A Simple and Robust Method for Long-Term Reconstruction of Landsat Vegetation Indices

Abstract

The application discloses a simple and robust long-time sequence reconstruction method for land satellite vegetation index, belonging to the technical field of land satellite, and comprises the following steps: acquiring a data source of a land satellite, pre-processing the data source, wherein the data source comprises Landsat NDVI data and MODIS NDVI data; performing data optimization on the Landsat NDVI data based on a 3sigma criterion and a bilinear interpolation method to obtain first optimized data; constructing a Prophet model based on the first optimized data to obtain annual, seasonal and residual components; and performing NDVI reconstruction on pixels marked as null based on the Prophet model to generate spatiotemporally continuous clear-sky Landsat NDVI data. The method has good fusion effect on the case where surface coverage (such as vegetation) changes suddenly, is simple in accurate registration, is not prone to systematic errors, is easy to accurately compare NDVI data, and is more accurate in analysis results. Meanwhile, by adjusting a prior sparsity, overfitting and underfitting phenomena are avoided.

Description

Technical Field

[0001] This invention relates to the field of land satellite technology, specifically to a simple and robust method for long-term reconstruction of land satellite vegetation indices. Background Technology

[0002] Vegetation indices are products of multiple bands in the visible and near-infrared light bands of satellites, providing a simple and effective measure of surface vegetation conditions. Currently, the Normalized Difference Vegetation Index (NDVI) has become the most widely used vegetation index due to its simple inversion algorithm and clear physical meaning. It is highly sensitive to the degree of surface vegetation cover and is one of the commonly used indicators for detecting and indicating vegetation growth status, vegetation cover, and eliminating some radiation errors.

[0003] Normalized Difference Vegetation Index (NDVI) time-series satellite data has been widely used for monitoring vegetation in terrestrial ecosystems over the past few decades. Currently, published NDVI time-series data products are mainly provided by satellites such as MODIS, SpOTVGT, and AVHRR. However, the spatial resolution of these NDVI products is relatively coarse (ranging from a few hundred meters to several kilometers), which limits their application in heterogeneous areas of the land surface. For example, MODIS sensor satellite data, with a revisit period of 1-2 days, has a high temporal resolution that can be used to identify and monitor dynamic changes in surface vegetation, but its spatial resolution reaches 250m-1000m, making it difficult to characterize high-precision surface information. Landsat satellite data, with a spatial resolution of 30m, can identify different land features with high accuracy and has rich historical data, making it highly valuable for calculating long-term and high-resolution NDVI vegetation indices. However, Landsat data has a relatively low temporal resolution (16-day revisit cycle), and due to cloud pollution interfering with optical imagery, Landsat NDVI data is severely lacking, making it difficult to obtain continuous and effective remote sensing images and thus failing to meet the requirements for vegetation dynamic monitoring.

[0004] To meet the requirements of high spatial and temporal resolution in dynamic remote sensing monitoring of land surface information, many scholars have proposed spatiotemporal fusion methods for remote sensing data. This involves fusing Landsat data (high spatial, low temporal resolution) with remote sensing data (low spatial, high temporal resolution, such as MODIS data) to generate remote sensing data with both high spatial and temporal resolution. Currently, representative spatiotemporal fusion algorithms include STARFM, FSDAF, and IFSDAF; however, these methods perform poorly in cases of abrupt changes in land cover (such as vegetation); precise registration is complex and prone to systematic errors; and because they typically use only a limited amount of remote sensing data at finite time points, it is difficult to compare interannual trends in NDVI. Furthermore, research on generating interannual NDVI trends by combining time-series decomposition and reconstruction models with Landsat NDVI is still lacking.

[0005] Based on this, the present invention designs a simple and robust method for long-term series reconstruction of land satellite vegetation indices to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a simple and robust method for reconstructing long-term series of vegetation indices from land satellites, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a simple and robust method for reconstructing long-term series of landsat vegetation indices, the reconstruction method comprising: acquiring a data source of landsat data, preprocessing the data source, wherein the data source includes Landsat NDVI data and MODIS NDVI data;

[0008] The Landsat NDVI data was optimized based on the 3σ criterion and bilinear interpolation method to obtain the first optimized data.

