Landsat NDVI time series image reconstruction method based on spatiotemporal fusion
By using clear pixel recovery in Landsat NDVI data with cloud volume ≤75% in the space-time fusion method, and combining the basic data, iterative filtering of Landsat NDVI time series data is solved, and the problem of low fusion accuracy of spatiotemporal data is improved. continuity and accuracy.
Patent Information
- Application Number
- CN202111418799.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-26
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-11-26
AI Technical Summary
The existing spatiotemporal fusion method ignores the clear pixels in some cloud-polluted images when reconstructing Landsat NDVI time series data, resulting in low fusion accuracy of spatiotemporal data. Due to the differences in imaging geometric conditions and band spectral response functions between MODIS and Landsat , affects the accuracy of the data.
By selecting Landsat NDVI data with a cloud volume of ≤75% as auxiliary data, the clear pixels with the highest correlation with the predicted image are restored, and combined with the basic data, the cloud pixels are restored to generate the initial Landsat NDVI fusion time data, and then The iterative SG filtering algorithm is used to smooth the time series and adjust the weight value to improve data accuracy.
Improve the accuracy of high-spatial-temporal resolution NDVI long-time series data, make full use of clear pixels in some cloud-polluted images, and enhance the continuity and accuracy of the data.
Smart Images

Figure CN114240773B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a Landsat NDVI time series image reconstruction method based on time-space fusion. Background Art
[0002] Normalized Difference Vegetation Index (NDVI) time series satellite data have been widely used for terrestrial ecosystem vegetation monitoring in the past few decades. At present, the published NDVI time series data products are mainly provided by satellites such as MODIS, SPOT VGT and AVHRR. The spatial resolution of these NDVI products is relatively coarse (from hundreds of meters to several kilometers), and there are great limitations when they are applied in areas with surface heterogeneity. There is a great demand for higher resolution satellite NDVI time series data products in the industry. Landsat satellites have rich historical data and a spatial resolution of 30m, which has great potential. However, the temporal resolution of Landsat data is relatively low (16-day re-entry period). At the same time, due to the interference of cloud pollution on optical images, Landsat NDVI data is seriously missing and the time series is quite discontinuous. Reconstructing continuous Landsat NDVI time series data is a technology that needs to be developed.
[0003] Spatiotemporal fusion technology is considered to be able to reconstruct Landsat NDVI time series data. This technology mainly generates remote sensing data with both high spatial resolution and high temporal resolution by fusing Landsat data with high spatial and low temporal resolution with remote sensing data with low spatial and high temporal resolution (such as MODIS data). Currently representative spatiotemporal fusion algorithms include STARFM, FSDAF, and IFSDAF. In order to generate a Landsat image for a certain predicted date, these spatiotemporal fusion methods usually use a pair of MODIS and Landsat images acquired on a certain date (called the base date), as well as a MODIS image acquired on the predicted date. However, for the generation of Landsat NDVI time series data, existing spatiotemporal fusion methods usually ignore the effective use of some cloud pollution images.
[0004] Landsat images with partial cloud pollution are widely available. Previous methods either ignored these cloud-polluted images and reconstructed them using spatiotemporal fusion algorithms, or replaced the clear pixels in these partially cloud-polluted images with the corresponding pixels in the fused image. Such fusion methods do not fully utilize the clear pixels in partially cloud-polluted images. Existing studies have shown that due to differences between MODIS and Landsat in imaging geometry, band spectral response functions and other factors, the inconsistency between the two will seriously affect the accuracy of spatiotemporal data fusion. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides a Landsat NDVI time series image reconstruction method based on spatiotemporal fusion, which solves the technical problem of low precision of spatiotemporal data fusion caused by directly using MODIS NDVI data and fusing it in the existing reconstruction of Landsat NDVI time series data.
