NDVI time series data reconstruction method and system

Abstract

This invention relates to the field of agricultural monitoring and discloses a method and system for reconstructing NDVI time-series data. The method includes: collecting Sentinel-1 SAR data, Sentinel-2 optical data, and initial NDVI time-series data of a target area; performing time node matching on the data to obtain an initial dataset; inputting the initial dataset into a trained Transformer-based time-series data reconstruction network, performing deep coupling between the Sentinel-1 SAR data and the Sentinel-2 optical data, and mapping the coupling result to the initial NDVI time-series data to reconstruct the NDVI time-series data. This invention enhances the learning between SAR and optical data, constructs the relationship between complete SAR data and missing optical data, and maps the contextual information of the coupling result to the initial NDVI time-series data to reconstruct the NDVI time-series data, filling the gaps in optical observation data and obtaining high-precision and complete NDVI time-series data.

Description

Technical Field

[0001] This invention relates to the field of agricultural monitoring, and more specifically, to a method and system for reconstructing NDVI time-series data. Background Technology

[0002] NDVI (Normalized Difference Vegetation Index) time-series satellite data is widely used in crop monitoring to obtain timely and accurate crop information, facilitating crop management, food production, and food security. NDVI is generally calculated based on optical remote sensing data. However, optical remote sensing data is affected by rapid crop growth and cloudy / rainy weather, resulting in a large portion of the data being invalid. Therefore, how to avoid the invalidity of most optical remote sensing data due to rapid crop growth and cloudy / rainy weather, and thus obtain complete and accurate NDVI time-series data, remains a problem that urgently needs to be solved by those skilled in the art.

[0003] An NDVI time-series data compensation and reconstruction method is available. It determines the NDVI time-series data and radar features of the area under study by acquiring remote sensing image data of the area under study, and determines the radar features that best match the NDVI time-series data by selecting the radar features from the radar features according to preset rules as the target radar features. Finally, the NDVI time-series data and the target radar features are fused using a preset fitting regression algorithm to obtain NDVI time-series data with fused radar features.

[0004] However, the above method for reconstructing NDVI time-series data using remote sensing image data does not fully utilize the joint complementarity between missing optical data and complete SAR data, and therefore cannot obtain complete NDVI time-series data. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies that cannot obtain complete NDVI time series data, this invention provides an NDVI time series data reconstruction method and system.

[0006] To solve the above-mentioned technical problems, the present invention provides the following:

[0007] In the first aspect, this invention proposes an NDVI time-series data reconstruction method, comprising:

[0008] Collect Sentinel-1 SAR data, Sentinel-2 optical data, and initial NDVI time series data for the target area.

[0009] Time nodes were matched with Sentinel-1 SAR data, Sentinel-2 optical data and initial NDVI time series data of the target area to obtain an initial dataset, which was then divided into a training set, a validation set and a test set.

[0010] The training set and validation set are used to train and tune the time series data reconstruction network based on the Transformer architecture until the loss function of the network converges, thus obtaining the trained time series data reconstruction network.

[0011] The test set is input into the trained time series data to reconstruct the network. The Sentinel-1 SAR data and Sentinel-2 optical data are deeply coupled, and the coupling result is mapped onto the initial NDVI time series data to reconstruct the NDVI time series data.

[0012] Secondly, this invention also proposes an NDVI time-series data reconstruction system, comprising:

[0013] The collection module is used to collect Sentinel-1 SAR data, Sentinel-2 optical data, and initial NDVI time-series data of the target area.

[0014] The matching module is used to perform time node matching on the Sentinel-1 SAR data, Sentinel-2 optical data and initial NDVI time series data of the target area to obtain an initial dataset, and divide the initial dataset into a training set, a validation set and a test set.

[0015] The training module is used to train and tune the time-series data reconstruction network based on the Transformer architecture using the training set and validation set, respectively, until the loss function of the network converges, thus obtaining the trained time-series data reconstruction network.

[0016] The reconstruction module is used to input the test set into the trained time-series data to reconstruct the network, perform deep coupling on the Sentinel-1 SAR data and Sentinel-2 optical data, and map the coupling result onto the initial NDVI time-series data to reconstruct the NDVI time-series data.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention enhances the learning between SAR and optical data by deeply coupling complete SAR data and missing optical data through the Transformer model, constructs the relationship between complete SAR data and missing optical data, and maps the context information of the coupling result to the initial NDVI time series data to reconstruct the NDVI time series data, filling the gaps in optical observation data, and obtaining high-precision and complete NDVI time series data. Attached Figure Description

[0018] Figure 1This is a flowchart of the NDVI time-series data reconstruction method according to an embodiment of this application.

