prediction model; Step 4: Use the trained model for NDVI prediction and evaluate the prediction performance based on relevant evaluation indicators.

2. The NDVI prediction method based on the spatiotemporal attention neural network model according to claim 1 is characterized in that: In step 1, the formula for calculating NDVI is: , where NIR and R are near infrared band and red band respectively.

3. The NDVI prediction method based on the spatiotemporal attention neural network model according to claim 1 is characterized in that: In step 2, a model STA-CNN-BiLSTM is constructed based on the attention mechanism and bidirectional convolutional long short-term memory network, including: A bidirectional convolutional long-term and short-term model network is constructed, remote sensing image data is deeply processed to obtain spatiotemporal information, the spatiotemporal attention module is integrated, and the weights of different months and different spaces are adjusted to improve the prediction accuracy.

4. The NDVI prediction method based on the spatiotemporal attention neural network model according to claim 3 is characterized in that: In step 2, the bidirectional convolutional long-term and short-term model network includes: At time t, ConvLSTM has two outputs C t and H t , that is, the output at the current moment and the state of the current unit; the gate control unit consists of three gates: input gate, forget gate and output gate; the state update process of the ConvLSTM hidden layer is: ； ； ； ； ； ； in, and Represent convolution and Hadamard product operations respectively; i t It is the input gate, which uses the sigmoid function to select the new input data and the output of the previous neuron to determine the input information of the neuron; f t It is the forget gate, and its input parameters are the same as those of the input gate. It forgets the output of the previous neuron in proportion, that is, it controls how much information of the previous neuron is retained to enter this neuron; o t is the output gate, which uses the sigmoid function to determine the proportional output unit state; X t for t The input of the neuron at time h t-1 for t -1 moment neuron output; W i , W f , W c , W o They are X in different convolution operations respectively. t The weight of R i , R f , R c , R o h t-1 The weight of i , f , c , C t-1 The weight of i , b f , b c , b o is the bias value of the convolution operation; C t for t The cell state of the neuron at that moment is a combination of the cell state of the previous neuron and t The new data input at any time forms a new cell state; finally, the state of the hidden layer h t , obtained through the output gate and the tanh function.

5. The NDVI prediction method based on the spatiotemporal attention neural network model according to claim 3 is characterized in that: In step 2, the fusion spatiotemporal attention module includes: Step 2.1, continuously compress the spatial dimension of the input image sequence, that is, the picture matrix, to form a one-dimensional time data corresponding to the picture sequence, and strengthen the feature extraction of different time periods; the purpose is to convert the time information in the data into the corresponding time attention matrix and store it; this module compresses the spatial dimension of the output Z of the bidirectional ConvLSTM module, leaving the time dimension feature; then maps it through the convolution layer and the attention layer, trains the weights to obtain the time attention weight matrix T, and performs Hadamard product with Z to output TA. In step 2.2, the output TA of the temporal attention module is used as input, and the data is reduced in dimension through maximum pooling and average pooling. Then, the convolutional layer and the attention layer are used to map the data, and the weights are trained to obtain the spatial attention weight matrix SA, which is then multiplied by the Hadamard product with the input to obtain the spatiotemporal attention weight matrix STA.

6. The NDVI prediction method based on the spatiotemporal attention neural network model according to claim 1 is characterized in that: In step 3, the specific process of training the STA-CNN-BiLSTM model is as follows: Step 3.1, dynamically adjust hyperparameters according to task characteristics and complexity to achieve task balance and optimize performance; Step 3.2, use the Sigmoid-Sigmoid combination method as the activation function; In step 3.3, use the Adam optimizer with adaptive learning rate as the gradient descent algorithm for training the model, minimize the loss function, and iteratively update the model.

7. The NDVI prediction method based on the spatiotemporal attention neural network model according to claim 1 is characterized in that: In step 4, the original data is given spatiotemporal weight information, and the data with spatiotemporal weight information is input into the STA-CNN-BiLSTM model. The model outputs the predicted NDVI value for a period of time in the future. Five indicators are used to measure the performance of the prediction model: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), mean absolute percentage error (MAPE), determination coefficient (R 2 ); where R 2 Measures the correlation between the predicted value and the actual value and the goodness of fit of the model; while MAE, MSE, RMSE and MAPE reflect the difference between the actual value and the predicted value. The calculation formula is as follows: ； ； ； ； ； ； in is the data with spatiotemporal attention weight information, is the Hadamard product, is the original NDVI data, n is the total number of input features, is the actual value, is the predicted value, is the average of the actual values.

Images