[0006] A Landsat NDVI time series image reconstruction method based on spatiotemporal fusion according to an embodiment of the present invention includes:
[0007] S1. Determine a data source and perform image preprocessing on the data source: the data source includes MODIS NDVI data and Landsat NDVI data, use the part of the Landsat NDVI data with cloud cover ≤75% as auxiliary data, select a completely clear image found in the auxiliary data as basic data, and generate surface category data of the area of the basic data, and obtain MODIS NDVI synthetic data after filtering the MODIS NDVI data through SG;
[0008] S2, determine the forecast date, then perform spatiotemporal fusion processing on the regional surface category data and MODIS NDVI synthetic data to generate a high spatiotemporal resolution forecast image for each forecast date;
[0009] S3. Perform Landsat NDVI fusion time series: including:
[0010] a. When there is no clear image in the predicted image, select the first few MODIS NDVI synthetic data with the highest correlation coefficient with the predicted image, use the corresponding auxiliary data to restore the corresponding clear pixels, and then sort the predicted images by the predicted date to obtain the initial Landsat NDVI fusion time data;
[0011] b. When there is a clear image in the predicted image, the first few scenes of data most relevant to the clear part of the predicted image are selected from the auxiliary data, and the clear pixels in these scenes are used in turn, together with the basic data as the auxiliary data of the last scene, to restore the cloud pixels on the predicted image, and then the predicted images are sorted by the predicted date to obtain the initial Landsat NDVI fusion time data;
[0012] S4, iterative SG filtering: Iterative SG filtering is used to smooth the fused time series, wherein the weight value of the clear pixel in the iterative SG filtering algorithm is determined by whether the clear pixel is obtained through step a or step b in step S3. If the clear pixel is obtained through step a, the weight coefficient is halved, so that the result of SG filtering is closer to the clear pixel obtained through step b, and high temporal and spatial resolution NDVI long time series data is obtained.
[0013] The technical principle of the present invention is: through the clear pixels of the clear part with the highest correlation coefficient with the predicted image in the auxiliary data with partial cloud pollution, and adding the basic data as the auxiliary data of the last scene, the cloud pixels on the predicted image are restored to obtain a clearer predicted image, and then the initial Landsat NDVI fused time data is obtained. When the initial Landsat NDVI fused time data is iteratively SG filtered, the weight value is judged for whether the auxiliary data and the basic data have been restored, that is, if yes, it is recorded as 1, and if no, it is recorded as 0.5, so that the final high temporal and spatial resolution NDVI long time series data results are more inclined to use the predicted image restored by the auxiliary data and the basic data.
[0014] Compared with the prior art, the present invention has the following beneficial effects: the predicted image is restored by auxiliary data with partial cloud pollution, and then the initial Landsat NDVI fused time data is obtained, which solves the technical problem of low precision of spatiotemporal data fusion that occurs when MODIS NDVI data is directly used to fuse with the existing reconstructed Landsat NDVI time series data, and improves the precision of the final high spatiotemporal resolution NDVI long time series data.
[0015] Furthermore, the cloud cover of the basic data in step S1 is ≤10%.
[0016] Furthermore, in the step S1, the data source is collected in units of years, and the projection coordinates of the collected MODIS NDVI data and Landsat NDVI data are consistent. The Landsat NDVI data needs to collect data for at least one year, and the MODIS NDVI data needs to collect data for at least 23 scenes in that year.
[0017] Furthermore, the basic data in step S1 is used to generate surface category data of the region through the ISODATA clustering algorithm.
[0018] Furthermore, the spatiotemporal fusion processing in step S2 adopts the FSDAF method.
[0019] Furthermore, in step a and b of step S3, the base date needs to be determined, and then the increment ΔF of the MODIS NDVI synthetic data between the prediction date and the base date needs to be obtained. After that, the value F0 of the MODIS NDVI synthetic data on the prediction date is equal to the sum of the value F0 of the MODIS NDVI synthetic data on the base date and the increment ΔF. The formula is: F β =F0+ΔF, used for pixel restoration in steps a and b;
[0020] It is also necessary to use the thin plate spline interpolation method to interpolate the MODIS NDVI data of the forecast date and the benchmark date, and obtain the interpolation results of the MODIS NDVI synthetic data of the forecast date and the benchmark date. and The spatial correlation ΔS(x j ,y j ), the formula is: Used for pixel restoration in step a.
[0021] Furthermore, in step b, the auxiliary data is searched for the first four scenes most relevant to the predicted image, and the basic data obtained in step S1 are used to perform a class-by-class linear regression of the regional surface category data and the MODIS NDVI synthetic data.
[0022] Furthermore, the linear regression method includes the formula:
[0023] ΔF clecr =F β (clear)-F0(clera);
[0024]
[0025]
[0026] Where ΔF clear and ΔS TPS clear The relationship is: ΔF clear =a*ΔS TPS clear +b;
[0027] ΔF cloud and ΔS TPS cloud The relationship is: ΔF cloud =a*ΔS TPS cloud +b;
[0028] a and b are regression coefficients, estimated from clear pixels in some cloud-polluted images;
[0029] Prediction result F0 pass: F β =F0+(ΔF cloud ∪ΔF cloud )get.
