Day-to-day ndvi reconstruction method and system combining self-supervised learning with variation

Abstract

The application discloses a daily NDVI reconstruction method combining self-supervised learning with variation, firstly constructs a global sample library based on a self-supervised learning strategy, effectively extracts space-time features through a global-local double-flow parallel time series reconstruction network, and produces seamless multi-day synthetic data, wherein the global branch adopts a Transformer to model the whole time series, and the local branch adopts a convolution long short-term memory module to depict local space-time laws. Accordingly, a daily NDVI pre-filling method guided by in-year phase neighborhood prior and inter-annual cycle similarity prior is developed. In view of the gross errors possibly introduced by factors such as surface cover change and image quality marks in the pre-filling process, a neighboring window screening method is designed to remove noise points. Finally, one-dimensional variation filtering is adopted to obtain high-quality seamless daily NDVI products. The application integrates the precision and efficiency advantages of a deep learning method and the robustness advantages of a variation method, and breaks through the bottleneck problem that existing methods are difficult to handle large-scale continuous missing of daily NDVI.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing image processing and computer vision, and particularly relates to a remote sensing missing information time series reconstruction framework, and in particular to a daily NDVI reconstruction method and system combining self-supervised learning and variation. BACKGROUND

[0002] Optical remote sensing satellites are affected by internal factors such as sensor hardware limitations and external factors such as poor atmospheric observation conditions during imaging, resulting in great challenges in obtaining long-time high-quality daily normalized difference vegetation index (NDVI) data. Reconstruction methods based on time series information establish mathematical models on local time neighborhood or overall time series, effectively extract vegetation growth rules from temporal and spatial changes before and after the time series, and thus achieve the purpose of reconstruction, which is widely used in multi-year NDVI sequence reconstruction.

[0003] At present, the reconstruction methods based on time series can be mainly divided into three categories: time domain filtering based methods, variation reconstruction based methods, and deep learning based methods. The time domain filtering based methods perform interpolation, fitting, filtering and other operations in time or space dimensions, summarize the time variation rules of NDVI and establish corresponding mathematical models, but this kind of method has certain requirements for the number of good observations, it is difficult to reconstruct daily data with few effective observations, and the introduction of a large number of artificial parameters also increases the subjectivity and uncertainty of the reconstruction task. The variation reconstruction based method regards the reconstruction process as an ill-posed inverse problem, and fills in the missing values by constructing an energy functional and numerically solving it. This kind of method can deal with more complex missing conditions, but it is still difficult to directly process daily data with extremely high missing rate, and the high computational cost limits its application in actual data production. The deep learning based method has strong nonlinear feature extraction and parallel computing ability, and has been preliminarily applied in NDVI time series reconstruction field due to its high precision and efficiency, but the effect of deep learning model depends on the number and quality of training samples, and the contradiction between degraded-clean sample pairs makes it unable to be fully explored in the field of remote sensing time series reconstruction. In addition, the huge data volume and extremely high missing rate of daily data further increase the difficulty of deep learning method in daily data reconstruction application.

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

[0005] To solve the problem that each type of method is still difficult to directly reconstruct due to the extremely frequent information missing of daily data, the application provides a daily NDVI reconstruction method combining self-supervised learning and variation, a self-supervised deep learning model with good regional generalization performance is developed to mine the supervision information of the network from large-scale original data to drive network training, realize high-precision reconstruction of multi-day synthetic data, and then construct a robust NDVI time series reconstruction one-dimensional variation framework for daily data to realize step-by-step effective reconstruction from multi-day synthesis to daily precision.

[0006] According to an aspect of the application, a daily NDVI reconstruction method combining self-supervised learning and variation is provided, comprising:

[0007] The NDVI data tiles are sampled and screened globally, and the obtained data blocks are classified and randomly matched according to different missing rates to construct a degraded-clean sample library for self-supervised network training;

[0008] Based on the self-supervised learning strategy, a global-local double-flow parallel NDVI spatio-temporal reconstruction network is developed, a loss function containing overall-local data consistency, multi-dimensional gradient prior and time smoothing prior is designed, and seamless multi-day synthetic NDVI data is obtained;

[0009] The obtained seamless multi-day synthetic NDVI data is used as reference information to pre-fill the original daily data, and the pre-filled data is subjected to outlier rejection, one-dimensional variation filtering is used for further denoising and smoothing, and the reconstructed daily seamless NDVI sequence is output.