[0009] Based on the first optimized data, a Prophet model is constructed to obtain interannual, seasonal, and residual components;

[0010] Based on the Prophet model, NDVI reconstruction is performed on pixels marked as null to generate spatiotemporally continuous clear sky Landsat NDVI data.

[0011] As a further aspect of the present invention, the preprocessing of the data source specifically includes:

[0012] The pixels with atmospheric influence coefficient DA≥N in the Landsat NDVI data are masked, and the pixels with atmospheric influence coefficient DA<N are atmospherically corrected to obtain the corrected pixels.

[0013] The corrected pixels are processed to remove cloud contamination and stripe differences, thus completing the preprocessing of the data source.

[0014] As a further aspect of the present invention, the specific steps for optimizing the Landsat NDVI data based on the 3σ criterion and bilinear interpolation method are as follows:

[0015] S1. Obtain Landsat NDVI data, remove outliers from the Landsat NDVI time series data, and obtain de-outliered Landsat NDVI data;

[0016] S2. Based on bilinear interpolation, the isolated Landsat VINDVI data is reconstructed using spatial resolution to obtain reconstructed Landsat VINDVI data.

[0017] S3. Evaluate the accuracy of the reconstructed Landsat NDVI data. If the spatial resolution of the reconstructed Landsat NDVI data meets the set threshold, then the reconstructed Landsat NDVI data is the first optimized data.

[0018] As a further aspect of the present invention, the calculation formula for the Prophet model is as follows:

[0019]

[0020] Where y(t) is the observed value at time t, g(t) is the interannual variation component, s(t) is the seasonal variation component, and ε i Let γ be the residual component, k and δ be the growth rate and the relevant adjustment coefficient, respectively, and m be the compensation parameter; γ is the -S j δ j S j δ represents the time when the j-th (j=1,2,…,S) inflection point occurs. j For S j rate of change at time, a n and b n These are the coefficients of the Fourier series, and P is the period.

[0021] As a further aspect of the present invention, the accuracy evaluation of the reconstructed Landsat NDVI data in step S3 includes both direct and indirect methods.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. This invention can solve the problem of abrupt changes in land cover by adding breakpoints in interannual variations, and can also handle high interannual NDVI dynamic changes by merging seasonal variations. It has a good fusion effect on the case of abrupt changes in land cover (such as vegetation).

[0024] 2. The model constructed in this invention is relatively simple, using only single-source data (such as Landsat) to avoid image registration and mismatch problems. Accurate registration is relatively simple and less prone to systematic errors.

[0025] 3. The model constructed in this invention uses all available images and effective pixels, making it easy to accurately compare NDVI data and resulting in more accurate analysis results.

[0026] 4. The method of the present invention avoids overfitting and underfitting by adjusting the prior sparsity.

[0027] Appendix Explanation

[0028] Table 1 shows the accuracy assessment of Landsat NDVI data reconstructed pixel by pixel in Nanjing; Attached Figure Description

[0029] Figure 1 This is a flowchart of the Landsat NDVI long-time series data reconstruction method of the present invention;

[0030] Figure 2 This is a schematic diagram of the pixel-by-pixel long-time Landsat NDVI reconstruction of the present invention;

[0031] Figure 3 This is a schematic diagram of the pixel-by-pixel long-time Landsat NDVI decomposition of the present invention;

[0032] Figure 4 Comparison of Landsat NDVI data reconstructed for this invention with MODIS satellite data: (a) NDVI spatial pattern scatter plot; (b) NDVI trend scatter plot;

[0033] Figure 5 A pixel-by-pixel comparison of the Landsat NDVI reconstructed in this invention and the observed MODIS NDVI: (a)-(b) comparison of NDVI spatial patterns; (c)-(d) comparison of NDVI trends;

[0034] Figure 6 This invention provides a comparison of the interannual variation of Nanjing's urban NDVI based on reconstructed Landsat NDVI and MODIS NDVI. Detailed Implementation

[0035] Please see Figure 1-6The present invention provides a technical solution: a simple and robust method for reconstructing long-term series of landsat vegetation indices, the reconstruction method comprising: acquiring a data source of landsat data, preprocessing the data source, the data source including Landsat NDVI data and MODIS NDVI data;

[0036] Specifically, the data sources include Landsat NDVI data and MODIS NDVI data. First, pixels in the Landsat NDVI data with strong atmospheric effects are masked, while pixels with weak atmospheric effects are atmospherically corrected. At the same time, pixels with cloud contamination and stripe differences are removed, thus completing the preprocessing of the data source.