[0030] Furthermore, in step a, the increment of MODIS TPS interpolation is directly used to make corresponding predictions, and the prediction result F0 is obtained by: β =F0+ΔS is obtained. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 The figure is an algorithm flow chart of a Landsat NDVI time series image reconstruction method based on spatiotemporal fusion according to an embodiment of the present invention.
[0032] Figure 2 Landsat images of three test areas according to an embodiment of the present invention.
[0033] Figure 3 It is a bar graph comparing the root mean square errors of the method according to the embodiment of the present invention and the IFSDAF method.
[0034] Figure 4 It is a bar graph comparing the image structure similarity indexes of the method according to the embodiment of the present invention and the IFSDAF method.
[0035] Figure 5 This is a visual comparison diagram of the images generated by the fusion of deciduous broad-leaved forests according to an embodiment of the present invention.
[0036] Figure 6 A visual comparison diagram of the images generated by the irrigated farmland fusion according to an embodiment of the present invention.
[0037] Figure 7 A visual comparison of images generated by evergreen forest fusion according to an embodiment of the present invention.
[0038] Figure 8 The following is a bar chart comparing the root mean square error between the method according to the embodiment of the present invention and the IFSDAF method in a deciduous broad-leaved forest under different cloud coverage ratios in the predicted image.
[0039] Fig. 9 A columnar comparison chart of the image structure similarity index of the method of the embodiment of the present invention and the IFSDAF method in a deciduous broad-leaved forest under different cloud coverage ratios of the predicted image.
[0040] Fig.10 A bar chart comparing the root mean square error between the method of the present invention and the IFSDAF method in predicting irrigated farmland under different cloud coverage ratios of the image.
[0041] Fig.11A bar graph comparing the image structure similarity index of the method of the present invention and the IFSDAF method in predicting irrigated farmland under different cloud coverage ratios of the image.
[0042] Fig.12 A bar chart comparing the root mean square errors of the method according to the embodiment of the present invention and the IFSDAF method in predicting evergreen forests under different cloud coverage ratios of the image.
[0043] Fig.13 A bar graph comparing the image structure similarity index of the method of the embodiment of the present invention and the IFSDAF method in predicting images with different cloud coverage ratios in evergreen forests.
[0044] Fig.14 The deciduous broad-leaved forest is a line comparison chart of Landsat NDVI time series data generated for one year by the method of the embodiment of the present invention and the IFSDAF method.
[0045] Fig.15 Irrigated farmland is a line comparison chart of Landsat NDVI time series data generated for one year by the method of the embodiment of the present invention and the IFSDAF method.
[0046] Fig.16 Changlulin is a line comparison chart of Landsat NDVI time series data generated for one year by the method of the embodiment of the present invention and the IFSDAF method. DETAILED DESCRIPTION
[0047] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0048] like Figure 1 The Landsat NDVI time series image reconstruction method based on time-space fusion shown includes:
[0049] S1. Determine the data source and perform image preprocessing on the data source: the data source includes MODIS NDVI data and Landsat NDVI data; Landsat NDVI data also contains its corresponding data quality band data, and then select auxiliary data according to the data quality; specifically, the part of Landsat NDVI data with cloud cover ≤75% is used as auxiliary data, and a completely clear image found in the auxiliary data is selected as the basic data. The cloud cover of the basic data needs to be ≤10%, and the basic data is used to generate the surface category data of the region through the ISODATA clustering algorithm, and the MODIS NDVI data is filtered through the SG to obtain the MODIS NDVI synthetic data.
[0050] The specific data source is collected on an annual basis. The projection coordinates of the collected MODIS NDVI data and Landsat NDVI data are kept consistent so that the grids of MODIS NDVI data and Landsat NDVI data correspond to each other. Landsat NDVI data needs to collect data for at least one year, while MODIS NDVI data needs to collect data for at least 23 scenes in that year. The 23 scenes of MODIS NDVI data are synthesized into 23 bands of data, and then the MODIS NDVI synthetic data is obtained after SG filtering.
[0051] S2. Determine the forecast date, then perform spatiotemporal fusion processing on the regional surface category data and MODIS NDVI synthetic data to generate a high spatiotemporal resolution forecast image for each forecast date. The specific spatiotemporal fusion processing adopts the FSDAF method.