[0010] As a further technical solution, the degraded-clean sample library for self-supervised network training further comprises:

[0011] The NDVI data tiles are blocked and cropped globally, and regions containing various ground object types and climate types are selected as sample candidate regions;

[0012] According to the quality label layer of the data, data blocks with a completeness rate greater than a first threshold and data blocks with a missing rate greater than a second threshold are selected according to the proportion of missing data;

[0013] The two types of data blocks are randomly matched, the missing values of the data blocks with a missing rate greater than the second threshold are assigned to the corresponding positions of the data blocks with a completeness rate greater than the first threshold, degraded data blocks are constructed, and quality labels are marked;

[0014] The constructed degraded data blocks and their quality labels are used as the input of the NDVI spatio-temporal reconstruction network, and the original data blocks with a completeness rate greater than the first threshold are used as the label data of self-supervised learning to construct degraded-clean sample pairs.

[0015] As a further technical solution, a global-local double-flow parallel NDVI spatio-temporal reconstruction network is developed, comprising:

[0016] A feature extraction module is constructed for feature extraction of the input degraded data block and its quality label;

[0017] A global-local time series processing module is constructed, including a local and global time series feature extraction branch parallel framework, wherein the local time series branch processes time series dependence, and the global time series branch models long time series relationship through multi-head self-attention, and the double-parallel branch simultaneously captures local time series details and global time series trends;

[0018] A dense connection module is constructed, including multiple layers of dense connection, wherein each layer receives the feature maps of all previous layers as input, and the output of itself is transmitted to all subsequent layers, and finally the reconstructed seamless image sequence is obtained by compressing the channel number;

[0019] A loss function is constructed, including overall-local data consistency loss, multi-dimensional gradient loss and time smoothing loss.

[0020] As a further technical solution, the local time series branch processes sequence information and establishes local time series dependence through a convolutional long short-term memory module, capturing neighborhood time series information.

[0021] As a further technical solution, the global time series branch adopts the multi-head self-attention mechanism in Transformer to capture key information of the entire time series, directly modeling the global time series relationship.

[0022] As a further technical solution, the original daily data is pre-filled, and the pre-filled data is subjected to outlier rejection, and one-dimensional variational filtering is used for further denoising and smoothing, comprising:

[0023] The missing data is pre-filled through the interannual cycle prior of multi-year data and the intra-annual phase neighborhood prior of synthetic data;

[0024] For each pixel of the pre-filled image, the mean and standard deviation of the range time neighborhood are set to screen and reject the gross error of the pre-filled data;

[0025] One-dimensional variational filtering is used to constrain the local smoothness prior and interannual similarity prior of the NDVI time series through regularization, further removing residual noise.

[0026] According to an aspect of the present application, a daily NDVI reconstruction system combining self-supervised learning and variation is provided, which is used to implement the daily NDVI reconstruction method combining self-supervised learning and variation.

[0027] As a further technical solution, the system comprises:

[0028] The first main module is used for sampling and screening NDVI data tiles globally, and classifying and randomly matching data blocks obtained according to different missing rates, to construct a degraded-clean sample library for self-supervised network training.

[0029] The second main module is used for developing a global-local double-flow parallel NDVI spatiotemporal reconstruction network based on a self-supervised learning strategy, designing a loss function containing overall-local data consistency, multi-dimensional gradient prior and time smoothing prior, and obtaining seamless multi-day synthetic NDVI data.

[0030] The third main module is used for using the obtained seamless multi-day synthetic NDVI data as reference information to pre-fill the original daily data, removing outliers after pre-filling, and adopting one-dimensional variational filtering for further denoising and smoothing processing, to output reconstructed daily seamless NDVI sequences.

[0031] According to an aspect of the present application, a self-supervised learning and variational combined daily NDVI reconstruction device is provided, comprising a memory and a processor, the memory storing program instructions executed by the processor, and the processor calling the program instructions to execute the self-supervised learning and variational combined daily NDVI reconstruction method.