[0037] The Landsat NDVI data was optimized based on the 3σ criterion and bilinear interpolation method to obtain the first optimized data.

[0038] Before data reconstruction, outliers in the NDVI time series data were removed using the 3σ criterion, and all Landsat imagery was resampled to a spatial resolution of 120m using bilinear interpolation. MODIS NDVI data, with a spatial resolution of 1km, was used to evaluate the accuracy of the reconstructed Landsat NDVI data.

[0039] Based on the first optimized data, a Prophet model is constructed to obtain interannual, seasonal, and residual components;

[0040] Based on the Prophet model, NDVI reconstruction is performed on pixels marked as null to generate spatiotemporally continuous clear sky Landsat NDVI data. The method adopted in this invention has good fusion effect in cases of abrupt changes in land cover (such as vegetation), is relatively simple to accurately register, is not prone to systematic errors, and is easy to accurately compare NDVI data, resulting in more accurate analysis results. At the same time, by adjusting the prior sparsity, overfitting and underfitting phenomena are avoided.

[0041] As a further aspect of the present invention, the preprocessing of the data source specifically includes:

[0042] The pixels with atmospheric influence coefficient DA≥N in the Landsat NDVI data are masked, and the pixels with atmospheric influence coefficient DA<N are atmospherically corrected to obtain the corrected pixels.

[0043] The corrected pixels are processed to remove cloud contamination and stripe differences, thus completing the preprocessing of the data source.

[0044] As a further aspect of the present invention, the specific steps for optimizing the Landsat NDVI data based on the 3σ criterion and bilinear interpolation method are as follows:

[0045] S1. Obtain Landsat NDVI data, remove outliers from the Landsat NDVI time series data, and obtain de-outliered Landsat NDVI data;

[0046] S2. Based on bilinear interpolation, the isolated Landsat VINDVI data is reconstructed using spatial resolution to obtain reconstructed Landsat VINDVI data.

[0047] S3. Evaluate the accuracy of the reconstructed Landsat NDVI data. If the spatial resolution of the reconstructed Landsat NDVI data meets the set threshold, then the reconstructed Landsat NDVI data is the first optimized data.

[0048] As a further aspect of the present invention, the calculation formula for the Prophet model is as follows:

[0049]

[0050] Where y(t) is the observed value at time t, g(t) is the interannual variation component, s(t) is the seasonal variation component, and ε i Let γ be the residual component, k and δ be the growth rate and the relevant adjustment coefficient, respectively, and m be the compensation parameter; γ is the -S j δ j S j δ represents the time when the j-th (j=1,2,…,S) inflection point occurs. j For S j rate of change at time, a n and b n These are the coefficients of the Fourier series, and P is the period.

[0051] As a further aspect of the present invention, the accuracy evaluation of the reconstructed Landsat NDVI data in step S3 includes both direct and indirect methods.

[0052] Specifically, the direct method involves comparing the coefficient of determination (R²) of the reconstructed time-series NDVI data with the original Landsat NDVI data. 2 The model was evaluated using mean absolute error (MAE) and root mean square error (RMSE). Furthermore, based on hundreds of pixel-level observation-reconstruction data pairs from all clear-sky observation dates, the spatiotemporal variation trends were characterized by comparing the reconstructed NDVI data with the original NDVI data, further evaluating the model performance.

[0053] Taking Nanjing as an example, the accuracy comparison between the reconstructed NDVI and the original Landsat NDVI is shown in Table 1. Among them, R... 2 =0.71±0.10, MAE=0.06±0.01, RMSE=0.09±0.02. The calculation results basically prove the effectiveness of the model. Figure 2 The original Landsat NDVI and the reconstructed NDVI were compared. In the figure, black dots represent available NDVI sample values, lines represent the reconstructed NDVI time series, and light gray represents the 95% confidence interval. The results show that the reconstructed NDVI and the original NDVI are in good agreement. Furthermore, the seasonal and trend components of the Prophet model can effectively reflect the seasonal and interannual variations of NDVI, such as... Figure 3 The "trend" item shows the interannual variation characteristics of NDVI, while the "yearly" item shows the seasonal variation characteristics of NDVI.

