Self-supervised learning and variation combined day-by-day NDVI reconstruction method and system

Abstract

The invention discloses a day-by-day NDVI reconstruction method combining self-supervised learning and variation, and the method comprises the steps: firstly constructing a global sample library based on a self-supervised learning strategy, effectively extracting spatial-temporal features through a global-local double-flow parallel time sequence reconstruction network, and producing seamless multi-day synthetic data, and the local branch depicts a local space-time law by adopting a convolution long-short time memory module. Therefore, a day-by-day NDVI pre-filling method with intra-year time adjacent domain priori and inter-year period similar priori as guidance is developed. For gross errors possibly introduced by factors such as surface coverage change and image quality marking in the pre-filling process, an adjacent window screening method is designed to remove noisy points. And finally, obtaining a high-quality seamless day-by-day NDVI product by adopting one-dimensional variation filtering. According to the invention, the precision and efficiency advantages of the deep learning method and the robustness advantage of the variational method are integrated, and the bottleneck problem that the conventional method is difficult to handle day-by-day NDVI large-range continuous missing is broken through.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image processing and computer vision technology, and in particular relates to a time series reconstruction framework for remote sensing missing information, and specifically relates to a daily NDVI reconstruction method and system combining self-supervised learning with variational methods. Background Art

[0002] Optical remote sensing satellite imaging is affected by internal factors such as sensor hardware limitations and external factors such as poor atmospheric observation conditions. This makes it extremely challenging to obtain high-quality daily Normalized Differnce Vegetation Index (NDVI) data over long time series. Time series reconstruction methods, which build mathematical models for local temporal neighborhoods or the entire time series, effectively extract vegetation growth patterns from the temporal and spatial variations of the preceding and subsequent time series, achieving the desired reconstruction goal. These methods are widely used in the reconstruction of multi-year NDVI series.

[0003] Currently, time series reconstruction methods can be categorized into three main categories: temporal filtering-based methods, variational reconstruction-based methods, and deep learning-based methods. Temporal filtering-based methods perform interpolation, fitting, and filtering operations in the temporal or spatiotemporal dimensions to summarize the temporal variation patterns of NDVI and establish corresponding mathematical models. However, these methods require a high number of good observations and are difficult to reconstruct daily data with very few valid observations. Furthermore, the introduction of numerous artificial parameters increases the subjectivity and uncertainty of the reconstruction task. Variational reconstruction-based methods treat the reconstruction process as an ill-posed inverse problem, constructing energy functionals and numerically solving them to fill in missing values. While these methods can handle more complex missing values, they still struggle to directly handle daily data with extremely high missing rates, and their high computational cost limits their application in practical data production. Deep learning-based methods, with their powerful nonlinear feature extraction and parallel computing capabilities, have seen initial application in NDVI time series reconstruction due to their high accuracy and efficiency. However, the effectiveness of deep learning models depends on the quantity and quality of training samples, and the inconsistency in constructing degraded-clean sample pairs has prevented their full application in remote sensing time series reconstruction. In addition, the huge amount of daily data and the extremely high missing rate further increase the difficulty of applying deep learning methods in daily data reconstruction.

[0004] In summary, although a large number of NDVI time series reconstruction methods have been proposed, the extremely frequent information loss of daily data makes it difficult for each type of method to directly reconstruct it. Summary of the Invention

[0005] To address the problem that daily data is frequently missing information, making it difficult for each type of method to directly reconstruct it, the present invention provides a daily NDVI reconstruction method that combines self-supervised learning with variational methods. By developing a self-supervised deep learning model with good regional generalization performance, the self-supervised information is mined from large-scale raw data to drive network training, achieving high-precision reconstruction of multi-day synthetic data. Furthermore, a robust one-dimensional variational framework for NDVI time series reconstruction is constructed for daily data, achieving effective step-by-step reconstruction from multi-day synthesis to daily fine-grained reconstruction.