[0032] According to an aspect of the present application, a non-transitory computer readable storage medium is provided, which stores computer instructions, and the computer instructions make the computer execute the self-supervised learning and variational combined daily NDVI reconstruction method.

[0033] The method of the present application can organically combine deep learning and variational methods, fully consider the temporal prior characteristics of NDVI, and realize step-by-step effective reconstruction of daily NDVI data. Compared with the prior art, the beneficial effects of the present application mainly include:

[0034] (1) The present application selects a variety of sample candidate regions globally, constructs sample pairs through real data, drives self-supervised network training, solves the problem that the prior art is limited by the construction of degraded-clean sample pairs, and the high precision and high parallel computing capability of deep learning technology cannot be fully explored in the field of NDVI temporal reconstruction.

[0035] (2) In the self-supervised spatio-temporal reconstruction network of synthetic data, the local-global double-flow parallel time series processing module is developed by combining the Transformer and ConvLSTM modules, which can fully extract the spatio-temporal features of NDVI image sequences and realize high-precision reconstruction of synthetic data.

[0036] (3) The step-by-step reconstruction framework combining self-supervised learning and one-dimensional variation is adopted, the advantages of the two reconstruction algorithms are fully combined, and the stability of NDVI data filling is improved through multi-day synthesis to daily fine step-by-step reconstruction in the time dimension, the intra-annual neighborhood priori and inter-annual cycle priori of NDVI time series are effectively mined, and the bottleneck problem that single method is difficult to process large-scale continuous missing of daily data is broken through. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0038] Figure 1 is the reconstruction flowchart provided by the embodiment of the present application.

[0039] Figure 2 is the visualization effect schematic diagram based on MODIS MOD09GQ product of the embodiment of the present application.

[0040] Figure 3 is the intra-annual daily time series change curve schematic diagram based on MODIS MOD09GQ product of the embodiment of the present application in the representative mixed forest and crop.

[0041] Figure 4 is the visualization effect schematic diagram based on MODIS MOD09GQ product of the embodiment of the present application. DETAILED DESCRIPTION

[0042] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments but not all of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application. In addition, the technical features in each embodiment or in a single embodiment provided by the present application can be combined with each other at will to form new technical solutions, and the combination is not restricted by the sequence of steps and / or the mode of structural composition, but should be based on the fact that a person of ordinary skill in the art can realize it. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the present application.

[0043] In view of the problem that extremely frequent cloud cover leads to a large number of long-term continuous missing of daily NDVI data, and the existing single reconstruction method is difficult to effectively reconstruct it, the present application provides a daily NDVI reconstruction method combining self-supervised learning and variation, first generates high-quality multi-day seamless synthesis data based on large-scale original data to assist reconstruction, then effectively pre-fills the daily NDVI by using in-phase neighborhood prior and inter-annual cycle similarity prior, and finally uses one-dimensional variation reconstruction method for further denoising and smoothing to obtain high-quality seamless daily NDVI data.

[0044] The present application combines deep learning and variation method, fully gives play to the advantages of the two methods, reconstructs multi-day synthesis data through deep learning algorithm, pre-fills by using latent in-phase neighborhood prior and inter-annual cycle prior, and further smoothes by using robust one-dimensional variation filtering, so that spatially and temporally seamless high-quality daily NDVI data can be obtained.

[0045] See Figure 1 The daily NDVI reconstruction method combining self-supervised learning and variation provided by the present application comprises the following steps:

[0046] Step 1: The NDVI data tiles are divided and cropped in a global range, and the obtained data blocks are classified and randomly matched according to different missing rates to construct a degraded-clean sample library for self-supervised network training.

[0047] Specifically, it comprises:

[0048] Step 1.1: The NDVI data tiles are divided and cropped in a global range, and regions containing various land cover types and climate types are selected as sample candidate regions.

[0049] Step 1.2: According to the quality marking layer of the data, that is, the layer reflecting the quality condition of each vegetation index pixel carried by the data itself, the data blocks with better quality and the data blocks with more cloud coverage are screened according to the proportion of missing data. It should be noted that the determination of the data blocks with better quality and the data blocks with more cloud coverage can select different thresholds to divide according to different regions, and the present application does not limit this. As a preferred embodiment, data with a completeness rate > 70% can be selected as the data block with better quality, and data with a missing rate > 30% can be selected as the data block with more cloud coverage.