[0019] Figure 2 This is an architecture diagram of a time-series data reconstruction network based on the Transformer architecture, as described in an embodiment of this application.

[0020] Figure 3 This is a schematic diagram illustrating the principle of random cloud masking in an embodiment of this application.

[0021] Figure 4 This is a schematic diagram of farmland and crops in the study area of ​​this application embodiment.

[0022] Figure 5 This is a schematic diagram illustrating the partitioning of the initial dataset in an embodiment of this application.

[0023] Figure 6 Visualizations of NDVI time-series data reconstruction using different methods.

[0024] Figure 7 This is an architecture diagram of the NDVI time-series data reconstruction system according to an embodiment of this application. Detailed Implementation

[0025] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0026] The technical invention of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Example 1

[0028] Please see Figure 1 This embodiment proposes an NDVI time series data reconstruction method, including:

[0029] Collect Sentinel-1 SAR data, Sentinel-2 optical data, and initial NDVI time-series data for the target area;

[0030] Time nodes are matched with Sentinel-1 SAR data, Sentinel-2 optical data and initial NDVI time series data of the target area to obtain an initial dataset, and the initial dataset is divided into training set, validation set and test set.

[0031] The training set and validation set are used to train and tune the time series data reconstruction network based on the Transformer architecture until the loss function of the network converges, thus obtaining the trained time series data reconstruction network.

[0032] The test set is input into the trained time series data to reconstruct the network. The Sentinel-1 SAR data and Sentinel-2 optical data are deeply coupled, and the coupling result is mapped onto the initial NDVI time series data to reconstruct the NDVI time series data.

[0033] The NDVI time-series data reconstruction method proposed in this embodiment constructs the relationship between complete SAR data and missing optical data by deeply coupling complete SAR data and missing optical data through a Transformer model, and maps the context information of the coupling result to the initial NDVI time-series data to reconstruct the NDVI time-series data, filling the gaps in optical observation data and obtaining high-precision and complete NDVI time-series data.

[0034] Example 2

[0035] This embodiment is an improvement on the NDVI time-series data reconstruction method proposed in Embodiment 1.

[0036] like Figure 2 As shown, in this embodiment, the time-series data reconstruction network based on the Transformer architecture includes a first linear layer, a second linear layer, a third linear layer, a decoder, and an encoder. The Sentinel-1 SAR data and Sentinel-2 optical data are input into the first linear layer for position encoding to obtain a SAR optical image time series with position information. The image time series is transmitted to the encoder to encode the context information of the image time series, obtaining SAR optical encoding including time context information. The initial NDVI time series data is input into the second linear layer for position encoding to obtain NDVI time series data with position information. The NDVI time series data with position information and the SAR optical encoding are input into the decoder, which maps the context information of the SAR optical encoding to the NDVI time series data with position information. The encoder outputs the mapping result to the third linear layer, and the third linear layer outputs the reconstructed NDVI time series data.

[0037] In this embodiment, the feedforward neural network includes a one-dimensional convolutional module, a linear transformation module, and a ReLU activation function layer connected in sequence.

[0038] In this embodiment, the first linear layer encodes the Sentinel-1 SAR data and Sentinel-2 optical data at the same location, as shown in the following expression:

[0039] ;

[0040]

[0041] in, For the input of the first i Sentinel-1 SAR data, For the input of the first i Sentinel-2 optical data, Encodes a predefined location.

[0042] In the specific implementation process, firstly, using the first linear layer and the same location coding, the input Sentinel-1 SAR data and Sentinel-2 optical data are converted into information embeddings with location information. A multivariate time series of length t contains all the features extracted from the SAR images, represented as... and .

[0043] In this embodiment, the initial dataset is subjected to random cloud masking to obtain an initial dataset that isolates local optical information.