[0006] According to one aspect of the present invention, a daily NDVI reconstruction method combining self-supervised learning and variational methods is provided, comprising: We sample and filter NDVI data tiles globally, classify and randomly match the resulting data blocks based on their missingness rates, and construct a degraded-clean sample library for self-supervised network training. Based on a self-supervised learning strategy, a global-local dual-stream parallel NDVI spatiotemporal reconstruction network is developed. A loss function that incorporates global-local data consistency, multi-dimensional gradient priors, and temporal smoothness priors is designed to obtain seamless multi-day synthetic NDVI data. The obtained seamless multi-day composite NDVI data is used as reference information to pre-fill the original daily data, and outliers are removed from the pre-filled data. One-dimensional variational filtering is used for further denoising and smoothing, and the reconstructed daily seamless NDVI sequence is output.

[0007] As a further technical solution, we build a degraded-clean sample library for self-supervised network training, which also includes: Divide and crop NDVI data tiles globally, and select areas with a variety of landform types and climate types as candidate sample areas; Based on the data quality marking layer, data blocks with a completeness rate greater than a first threshold and data blocks with a missing rate greater than a second threshold are screened out according to the proportion of missing data; Randomly match the two types of data blocks, assign missing values ​​of data blocks with missing rates greater than the second threshold to data blocks with integrity rates greater than the first threshold at the corresponding positions, construct degraded data blocks, and perform quality marking. The constructed degraded data blocks and their quality labels are used as the input of the NDVI spatiotemporal reconstruction network, and the original data blocks with a completeness rate greater than the first threshold are used as label data for self-supervised learning to construct degraded-clean sample pairs.

[0008] As a further technical solution, a global-local dual-stream parallel NDVI spatiotemporal reconstruction network is developed, including: Construct a feature extraction module for extracting features from the input degraded data blocks and their quality marks; Construct a global-local temporal processing module, which includes a parallel framework of local and global temporal feature extraction branches. The local temporal branch processes temporal dependencies moment by moment, while the global temporal branch models long-term temporal relationships through multi-head self-attention. The two parallel branches simultaneously capture local temporal details and global temporal trends. Construct a dense connection module, including multiple layers of dense connections, where each layer receives the feature maps of all previous layers as input and passes its own output to all subsequent layers. Finally, by compressing the number of channels, a seamless image sequence is obtained after reconstruction. Construct loss functions, including global-local data consistency loss, multi-dimensional gradient loss, and temporal smoothing loss.

[0009] As a further technical solution, the local timing branch processes sequence information moment by moment through a convolutional long short-term memory module and establishes local timing dependencies to capture neighborhood timing information.

[0010] As a further technical solution, the global timing branch adopts the multi-head self-attention mechanism in Transformer to capture the key information of the entire timing and directly model the global timing relationship.

[0011] As a further technical solution, the original daily data is pre-filled, and outliers are removed from the pre-filled data. One-dimensional variational filtering is used for further denoising and smoothing, including: Missing data are pre-filled using interannual cycle priors from multi-year data and intra-annual temporal neighbor domain priors from synthetic data. For each pixel of the pre-filled image, the gross errors of the pre-filled data are screened and eliminated by setting the mean and standard deviation within the range and time neighborhood; One-dimensional variational filtering is used to constrain the local smoothness prior and interannual similarity prior of the NDVI time series through regularization to further remove residual noise.

[0012] According to one aspect of the present invention, a daily NDVI reconstruction system combining self-supervised learning with variational analysis is provided, wherein the system is used to implement the daily NDVI reconstruction method combining self-supervised learning with variational analysis.