[0050] 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 corresponding positions of the data blocks with better quality, construct the degraded data blocks, and make corresponding processing to the quality marks.

[0051] Step 1.4: The degraded data blocks and the quality marks constructed are taken as the input of step 2, and the original data blocks with better quality are taken as the label data of step 2, to construct the degraded-clean sample pairs.

[0052] Step 2: Based on the self-supervised learning strategy, a global-local double-flow parallel NDVI spatio-temporal reconstruction network is developed, a loss function containing overall-local data consistency, multi-dimensional gradient prior and time smoothing prior is designed, and seamless multi-day synthetic NDVI data is obtained.

[0053] The reconstruction network takes the degraded data blocks and the corresponding quality marks as the input, first encodes the original data under different receptive fields through a multi-scale information extraction module to obtain multi-scale features, and inputs the extracted feature information into a local spatio-temporal processing branch and a global time series processing branch. The local time series branch processes the sequence information and establishes the local time series dependency through a convolutional long short-term memory module (ConvLSTM) at each time, and captures the neighborhood time series information. The global time series processing branch uses the multi-head self-attention mechanism in the Transformer to capture the key information of the entire time series, and directly models the global time series relationship. The output features of the two branches will be strengthened through a dense connection module to enhance feature reuse and significantly improve feature representation capability. Finally, the network obtains the reconstruction result through feature dimension reduction. The reconstruction result and the label data are calculated in the loss in the cloud-free area and the pixel quality uncertain area, including overall-local data consistency loss, multi-dimensional gradient loss and time smoothing loss.

[0054] Specifically, it includes:

[0055] Step 2.1: Construct a feature extraction module. The degraded NDVI data blocks and the quality marks constructed in step 1.4 are input: N ∈ℝW×H×T With F ∈ ℝ W×H×T where T is the time length of the time series image, H and W are the height and width of the image. The two data are respectively extracted for each time phase:

[0056]

[0057] where N and are the degraded NDVI data and the extracted image features, F and are the corresponding quality labels and the extracted label features, w is the convolution kernel, b represents the bias, is the convolution operation, R represents the ReLU activation function.

[0058] Step 2.2: Construct a global-local time series processing module. It contains a parallel framework of local and global time series feature extraction branches. The local branch processes time series dependence for each time, and the global branch models long time series relationship through multi-head self-attention. Both branches capture local time series details and global time series trends at the same time.

[0059] Step 2.2.1: Local time series feature extraction branch. In this branch, the extracted image features and label features are first multiplied element by element, then input into the ConvLSTM unit to effectively extract local spatio-temporal features, and finally the time dimension is adjusted to a specific length using convolution operation. The process can be represented as:

[0060]

[0061] where represents the ConvLSTM time series processing unit, is the convolution kernel, is the local time series feature extracted by this branch.

[0062] Step 2.2.2: Global time series feature extraction branch. In this branch, first, N and are dimensionally transformed to unify Batchsize, H and W to the same dimension, keeping the time dimension T and the feature dimension C , the dimensionally transformed features ∈ ℝ (B×H×W)×T×C and ∈ ℝ (B×H×W)×T×CElement-wise multiplication is performed and input into the Transformer unit to extract global temporal information through multi-head self-attention mechanism Finally, the local and global information of the two branches are spliced and convoluted to adjust the time dimension to a specific length, outputting the results of the global-local temporal processing module The process can be represented by equations (4) and (5), where Concat represents the splicing operation along the time dimension:

[0063]

[0064] Step 2.3: Building a dense connection module. The reuse of temporal features and gradient flow is significantly improved through multiple layers of dense connections. Each layer receives the feature maps of all previous layers as input and passes its output to all subsequent layers. Finally, a 1x1 convolution is used to compress the channel number and obtain the reconstructed seamless image sequence The process can be represented by:

[0065]

[0066] where H l is the l th dense connection block, which includes a batch normalization layer, a ReLU activation function, and a convolutional layer, N l ( l =1,…, L ) represents the dense features of the l th layer, is the convolution kernel of the last layer.