[0054] Indirect method: The reconstructed Landsat NDVI data is evaluated using the original MODIS NDVI data to demonstrate the feasibility of using reconstructed NDVI data to characterize interannual variation trends. To reduce the impact of the uneven distribution of effective NDVI data across different seasons on the calculation of the annual average NDVI, the average NDVI for each season is first calculated using the effective NDVI data, and then the annual average NDVI is calculated by averaging the average NDVI of the four seasons. Figure 4 The images show a comparison between reconstructed Landsat NDVI data and MODIS satellite data, where (a) is a scatter plot of NDVI spatial pattern and (b) is a scatter plot of NDVI trend.

[0055] Figure 5 The images show a pixel-by-pixel comparison between the reconstructed Landsat NDVI and the observed MODIS NDVI, with (a)-(b) showing the NDVI spatial pattern comparison and (c)-(d) showing the NDVI trend comparison. The results indicate that the NDVI spatial patterns and interannual trends obtained by the two methods are very similar. Figure 6 This study compares the interannual variations of Nanjing's urban NDVI based on reconstructed Landsat NDVI and MODIS NDVI. The results show that the interannual variations and trends of the reconstructed NDVI are similar to those calculated using MODIS NDVI.

[0056] The accuracy assessment results of the reconstructed Landsat NDVI data are shown in Table 1:

[0057] Table 1. Accuracy assessment of Landsat NDVI data reconstructed pixel-by-pixel from Nanjing

[0058] average value Standard deviation Coefficient of determination 0.71 0.10 Mean Absolute Error 0.06 0.01 Root mean square error 0.09 0.02

Claims

1. A simple and robust method for long time series reconstruction of land satellite vegetation indices, characterized in that: The reconstruction method comprises: acquiring a data source of a land satellite, pre-processing the data source, the data source comprising Landsat NDVI data and MODIS NDVI data; Data optimization is performed on the Landsat NDVI data based on a 3σ criterion and a bilinear interpolation method to obtain first optimized data; A Prophet model is constructed based on the first optimized data to obtain annual, seasonal and residual components; Based on the Prophet model, NDVI reconstruction is performed on pixels marked as null to generate spatiotemporally continuous clear-sky Landsat NDVI data; The specific steps of data optimization on the Landsat NDVI data based on the 3σ criterion and the bilinear interpolation method are as follows: S1. Acquire Landsat NDVI data, remove outliers in the Landsat NDVI time series data to obtain outlier-removed Landsat NDVI data; S2. Perform spatial resolution reconstruction on the outlier-removed Landsat NDVI data based on the bilinear interpolation method to obtain reconstructed Landsat NDVI data; S3. Perform accuracy evaluation on the reconstructed Landsat NDVI data, if the spatial resolution of the reconstructed Landsat NDVI data meets a set threshold, the reconstructed Landsat NDVI data is first optimized data; The Prophet model calculation formula is as follows: ; where, y t is the observation value at time, t g is the interannual variation component, t s is the seasonal variation component, t is the residual component, k and are the growth rate and the correlation adjustment coefficient, respectively, m is the compensation parameter; where S j is the time of the appearance of the j j= 1, 2, …, S inflection points ， is the rate of change at time, , a n and b n are the coefficients of the Fourier series, P is the period.​​​​​​​ 2. The simple and robust long time series reconstruction method of land satellite vegetation index according to claim 1, characterized in that: The pre-processing of the data source specifically comprises: The atmospheric influence coefficient DA in the Landsat NDVI data is ≥ N The pixels are masked, and the atmospheric effect DA < N Atmospheric correction is performed on the pixels to obtain the corrected pixels; The corrected pixels are subjected to cloud pollution and strip difference removal processing to complete the pre-processing of the data source.

3. The simple and robust long time series reconstruction method of land satellite vegetation index according to claim 1, characterized in that: The accuracy evaluation of the reconstructed Landsat NDVI data in step S3 comprises a direct method and an indirect method.

Images