[0013] As a further technical solution, the system includes: The first main module is used to sample and filter NDVI data tiles globally, classify and randomly match the obtained data blocks according to their missing rates, and construct a degraded-clean sample library for self-supervised network training; The second main module is used to develop a global-local dual-stream parallel NDVI spatiotemporal reconstruction network based on a self-supervised learning strategy, design a loss function that includes global-local data consistency, multi-dimensional gradient priors, and temporal smoothness priors to obtain seamless multi-day synthetic NDVI data; The third main module is used to use the obtained seamless multi-day synthetic NDVI data as reference information to pre-fill the original daily data, remove outliers from the pre-filled data, use one-dimensional variational filtering for further denoising and smoothing, and output the reconstructed daily seamless NDVI sequence.

[0014] According to one aspect of the present invention, a daily NDVI reconstruction device combining self-supervised learning and variational analysis is provided, comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the daily NDVI reconstruction method combining self-supervised learning and variational analysis.

[0015] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the daily NDVI reconstruction method combining self-supervised learning and variational analysis.

[0016] The proposed method combines deep learning with variational methods, fully taking into account the temporal prior characteristics of NDVI, and effectively reconstructing daily NDVI data in a step-by-step manner. Compared with existing technologies, the advantages of this invention are mainly reflected in: (1) This invention selects diverse sample candidate areas worldwide, constructs sample pairs using real data, and drives self-supervised network training. This solves the problem that the existing technology is limited by factors such as the contradiction in constructing degraded-clean sample pairs, and the high-precision and high-parallel computing capabilities of deep learning technology have not been fully explored in the field of NDVI time series reconstruction.

[0017] (2) In the self-supervised spatiotemporal reconstruction network of synthetic data, the present invention combines the Transformer and ConvLSTM modules to develop a local-global dual-stream parallel time series processing module, which can fully extract the spatiotemporal characteristics of the NDVI image sequence and achieve high-precision reconstruction of the synthetic data.

[0018] (3) The present invention adopts a step-by-step reconstruction framework that combines self-supervised learning with one-dimensional variational methods, fully combining the respective advantages of the two reconstruction algorithms. In the time dimension, from coarse to fine, the stability of NDVI data filling is improved through multi-day synthesis to daily fine step-by-step reconstruction, and the intra-annual neighborhood prior and inter-annual cycle prior of the NDVI time series are effectively mined, breaking through the bottleneck problem that a single method is difficult to handle large-scale continuous missing daily data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction will be given below to the drawings used in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 This is a reconstruction flowchart provided by an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the visualization effect of the MODIS MOD09GQ product according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the daily time series change curves of representative land features such as mixed forests and crops based on the MODIS MOD09GQ product in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the visualization effect of the MODIS MOD09GQ product according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments or single embodiments provided by the present invention are arbitrarily combined with each other to form a new technical solution. This combination is not restricted by the sequence of steps and / or structural composition mode, but must be based on the ability of ordinary technicians in this field to implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that this combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0025] To address the problem that extremely frequent cloud cover causes a large number of long-term continuous missing data in daily NDVI data, and existing single reconstruction methods are difficult to effectively reconstruct them, the present invention proposes a daily NDVI reconstruction method that combines self-supervised learning and variational learning. First, high-quality multi-day seamless synthetic data is generated based on large-scale original data to assist in reconstruction. Then, the daily NDVI is effectively pre-filled using the intra-annual time-adjacent domain prior and the inter-annual cycle similarity prior. Then, a one-dimensional variational reconstruction method is used to further denoise and smooth it, ultimately obtaining high-quality seamless daily NDVI data.

[0026] This paper combines deep learning with variational methods to give full play to the advantages of both methods. It reconstructs multi-day synthetic data through a deep learning algorithm, pre-fills it with potential intra-annual neighborhood priors and inter-annual cycle priors, and further smoothes it with robust one-dimensional variational filtering, thus obtaining high-quality daily NDVI data that is seamless in time and space.

[0027] Please see Figure 1 The present invention provides a daily NDVI reconstruction method combining self-supervised learning and variational methods, comprising the following steps: Step 1: Divide and crop NDVI data tiles globally, classify and randomly match the resulting data blocks according to their missing rates, and construct a degraded-clean sample library for self-supervised network training.