[0067] Step 2.4: Building a loss function.

[0068] Step 2.4.1: Global-local data consistency loss. The local mask loss and global loss between the reconstructed results and the label data are calculated using the 1-norm, which can be represented as:

[0069]

[0070] where B represents the Batchsize size. M 1 represents the quality of the label data as a well-observed position mask, i.e., the global mask; M 2 represents the input data as missing data and the label data as a well-observed position mask, i.e., the local mask of the missing area.

[0071] Step 2.4.2: Multi-dimensional gradient loss. It includes H , W , TThe gradient loss in three directions can be expressed as:

[0072]

[0073] wherein represents the gradient operation on the image along H , W , T the dimension.

[0074] Step 2.4.3: Time smoothing loss. Considering the gradual change of NDVI in the time dimension, i.e., the vegetation growth generally follows the gradual change rule, the time smoothing loss is introduced as the prior constraint of the sequence, which can be expressed as:

[0075]

[0076] wherein represents the reconstructed image at the t time.

[0077] Step 2.4.4: The total loss function is constructed by weighted sum of the three loss functions, and are the balance coefficients:

[0078]

[0079] Step 3: Using the seamless multi-day synthetic NDVI data obtained in step 2 as reference information, the original daily data is pre-filled, and the pre-filled data is subjected to outlier rejection, and one-dimensional variational filtering is used for further denoising and smoothing processing. The output seamless daily data is the required high-quality NDVI sequence.

[0080] Specifically includes:

[0081] Step 3.1: Effective pre-filling of daily data. Through the interannual cycle prior of multi-year data and the intra-annual neighborhood prior of synthetic data, the missing data is pre-filled as much as possible to increase the number of effective observations.

[0082] Step 3.1.1: Intra-annual temporal neighborhood prior extraction. For the image on the acquisition date (day of year, DOY), search for the adjacent seamless synthetic NDVI image generated in step 2 within a multi-day window as the intra-annual neighborhood prior .

[0083] Step 3.1.2: Interannual cycle prior information extraction. For a certain observation date DOY, traverse the images of all years Y on this date, and take the median of multi-year effective observations as the interannual cycle prior information , which can be specifically expressed as:

[0084]

[0085] in I Y is the image of year Y on date DOY, F Y is the corresponding effective observation mask.

[0086] 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:

[0087]

[0088] 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:

[0089]

[0090] 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:

[0091]

[0092] 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 a regularization parameter. The final sequence obtained through the variational model is a seamless daily NDVI sequence.

[0093] The visualization effect of the embodiment of the application based on the MODIS MOD09GQ product is as shown in Figure 2 and 4 . 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, and the second row is the local enlarged view in the red box in the first row; the last row is a legend. Figure 4 In the figure, the left side is the original image and the reconstruction result of six dates in 2004, and the right side is the original image and the reconstruction result of six dates in 2010. The embodiment of the application based on the MODIS MOD09GQ product is as shown in Figure 3 .

[0094] The implementation basis of each embodiment of the application is that the implementation is realized through programmed processing of a device with a processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the application are packaged into various modules. Based on this actual situation, on the basis of each of the above embodiments, the embodiment of the application provides a daily NDVI reconstruction system combining self-supervised learning and variation, which is used to execute the daily NDVI reconstruction method combining self-supervised learning and variation in the above method embodiment.

[0095] The system comprises: a first main module, which is used to sample and screen NDVI data tiles in a global range, 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 double-flow parallel NDVI spatio-temporal reconstruction network based on a self-supervised learning strategy, design a loss function comprising overall-local data consistency, multi-dimensional gradient prior and time smoothing prior, and obtain seamless multi-day synthetic NDVI data; and 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 after pre-filling, and adopt one-dimensional variation filtering to further denoise and smooth the data, and output the reconstructed seamless daily NDVI sequence.