[0044] In this embodiment, the encoder comprises N stacks of identical layers, where N is a positive integer. Each stack layer includes a first multi-head self-attention module, a feedforward neural network, a first normalization layer, a first MLP layer, and a second normalization layer connected in sequence. The Sentinel-1 SAR data is input to the encoder as the VV and VH bands of the SAR image metadata time series to learn the temporal characteristics of the SAR data. Each time point in the SAR image data corresponds to a token in natural language processing. The two input bands provide two input dimensions, which are converted into hidden dimensions in a linear layer. The decoder comprises a second multi-head self-attention module, a third normalization layer, a fourth normalization layer, a second MLP layer, and a fifth normalization layer connected in sequence. The output of the second normalization layer is connected to the input of the third multi-head self-attention module. The encoder is used to encode the contextual information embedded in the Sentinel-1 SAR data and Sentinel-2 optical data.

[0045] like Figure 3 As shown, the input data to the decoder is NDVI temporal data with artificially randomized cloud masks. Similar to the encoder, data with additional positional encoding information first enters the MHSA block of the cloud mask. During natural language processing training, the decoder's self-attention block needs to be applied together with subsequent masks to cover the positions after the block to be predicted, ensuring that the prediction depends only on previously known outputs. However, NDVI temporal data does not strictly follow the semantic sequenceness of natural language, meaning that the prediction of the data can depend on valid data from previous and subsequent time points. This invention proposes an artificially randomized cloud masking method that can isolate local optical information, forcing the model to reconstruct under-cloud information by learning the contextual information of optical data and the temporal correlation of SAR. When calculating self-attention, the decoder obtains the artificially set randomized cloud mask positions and removes the information of the cloud mask points when calculating the self-attention score. This ensures that during the prediction process, only partial NDVI information is needed to reconstruct the complete NDVI sequence.

[0046] In this embodiment, the first multi-head self-attention module, the second multi-head self-attention module, and the third multi-head self-attention module are self-attention-based multi-head self-attention modules, and the mathematical expression of the self-attention-based multi-head self-attention module is as follows:

[0047]

[0048]

[0049] , ,

[0050] in, For the first i A person's attention, H The number of self-attention heads. For the output of the self-attention head, This is a time series of SAR optical images with location information. , and The first i The query matrix, key matrix, and value matrix of each self-attention head, where D is the dimension of the query matrix, key matrix, and value matrix. The weights of the first linear layer, The weights of the second linear layer, The weights of the third linear layer, T Indicates the period. for , , The dimension can prevent The function returns a very small gradient, causing the gradient to vanish.

[0051] In this embodiment, The results are input into the feedforward network. And in Blocks and Residual skip connections and layer normalization are used after each layer.

[0052] Multi-head attention module The output is the result of the action of each self-attention point. A linear mapping graph. Self-attention is achieved through querying the matrix. Bond matrix AND-value matrix The proportional dot product is calculated by multiplying the products. This is obtained by applying different fully connected... The sequence elements of the layer are processed to obtain each self-attention head. , and Matrix

[0053] In this embodiment, the output of the encoder is expressed as follows:

[0054]

[0055] where represents normalization processing, is the output of the feedforward neural network, is the output of the multi-head self-attention module. The output of the nth layer (n < N) is used as the input of the (n + 1)th layer, and at the same time, the output of the Nth layer is used as the input of the encoder multi-head self-attention module.

[0056] For training samples, the loss function of the network is defined as the weighted sum of the mean absolute percentage error SMAPE is expressed as

[0057]

[0058] Although the model only needs to predict the data at the missing time points, in order to make the model better fit the overall curve, the invention uses weighted as the total loss function.

[0059] In this embodiment, the expression of the weighted loss function is as follows:

[0060]

[0061] where is the weight for predicting the missing data masked by the cloud mask,<000​​​​​​​​​​​​​​​​​​​​​​​​​The NDVI time series data reconstruction method proposed in this application utilizes Sentinel-1 SAR data, which can penetrate clouds and is not affected by fog, providing crucial supplementary information below the clouds. Moreover, the time series information of SAR data is significantly correlated with crop phenological rhythms. Although the data from Sentinel-2 optical observations is incomplete, its existing data and spatiotemporal constraints are of great significance for constructing NDVI data. By deeply coupling complete SAR data and partially missing optical data, the advantages of both types of data can be utilized to reconstruct dense NDVI time series.

[0063] We utilize a time-series data reconstruction network based on the Transformer architecture, construct a fusion mechanism for heterogeneous data through three different attention heads, use the attention of the SAR data branch to learn the temporal changes of SAR data itself, use the attention of the optical branch to learn the rules of optical data self-filling, and use fusion attention to learn the complementarity and conversion mechanism between SAR and missing optics.