[0028] Specifically include: Step 1.1: Divide and crop the NDVI data tiles globally, and select areas with a variety of landform types and climate types as candidate sample areas.

[0029] Step 1.2: Based on the data quality marker layer (i.e., the layer inherent in the data that reflects the quality of each vegetation index pixel), select high-quality data blocks and data blocks with high cloud coverage according to the proportion of missing data. It should be noted that the determination of high-quality data blocks and data blocks with high cloud coverage can be based on different thresholds depending on the region, and this is not limited by the present invention. As a preferred embodiment, data with a completeness rate greater than 70% can be selected as high-quality data blocks, and data with a missing rate greater than 30% can be selected as data blocks with high cloud coverage.

[0030] Step 1.3: Randomly match the two types of data blocks, assign the missing values ​​of the data blocks with more cloud coverage to the data blocks with better quality at the corresponding positions, construct degraded data blocks, and perform corresponding processing on the quality marks.

[0031] Step 1.4: Use the constructed degraded data block and its quality mark as the input of step 2, and the original good quality data block as the label data of step 2 to construct the degraded-clean sample pair.

[0032] Step 2: Based on the self-supervised learning strategy, a global-local dual-stream parallel NDVI spatiotemporal reconstruction network is developed, and a loss function that includes global-local data consistency, multidimensional gradient prior, and temporal smoothness prior is designed to obtain seamless multi-day synthetic NDVI data.

[0033] The reconstruction network takes degraded data blocks and corresponding quality labels as input. It first encodes the original data at different receptive fields using a multi-scale information extraction module to obtain multi-scale features. The extracted feature information is then fed into the local spatiotemporal processing branch and the global temporal processing branch, respectively. The local temporal processing branch uses a Convolutional Long Short-Term Memory (ConvLSTM) module to process sequence information moment by moment and establish local temporal dependencies, capturing neighborhood temporal information. The global temporal processing branch utilizes the multi-head self-attention mechanism within the Transformer to capture key information across the entire temporal sequence and directly model global temporal relationships. The output features of both branches are then processed through a densely connected module to enhance feature reuse and significantly improve feature representation capabilities. Finally, the network achieves reconstruction results through feature dimensionality reduction. Losses are calculated between the reconstruction results and the labeled data in cloud-free areas and areas of uncertain pixel quality, including global-local data consistency loss, multidimensional gradient loss, and temporal smoothing loss.

[0034] Specifically include: Step 2.1: Build a feature extraction module. Input the degraded NDVI data block constructed in step 1.4 and its quality mark: N ∈ℝ W×H×T and F ∈ℝ W×H×T ,in T is the time length of the time series image, H and W is the height and width of the image. The feature extraction of each phase is performed on these two data respectively:

[0035] in N and To degrade NDVI data and extract image features, F and is the corresponding quality mark and the extracted mark feature, w is the convolution kernel, b represents the bias, is the convolution operation, RRepresents the ReLU activation function.

[0036] Step 2.2: Build a global-local temporal processing module. This includes a parallel framework for local and global temporal feature extraction branches. The local branch processes temporal dependencies moment by moment, while the global branch models long-term temporal relationships through multi-head self-attention. This dual-branch parallel architecture captures both local temporal details and global temporal trends.

[0037] Step 2.2.1: Local temporal feature extraction branch. In this branch, the extracted image features and marker features are first element-wise multiplied and then input into the ConvLSTM unit to effectively extract local spatiotemporal features. Finally, a convolution operation is used to adjust the time dimension to a specific length. This process can be expressed as:

[0038] in Represents the ConvLSTM time series processing unit, is the convolution kernel, is the local timing feature extracted from this branch.