[0096] The self-supervised learning and variational combined daily NDVI reconstruction system provided by the embodiment of the present application faces the problem that each type of method is still difficult to directly reconstruct due to extremely frequent information missing of daily data, adopts the foregoing modules, mines the self-supervised information from large-scale original data to drive network training by developing a self-supervised deep learning model with good regional generalization performance, realizes high-precision reconstruction of multi-day synthetic data, and then constructs a robust NDVI time series reconstruction one-dimensional variational framework for daily data to realize effective step-by-step reconstruction from multi-day synthesis to daily precision.

[0097] It should be noted that the system embodiments provided by the present application are used to implement the methods in the method embodiments of the present application, and are also used to implement the methods in other method embodiments provided by the present application. The difference is only that the corresponding function modules are set, and the principle is basically the same as that of the above-mentioned system embodiments provided by the present application. As long as the person skilled in the art improves the equipment in the above-mentioned system embodiments on the basis of the above-mentioned system embodiments, refers to the specific technical solutions in other method embodiments, obtains the corresponding technical means by combining technical features, and the technical solutions composed of these technical means, as long as the technical solutions have practicality, the corresponding system class embodiments are obtained, which are used to implement the methods in other method class embodiments.

[0098] Based on the same inventive concept as the foregoing embodiments, the embodiment of the present application also provides a self-supervised learning and variational combined daily NDVI reconstruction device, comprising a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the self-supervised learning and variational combined daily NDVI reconstruction method.

[0099] In the embodiment of the present application, the memory can be a non-volatile memory such as a hard disk (HDD) or a solid-state drive (SSD), and can also be a volatile memory such as a random-access memory (RAM). The memory can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but is not limited to this. The memory in the embodiment of the present application can also be a circuit or any other device capable of realizing a storage function, used for storing program instructions and / or data.

[0100] In the embodiments of the present application, the processor can be a general 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, a discrete hardware component, and can implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application. The general processor can be a microprocessor or any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as execution completed by a hardware processor, or executed by a combination of hardware and software modules in the processor.

[0101] According to an aspect of the present application, a non-transitory computer readable storage medium is provided, which stores computer instructions for performing the self-supervised learning and variational combined daily NDVI reconstruction method, comprising:

[0102] Sampling and screening NDVI data tiles globally, and classifying and randomly matching the obtained data blocks according to different missing rates to construct a degraded-clean sample library for self-supervised network training;

[0103] Based on the self-supervised learning strategy, a global-local dual-flow parallel NDVI spatio-temporal reconstruction network is developed, a loss function including overall-local data consistency, multi-dimensional gradient prior and time smoothing prior is designed, and seamless multi-day synthetic NDVI data is obtained;

[0104] Using the obtained seamless multi-day synthetic NDVI data as reference information, the original daily data is pre-filled, and the pre-filled data is subjected to outlier rejection, and one-dimensional variational filtering is used for further denoising and smoothing processing, and the reconstructed daily seamless NDVI sequence is output.

[0105] In summary of the embodiments, the present application firstly constructs a global sample library based on a self-supervised learning strategy, effectively extracts spatio-temporal features through a global-local dual-flow parallel time series reconstruction network, and produces seamless multi-day synthetic data. The global branch uses Transformer to model the entire time series, while the local branch uses a convolutional long short-term memory module to depict local spatio-temporal laws. On this basis, a daily NDVI pre-filling method guided by intra-annual temporal neighborhood prior and inter-annual cycle similarity prior is developed. Aiming at the gross errors possibly introduced by surface cover change, image quality marks and other factors in the pre-filling process, a neighboring window screening method is designed to reject noise points. Finally, one-dimensional variational filtering is used to obtain high-quality seamless daily NDVI products. The present application can integrate the precision and efficiency advantages of deep learning methods and 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 of daily NDVI.

[0106] It should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present application.