[0064] A random cloud mask structure is introduced into the time series data reconstruction network based on the Transformer architecture, which improves the reconstruction performance of missing optical data time series. The use of random masks enhances the ability of optical data to learn its own data characteristics, and a scheme similar to self-supervised learning enables the model to model the contextual information of the optical data itself, fully exploring the patterns in the optical data time series.

[0065] Example 3

[0066] To further illustrate the technical effects achieved by the present invention, this embodiment, based on the method proposed in Embodiment 2, takes a farmland research area in southeastern Missouri, USA as an example to further explain the technical solution of the present invention.

[0067] like Figure 4 As shown, southeastern Missouri, USA, has a large area of ​​plains farmland, with arable land accounting for about 90%. The crops are diverse, with soybeans accounting for the largest share at 39.22%, followed by cotton and corn, accounting for 20% and 18.44% respectively. The agricultural season in this region from January to December 2021 is the timeline for this study.

[0068] In this embodiment, 25 Sentinel-1 SAR images and 26 Sentinel-2 images were collected, so that the resulting dataset can cover the entire study area and the proportion of invalid pixels is less than 10%.

[0069] In this embodiment, the initial NDVI time series data of the cloudless area is selected based on the optical image data of the study area.

[0070] In this embodiment, based on the spatial continuity of cloud coverage, the initial labeled dataset is divided into a training set, a validation set, and a test set. These three sets are non-boundary regions with an area ratio of 3:1:1, as shown in the figure. Figure 5 As shown in the diagram. For the test set, the initial NDVI time-series data were artificially set to NA (null value), with proportions of 20%, 30%, 40%, and 50% respectively, corresponding to data gaps caused by cloud cover in actual situations. These are the data that need to be predicted, thus constructing four test datasets. Simultaneously, the time periods with the largest fluctuations in greenness values ​​for several common crops were set to NA, which are the data that need to be predicted, thus constructing a dataset related to the key growth stages of crops.

[0071] In this embodiment, the steps are... The parameter settings for the time series data reconstruction network based on the Transformer architecture are as follows: the input data size before the encoder is 20000×25×2, and the size after linear projection is 20000×25×128. Considering both accuracy and computational resources, both the encoder and decoder contain two layers. In the encoder, the input size of the multi-head self-attention module is set to 128. In the MLP layer, the input data is first converted to 512, and then returned to 128 dimensions through two one-dimensional convolutional layers, meaning the output dimension of each encoder layer is 128. In the decoder, it has the same two-layer structure and parameters as the encoder. The input size of the decoder is 20000×25×128. For the input data, in each time series, a portion of the data is set to NA to simulate partial time series data loss caused by clouding during the prediction stage. Before being fed into the decoder, the input dimension of the data is converted to 128 dimensions. The input data size of the masked multi-head self-attention module and the subsequent multi-head attention module are both 128-dimensional and have 3 heads. The mask multi-head self-attention module receives and , , The same input; while the subsequent multi-head attention module receives the decoder input. and encoder output , Regarding the optimizer, this embodiment selects the Adam optimizer and sets the model learning rate to 0.001. For model convergence settings, this example sets 100 convergence parameters. The model was run on a GeForce RTX 3090 GPU, and the convergence time was approximately 1.5 hours.

[0072] Table 1 Different models in Evaluation of prediction results for each test set

[0073]

[0074] This embodiment uses the trained model to predict test set data and compares the test results of this application with MCNN-Seq, SenRVM, Random Forest (RF) Regression, and Linear Interpolation, based on evaluation metrics. and The prediction performance of different models on the four test sets is evaluated as shown in Table 1.

[0075] As shown in Table 1, the results indicate that this application... Less than 0.09 Greater than 0.90, achieving the best performance compared to other methods. Compared to MCNN-Seq, SenRVM, Random Forest (RF) Regression, and Linear Interpolation, RM-Trans... The values ​​increased by approximately 38%, 39%, 45%, and 50%, respectively. They increased by 16%, 17%, 25%, and 35% respectively.