[0039] Step 2.2.2: Global temporal feature extraction branch. In this branch, first N and Perform dimension transformation and change Batchsize, H and W Unify to the same dimension, retaining the time dimension T and feature dimension C , the features after dimension transformation ∈ℝ (B×H×W)×T×C and ∈ℝ (B×H×W)×T×C Perform element-wise multiplication and input into the Transformer unit to extract global timing information through the multi-head self-attention mechanism Finally, the local and global information of the two branches are concatenated and convolved to adjust the time dimension to a specific length, and the result of the global-local timing processing module is output. The process can be expressed as (4) and (5), where Concat represents the concatenation operation along the time dimension:

[0040] Step 2.3: Construct a dense connection module. Through multi-layer dense connections, the reuse of temporal features and gradient flow are significantly improved. Each layer receives the feature maps of all previous layers as input and passes its own output to all subsequent layers. Finally, the number of channels is compressed by 1×1 convolution to obtain the reconstructed seamless image sequence. The process can be expressed as:

[0041] in H l For the l densely connected blocks, including batch normalization layers, ReLU activation functions, and convolutional layers, N l ( l =1,…, L ) represents the l Dense features of the layer, is the convolution kernel of the last layer.

[0042] Step 2.4: Construct the loss function.

[0043] Step 2.4.1: Global-local data consistency loss. The local mask loss and global loss between the reconstruction result and the labeled data are calculated using the 1-norm, which can be expressed as:

[0044] in B Indicates the batch size. M 1 represents the quality of the label data as a well-observed position mask, i.e., the overall mask; M 2 represents the position mask where the input data is missing data and the label data is a good observation, that is, the local mask of the missing area.

[0045] Step 2.4.2: Multidimensional gradient loss. Including H 、 W 、 T The gradient loss in three directions can be expressed as:

[0046] in Indicates the image edge H 、 W 、 T Dimensional gradient operation.

[0047] Step 2.4.3: Temporal smoothing loss. Considering the gradual change of NDVI in the temporal dimension, that is, vegetation growth generally follows a gradual change law, temporal smoothing loss is introduced as a priori constraint of the sequence, which can be expressed as:

[0048] in Indicates the t Reconstructed image at a moment in time.

[0049] Step 2.4.4: Perform weighted summation of the three loss functions to construct the overall loss function. and is the balance coefficient:

[0050] Step 3: Use the seamless multi-day composite NDVI data obtained in step 2 as reference information to pre-fill the original daily data, remove outliers from the pre-filled data, and use one-dimensional variational filtering for further denoising and smoothing. The final seamless daily data output is the required high-quality NDVI sequence.

[0051] Specifically include: Step 3.1: Effectively pre-fill daily data. Pre-fill missing data as much as possible using the inter-annual cycle prior of multi-year data and the intra-annual neighborhood prior of synthetic data to increase the number of valid observations.

[0052] Step 3.1.1: Extraction of intra-year neighboring domain priors. Traverse each daily NDVI image. For the image acquired on the day of year (DOY), search for a neighboring seamless composite NDVI image generated by step 2 within the multi-day window as the intra-year neighboring domain prior. ; Step 3.1.2: Extraction of interannual cycle prior information. For a certain observation date DOY, traverse the images of all data of year Y on that date and take the median of the valid observation values ​​over many years as the interannual cycle prior information. , which can be specifically expressed as:

[0053] in I Y is the image of year Y on date DOY, F Y is the corresponding effective observation mask.