Claims

1. A method for daily NDVI reconstruction combining self-supervised learning with variational, characterized in that, The application comprises the following steps: Sampling and screening NDVI data tiles globally, and classifying and randomly matching the obtained data blocks according to different missing rates to construct a degraded-clean sample library for self-supervised network training; Based on the self-supervised learning strategy, a global-local dual-flow parallel NDVI spatio-temporal reconstruction network is developed to obtain seamless multi-day synthetic NDVI data; The global-local dual-flow parallel NDVI spatio-temporal reconstruction network comprises the following steps: a feature extraction module is constructed to extract features from the input degraded data block and its quality label; a global-local temporal processing module is constructed, which comprises a local and global temporal feature extraction branch parallel framework, wherein the local temporal branch processes the temporal dependence of each time step, and the global temporal branch models the long temporal relationship through a multi-head self-attention mechanism, and the two branches are connected in parallel to capture local temporal details and global temporal trends; a dense connection module is constructed, which comprises multiple layers of dense connection, wherein each layer receives the feature maps of all previous layers as input, and the output of the layer is transmitted to all subsequent layers, and finally the channel number is compressed to obtain the reconstructed seamless image sequence; a loss function is constructed, which comprises overall-local data consistency loss, multi-dimensional gradient loss and time smoothing loss; The obtained seamless multi-day synthetic NDVI data is used as reference information to pre-fill the original daily data, and the pre-filled data is subjected to outlier rejection, one-dimensional variational filtering for further denoising and smoothing, and the reconstructed daily seamless NDVI sequence is output.

2. The method according to claim 1, wherein, The degraded-clean sample library for self-supervised network training also comprises the following steps: The NDVI data tiles are blocked and cropped globally, and regions containing various land cover types and climate types are selected as sample candidate regions; According to the quality label layer of the data, 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 according to the proportion of missing data; The two types of data blocks are randomly matched, the missing values of the data blocks with a missing rate greater than the second threshold are assigned to the corresponding positions of the data blocks with a completeness rate greater than the first threshold, degraded data blocks are constructed, and quality labels are marked; The constructed degraded data blocks and their quality labels are used as the input of the NDVI spatio-temporal reconstruction network, and the original data blocks with a completeness rate greater than the first threshold are used as the label data of self-supervised learning to construct degraded-clean sample pairs.

3. The method according to claim 1, wherein, The local temporal branch processes sequence information and establishes local temporal dependence through a convolutional long short-term memory module, and captures neighborhood temporal information.

4. The method of claim 1, wherein the method further comprises: The global temporal branch adopts a multi-head self-attention mechanism in the Transformer to capture key information of the entire time sequence and directly model the global temporal relationship.

5. The method of claim 1, wherein the method further comprises: The original daily data is pre-filled, and the pre-filled data is subjected to outlier rejection, one-dimensional variational filtering for further denoising and smoothing, which comprises the following steps: The missing data is pre-filled through the interannual cycle prior of multi-year data and the intra-annual phase neighborhood prior of synthetic data; For each pixel of the pre-interpolated image, the mean and standard deviation of the time neighborhood within the set range are used to filter and remove the gross errors of the pre-interpolated data; A one-dimensional variational filter is used to further remove residual noise by constraining the local smoothness prior and interannual similarity prior of the NDVI time series through regularization.

6. A day-to-day NDVI reconstruction system combining self-supervised learning with variational, characterized in that, The system is used to implement the daily NDVI reconstruction method of self-supervised learning combined with variation, as claimed in any one of claims 1 to 5.

7. The system for day-to-day NDVI reconstruction using self-supervised learning combined with variational inference according to claim 6, wherein, The system comprises: A first main module for sampling and screening NDVI data tiles globally, and classifying and randomly matching 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 for developing a global-local double-flow parallel NDVI spatio-temporal reconstruction network based on a self-supervised learning strategy, designing a loss function containing overall-local data consistency, multi-dimensional gradient prior and time smoothness prior, and obtaining seamless multi-day synthetic NDVI data; A third main module for using the obtained seamless multi-day synthetic NDVI data as reference information to pre-interpolate the original daily data, removing outliers after pre-interpolation, and using a one-dimensional variational filter for further denoising and smoothing to output the reconstructed daily seamless NDVI sequence.

8. A device for reconstructing NDVI day by day by self-supervised learning combined with variation, characterized in that, A memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor invokes the program instructions to execute the daily NDVI reconstruction method of self-supervised learning combined with variation, as claimed in any one of claims 1 to 5.

9. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium stores computer instructions, which make the computer execute the daily NDVI reconstruction method of self-supervised learning combined with variation, as claimed in any one of claims 1 to 5.

Images