[0076] To further explore the performance of the temporal data reconstruction network proposed in this application and the coupling effect of SAR, optical, and NDVI inputs, two ablation experiments were conducted in this embodiment. The first ablation experiment aimed to verify the effectiveness of the random mask proposed in this invention and the serialized mask used in the temporal data reconstruction network, both compared with the maskless decoder MHSA. The second ablation experiment discussed the settings of optical data including five different cloud proportions (CP) and whether SAR data was used.

[0077] Table 2 Comparison Results of Mask Ablation Experiments

[0078]

[0079] The ablation results for the mask structures are shown in Table 2, with bold values ​​representing the highest evaluation metrics. MHSA with a serialized mask (MHSA + Subsequent) performs the worst, indicating that this mask structure is unsuitable for NDVI sequence processing. MHSA without a mask achieves better results than MHSA + Subsequent, suggesting that the model can learn to ignore the information provided by the mask tokens. MHSA with a random mask (MHSA + random) is the best structure, indicating that the random mask directly prevents the model from learning information from the masked locations, thus contributing to better representation learning.

[0080] Table 3 Comparison results of experiments with different input combinations

[0081]

[0082] After discussing the performance of the masking, this embodiment also discusses different input combination strategies. As shown in Table 3, among all optical and SAR data inputs, a 50% cloud ratio is the best masking strategy during training. Comparing different modalities, using optical data alone yields better results than using SAR data alone, indicating the contribution of information from the same modality (optical). However, since the SAR+optical input strategy (“50%CP_Opt+SAR”) outperforms 50%CP_Opt, SAR data still provides useful supplementary information when some NDVI data is lost.

[0083] like Figure 6 As shown, the visualization results of the above methods are displayed. It can be seen that the results of the three deep learning methods (i.e., MCNN-Seq, SenRVM, and this application) have similar predictions in terms of plot color and boundaries, far outperforming the predictions of RF and linear methods. The RF method fails to recover scene colors, especially in July and August, possibly because the NDVI differences of plots are smaller at this time. However, when some plots begin to be harvested, the spatial differences in NDVI are larger, and its performance is better. For the linear method, the predictions in July and August are less accurate than in September. This is mainly because the input to the linear interpolation function is before or after the growth phase, when NDVI is relatively small. Therefore, the linear prediction is much smaller than the GT, and the prediction is more accurate when GT_NDVI decreases in September. Regarding RMSE and R... 2 In this application, the best reconstruction results are achieved across all dates. Furthermore, noise in the mappings obtained in this application is suppressed compared to the suboptimal MCNN-Seq method. This indicates that the optical data used in this application helps mitigate the noise problem in SAR data.

[0084] Example 4

[0085] This embodiment proposes an NDVI time-series data reconstruction system, including:

[0086] The collection module is used to collect Sentinel-1 SAR data, Sentinel-2 optical data, and initial NDVI time-series data of the target area.

[0087] The matching module is used to perform time node matching on the Sentinel-1 SAR data, Sentinel-2 optical data and initial NDVI time series data of the target area to obtain an initial dataset, and divide the initial dataset into a training set, a validation set and a test set.

[0088] The training module is used to train and tune the time-series data reconstruction network based on the Transformer architecture using the training set and validation set, respectively, until the loss function of the network converges, thus obtaining the trained time-series data reconstruction network.

[0089] The reconstruction module is used to input the test set into the trained time-series data to reconstruct the network, perform deep coupling on the Sentinel-1 SAR data and Sentinel-2 optical data, and map the coupling result onto the initial NDVI time-series data to reconstruct the NDVI time-series data.

[0090] The NDVI time-series data reconstruction system proposed in this application constructs the relationship between complete SAR data and missing optical data by deeply coupling complete SAR data and missing optical data through a Transformer model, and maps the context information of the coupling result to the initial NDVI time-series data to reconstruct the NDVI time-series data, filling the gaps in optical observation data and obtaining high-precision and complete NDVI time-series data.