[0054] Step 3.1.3: Daily data pre-filling. For the image on the observation date DOY, the pre-filling result is obtained by weighting the inter-annual cycle prior and the intra-annual neighborhood prior. I DOY Considering that missing data of NDVI in cloud-covered areas are generally negatively skewed, the larger value of the two prior information is given a greater weight, specifically:

[0055] Step 3.2: Remove abnormal noise from the pre-filled data. Uncertainties caused by factors such as errors in the quality mark of the data itself, changes in vegetation phenology within a multi-day time window, and changes in ground features between years will lead to some obvious gross errors in the pre-filled data. For each pixel of the pre-filled image, p , through a certain range of time neighborhood Mean within µ and standard deviation σ Screen and remove gross errors. For pre-filled pixels on the observation date DOY ,like or , then the denoising result for:

[0056] Step 3.3: Use a one-dimensional variational filter to further smooth the pre-filled and gross error-removed NDVI image series. Although the daily data obtained in step 3.2 has undergone preliminary gross error removal, a certain range of noise still exists. This paper uses a robust one-dimensional variational filter to constrain the local smoothness prior and interannual similarity prior of the NDVI time series through regularization to further remove residual noise. The matrix-vector form of this model is:

[0057] In the formula x and y are the ideal clean sequence and the original missing sequence, W is the weight matrix, D is the time series second-order difference matrix, L is the inter-annual first-order difference matrix, and is the regularization parameter. The final sequence obtained through this variational model is the seamless daily NDVI sequence.

[0058] The visualization effect of the embodiment of the present invention based on MODIS MOD09GQ product is as follows Figure 2 and 4 shown. Figure 2 In the figure, the left side is the image of the 128th day of 2011, and the right side is the image of the 250th day of 2011; the first row is the whole scene image, the second row is the enlarged view of the part in the red box in the first row; the bottom row is the legend. Figure 4 In the figure, the left side shows the original image and reconstruction results of 6 dates in 2004, and the right side shows the original image and reconstruction results of 6 dates in 2010. The embodiment of the present invention is based on the MODISMOD09GQ product in the annual daily time series change curve of representative landforms such as mixed forests and crops. Figure 3 shown.

[0059] The implementation of each embodiment of the present invention is based on programmed processing performed by a device with processor functionality. Therefore, in practical engineering applications, the technical solutions and functions of each embodiment of the present invention are packaged into various modules. Based on this reality, and building upon the aforementioned embodiments, an embodiment of the present invention provides a daily NDVI reconstruction system that combines self-supervised learning with variational analysis. This system is used to implement the daily NDVI reconstruction method combining self-supervised learning with variational analysis described in the aforementioned method embodiments.

[0060] The system includes: a first main module, which is used to sample and screen NDVI data tiles on a global scale, and classify and randomly match the obtained data blocks according to different missing rates to construct a degraded-clean sample library for self-supervised network training; a second main module, which is used to develop a global-local dual-stream parallel NDVI spatiotemporal reconstruction network based on a self-supervised learning strategy, and design a loss function that includes global-local data consistency, multi-dimensional gradient prior and time smoothing prior to obtain seamless multi-day synthetic NDVI data; a third main module, which is used to use the obtained seamless multi-day synthetic NDVI data as reference information to pre-fill the original daily data, remove outliers from the pre-filled data, use one-dimensional variational filtering for further denoising and smoothing, and output the reconstructed daily seamless NDVI sequence.

[0061] The daily NDVI reconstruction system combining self-supervised learning and variational methods provided in the embodiments of the present invention addresses the problem that daily data is frequently missing information, making it difficult for each type of method to directly reconstruct it. By adopting the aforementioned modules and developing a self-supervised deep learning model with good regional generalization performance, the system mines its own supervisory information from large-scale raw data to drive network training, achieving high-precision reconstruction of multi-day synthetic data. Furthermore, a robust one-dimensional variational framework for NDVI time series reconstruction is constructed for daily data, achieving effective step-by-step reconstruction from multi-day synthesis to daily fine-grained reconstruction.

[0062] It should be noted that the system embodiments provided by the present invention are not only used to implement the methods in the above-mentioned method embodiments, but also used to implement the methods in other method embodiments provided by the present invention. The only difference lies in the setting of corresponding functional modules, and its principles are basically the same as the principles of the above-mentioned system embodiments provided by the present invention. As long as those skilled in the art refer to the specific technical solutions in other method embodiments on the basis of the above-mentioned system embodiments, obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and on the premise of ensuring the practicality of the technical solutions, they will improve the equipment in the above-mentioned system embodiments to obtain corresponding system embodiments for implementing the methods in other method embodiments.

[0063] Based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention also provides a daily NDVI reconstruction device combining self-supervised learning and variational analysis, comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the daily NDVI reconstruction method combining self-supervised learning and variational analysis.

[0064] In an embodiment of the present invention, the memory may be a non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or a volatile memory (volatile memory), such as a random-access memory (RAM). The memory is any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory in an embodiment of the present invention may also be a circuit or any other device that can implement a storage function, for storing program instructions and / or data.

[0065] In the embodiments of the present invention, the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field-programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of the present invention may be directly implemented and executed by a hardware processor, or by a combination of hardware and software modules within the processor.

[0066] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions cause the computer to execute the daily NDVI reconstruction method combining self-supervised learning and variational analysis, including: We sample and filter NDVI data tiles globally, classify and randomly match the resulting data blocks based on their missingness rates, and construct a degraded-clean sample library for self-supervised network training. Based on a self-supervised learning strategy, a global-local dual-stream parallel NDVI spatiotemporal reconstruction network is developed. A loss function that incorporates global-local data consistency, multi-dimensional gradient priors, and temporal smoothness priors is designed to obtain seamless multi-day synthetic NDVI data. The obtained seamless multi-day composite NDVI data is used as reference information to pre-fill the original daily data, and outliers are removed from the pre-filled data. One-dimensional variational filtering is used for further denoising and smoothing, and the reconstructed daily seamless NDVI sequence is output.

[0067] In summary, the present invention first constructs a global sample library based on a self-supervised learning strategy, effectively extracts spatiotemporal features through a global-local dual-stream parallel time series reconstruction network, and produces seamless multi-day synthetic data, wherein the global branch uses Transformer to model the entire time series, while the local branch uses a convolutional long short-term memory module to characterize local spatiotemporal laws. On this basis, a daily NDVI pre-filling method guided by intra-year time adjacent domain priors and inter-annual cycle similarity priors is developed. In view of the gross errors that may be introduced by factors such as surface cover changes and image quality marks during the pre-filling process, an adjacent window screening method is designed to remove noise points. Finally, a one-dimensional variational filter is used to obtain a high-quality seamless daily NDVI product. The present invention can integrate the accuracy and efficiency advantages of deep learning methods with the robustness advantages of variational methods, promote the production application of time series reconstruction algorithms in daily data reconstruction, and break through the bottleneck problem that existing methods are difficult to handle large-scale continuous missing daily NDVI.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. The daily NDVI reconstruction method combining self-supervised learning and variational analysis is characterized by: include: We sample and filter NDVI data tiles globally, classify and randomly match the resulting data blocks based on their missingness rates, and construct a degraded-clean sample library for self-supervised network training. Based on a self-supervised learning strategy, a global-local dual-stream parallel NDVI spatiotemporal reconstruction network is developed. A loss function that incorporates global-local data consistency, multi-dimensional gradient priors, and temporal smoothness priors is designed to obtain seamless multi-day synthetic NDVI data. The obtained seamless multi-day composite NDVI data is used as reference information to pre-fill the original daily data, and outliers are removed from the pre-filled data. One-dimensional variational filtering is used for further denoising and smoothing, and the reconstructed daily seamless NDVI sequence is output.

2. The daily NDVI reconstruction method combining self-supervised learning and variational analysis according to claim 1 is characterized in that: Constructing a degraded-clean sample library for self-supervised network training, including: Divide and crop NDVI data tiles globally, and select areas with a variety of landform types and climate types as candidate sample areas; Based on the data quality marking layer, data blocks with a completeness rate greater than a first threshold and data blocks with a missing rate greater than a second threshold are screened out according to the proportion of missing data; Randomly match the two types of data blocks, assign missing values ​​of data blocks with missing rates greater than the second threshold to data blocks with integrity rates greater than the first threshold at the corresponding positions, construct degraded data blocks, and perform quality marking. The constructed degraded data blocks and their quality labels are used as the input of the NDVI spatiotemporal reconstruction network, and the original data blocks with a completeness rate greater than the first threshold are used as label data for self-supervised learning to construct degraded-clean sample pairs.