[0091] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0092] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for reconstructing NDVI time-series data, characterized in that, include: Collect Sentinel-1 SAR data, Sentinel-2 optical data, and initial NDVI time-series data for the target area; Time nodes were matched with Sentinel-1 SAR data, Sentinel-2 optical data and initial NDVI time series data of the target area to obtain an initial dataset, and the initial dataset was divided into training set, validation set and test set. The training set and validation set are used to train and tune the time series data reconstruction network based on the Transformer architecture until the loss function of the network converges, thus obtaining the trained time series data reconstruction network; the time series data reconstruction network based on the Transformer architecture includes a first linear layer, a second linear layer, a third linear layer, a decoder, and an encoder. The Sentinel 1 SAR data and Sentinel 2 optical data are input into the first linear layer for location-coordinated encoding to obtain a SAR optical image time series with location information, the expression of which is as follows: ； in, For the input of the first i Sentinel-1 SAR data, For the input of the first i Sentinel-2 optical data, Encode the predefined location; The image time series is transmitted to the encoder to encode the context information of the image time series, thereby obtaining SAR optical coding that includes time context information; The initial NDVI timing data is input into the second linear layer for position encoding to obtain NDVI timing data with position information. The NDVI time-series data with location information and the SAR optical code are input into the decoder, and the context information of the SAR optical code is mapped into the NDVI time-series data with location information. The encoder outputs the mapping result to the third linear layer, and the third linear layer outputs the reconstructed NDVI time-series data. The test set is input into the trained time series data to reconstruct the network. The Sentinel-1 SAR data and Sentinel-2 optical data are deeply coupled, and the coupling result is mapped onto the initial NDVI time series data to reconstruct the NDVI time series data.

2. The NDVI time-series data reconstruction method according to claim 1, characterized in that, After obtaining the initial dataset, but before dividing it into training, validation, and test sets, the method further includes: performing random cloud masking on the initial dataset to obtain an initial dataset that isolates local optical information.

3. The NDVI time-series data reconstruction method according to claim 1, characterized in that, The encoder includes a one-dimensional convolution module, a linear transformation module, and a ReLU activation function layer connected in sequence.

4. The NDVI time-series data reconstruction method according to claim 3, characterized in that, The encoder comprises N stacks of identical layers, where N is a positive integer. Each stack layer comprises a first multi-head self-attention module, a feedforward neural network, a first normalization layer, a first MLP layer, and a second normalization layer connected in sequence. The decoder comprises a second multi-head self-attention module, a third normalization layer, a third multi-head self-attention module, a fourth normalization layer, a second MLP layer, and a fifth normalization layer connected in sequence. The output of the second normalization layer is connected to the input of the third multi-head self-attention module.

5. The NDVI time-series data reconstruction method according to claim 4, characterized in that, The first, second, and third multi-head self-attention modules are self-attention-based multi-head self-attention modules, and their mathematical expressions are as follows: , , in, For the first i A person's attention, H The number of self-attention heads. For the output of the self-attention head, This is a time series of SAR optical images with location information. , and The first i The query matrix, key matrix, and value matrix of each self-attention head, where D is the dimension of the query matrix, key matrix, and value matrix. For the first i One attention head is used to generate the weight matrix of the query vector. For the first i One attention head is used to generate the weight matrix of the key vector. For the first i One attention head is used to generate the weight matrix of the value vector. T Indicates the period.

6. The NDVI time-series data reconstruction method according to claim 5, characterized in that, The output of the encoder The expression is as follows: in, This indicates normalization processing. This is the output of the feedforward neural network. This is the output of the multi-head self-attention module.

7. The NDVI time-series data reconstruction method according to claim 6, characterized in that, The expression for the loss function is as follows: in, Weights for predicting missing data masked by cloud masking. For timing position, For the sample i The network prediction value, For the sample i The true value, As a marker of validity, ,when When it is indicated that it is a cloud mask, This indicates that it is a non-cloud mask.

8. An NDVI time-series data reconstruction system, applied to the DVI time-series data reconstruction method as described in any one of claims 1 to 7, characterized in that, include: The collection module is used to collect Sentinel-1 SAR data, Sentinel-2 optical data, and initial NDVI time-series data of the target area; The matching module is used to perform time node matching on the Sentinel-1 SAR data, Sentinel-2 optical data and initial NDVI time series data of the target area to obtain an initial dataset, and divide the initial dataset into a training set, a validation set and a test set. The training module is used to train and tune the time series data reconstruction network based on the Transformer architecture using the training set and the validation set, respectively, until the loss function of the network converges, and thus obtain the trained time series data reconstruction network. The reconstruction module is used to input the test set into the trained time-series data to reconstruct the network, perform deep coupling of Sentinel-1 SAR data and Sentinel-2 optical data, and map the coupling result to the initial NDVI time-series data to reconstruct the NDVI time-series data.

Images