3. The daily NDVI reconstruction method combining self-supervised learning and variational analysis according to claim 1 is characterized in that: Develop a global-local dual-stream parallel NDVI spatiotemporal reconstruction network, including: Construct a feature extraction module for extracting features from the input degraded data blocks and their quality marks; Construct a global-local temporal processing module, which includes a parallel framework of local and global temporal feature extraction branches. The local temporal branch processes temporal dependencies moment by moment, while the global temporal branch models long-term temporal relationships through multi-head self-attention. The two parallel branches simultaneously capture local temporal details and global temporal trends. Construct a dense connection module, including multiple layers of dense connections, where each layer receives the feature maps of all previous layers as input and passes its own output to all subsequent layers. Finally, by compressing the number of channels, a seamless image sequence is obtained after reconstruction. Construct loss functions, including global-local data consistency loss, multi-dimensional gradient loss, and temporal smoothing loss.

4. The daily NDVI reconstruction method combining self-supervised learning and variational analysis according to claim 3 is characterized in that: The local temporal branch processes sequence information moment by moment through a convolutional long short-term memory module and establishes local temporal dependencies to capture neighborhood temporal information.

5. The daily NDVI reconstruction method combining self-supervised learning and variational analysis according to claim 3 is characterized in that: The global timing branch uses the multi-head self-attention mechanism in Transformer to capture the key information of the entire timing and directly model the global timing relationship.

6. The daily NDVI reconstruction method combining self-supervised learning and variational analysis according to claim 1 is characterized in that: The original daily data is pre-filled, and outliers are removed from the pre-filled data. One-dimensional variational filtering is used for further denoising and smoothing, including: Missing data are pre-filled using interannual cycle priors for multi-year data and intra-annual temporal neighbor domain priors for synthetic data. For each pixel of the pre-filled image, the gross errors of the pre-filled data are screened and eliminated by setting the mean and standard deviation within the range and time neighborhood; One-dimensional variational filtering is used to constrain the local smoothness prior and interannual similarity prior of the NDVI time series through regularization to further remove residual noise.

7. A daily NDVI reconstruction system combining self-supervised learning and variational methods, characterized by: The system is used to implement the daily NDVI reconstruction method combining self-supervised learning and variational analysis as described in any one of claims 1 to 6.

8. The daily NDVI reconstruction system combining self-supervised learning and variational analysis according to claim 7, characterized in that: The system comprises: The first main module is used to sample and filter NDVI data tiles globally, classify and randomly match the obtained data blocks according to their missing rates, and construct a degraded-clean sample library for self-supervised network training; The second main module is used to develop a global-local dual-stream parallel NDVI spatiotemporal reconstruction network based on a self-supervised learning strategy, design a loss function that includes global-local data consistency, multi-dimensional gradient priors, and temporal smoothness priors to obtain seamless multi-day synthetic NDVI data; The third main module is used to use the obtained seamless multi-day synthetic NDVI data as reference information to pre-fill the original daily data, remove outliers from the pre-filled data, use one-dimensional variational filtering for further denoising and smoothing, and output the reconstructed daily seamless NDVI sequence.

9. A daily NDVI reconstruction device combining self-supervised learning and variational methods, characterized by: The invention comprises a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the daily NDVI reconstruction method combining self-supervised learning and variational analysis as described in any one of claims 1 to 6.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, which enable the computer to execute the daily NDVI reconstruction method combining self-supervised learning and variational analysis according to any one of claims 1 to 6.

Images