[0052] S3. Perform Landsat NDVI fusion time series: including:
[0053] First, we need to determine the base date, and then we need to get the increment of the MODIS NDVI composite data between the forecast date and the base date. Then, the value of the MODIS NDVI composite data on the forecast date is equal to the sum of the value of the MODIS NDVI composite data on the base date and the increment. The formula is: β =F0+ΔF, used for pixel restoration in steps a and b.
[0054] It is also necessary to use the thin plate spline interpolation method to interpolate the MODIS NDVI data of the forecast date and the benchmark date, and obtain the interpolation results of the MODIS NDVI data of the forecast date and the benchmark date. and The spatial correlation ΔS(x j ,y j ), the formula is: Used for pixel restoration in step a
[0055] a. When there is no clear image in the predicted image, the increment ΔS of MODIS TPS interpolation is directly used for corresponding prediction. The MODIS NDVI synthetic data of the top five scenes with the highest correlation coefficient with the predicted image are selected, and the corresponding auxiliary data are used to restore the corresponding clear pixels. The prediction result is obtained by: F β =F0+ΔS, and then the predicted images are sorted by the predicted date to obtain the initial Landsat NDVI fusion time data.
[0056] b. When there is a clear image in the predicted image, the first four scenes of data most relevant to the clear part of the predicted image are selected from the auxiliary data, and the clear pixels in these four scenes are used in turn, together with the basic data as the auxiliary data of the last scene, to perform a class-by-class linear regression of the regional surface category data and the MODIS NDVI synthetic data, restore the cloud pixels on the predicted image, and then sort the predicted images by the predicted date to obtain the initial Landsat NDVI fusion time data.
[0057] The specific linear regression method includes the formula:
[0058] ΔF clear =F β (clear)-F0(clear);
[0059]
[0060]
[0061] where ΔF clear and ΔS TPS clear The relationship is: ΔF clear =a*ΔS TPS clear +b;
[0062] ΔF cloud and ΔS TPS cloud The relationship is: ΔF cloud =a*ΔS TPS cloud +b;
[0063] a and b are regression coefficients, estimated from clear pixels in some cloud-polluted images;
[0064] Prediction result F0 pass: F β =F0+(ΔF cloud ∪ΔF clear )get.
[0065] S4, iterative SG filtering: Iterative SG filtering is used to smooth the fused time series, wherein the weight value of the clear pixel in the iterative SG filtering algorithm is determined by whether the clear pixel is obtained through step a or step b in step S3. If the clear pixel is obtained through step a, the weight coefficient is halved, so that the result of SG filtering is closer to the clear pixel obtained through step b, and high temporal and spatial resolution NDVI long time series data is obtained.
[0066] like Figure 2As shown in the figure, we tested the Landsat NDVI time series image reconstruction method based on spatiotemporal fusion in three typical areas and compared it with the advanced algorithm IFSDAF developed in 2019. The three test areas have different vegetation cover types:
[0067] The first area is the deciduous broad-leaved forest (Shenlongjia area, Hubei);
[0068] The second region is irrigated farmland (New South Wales, Australia) with high surface heterogeneity;
[0069] The third region is the evergreen forests in Southeast Asia, which is cloudy and foggy, with fewer cloud-free images in a year;
[0070] Landsat NDVI and MODIS NDVI data were collected for one year in three regions, see Tables 1 and 2:
[0071] Table 1. Landsat data list
[0072] Experimental Area date Row and column number Image size Data Source Quantity (pieces) Deciduous forest (SNJ) 2015 / 01 / 01-2015 / 12 / 31 125 / 38 576×552 Landsat 8 9 Farmland (CIA) 2011 / 06 / 01-2012 / 06 / 01 93 / 84 800×800 Landsat 7 17 Evergreen forest (SEA) 2019 / 01 / 01-2019 / 12 / 31 127 / 45 800×800 Landsat 8 6
[0073] Table 2. MODIS data list
[0074] Experimental Area date Image size Data Source Quantity (pieces) Deciduous forest (SNJ) 2015 / 01 / 01-2015 / 12 / 31 72×69 MOD13Q1 23 Farmland (CIA) 2011 / 06 / 01-2012 / 06 / 01 100×100 MOD13Q1 23 Evergreen forest (SEA) 2019 / 01 / 01-2019 / 12 / 31 100×100 MOD13Q1 23
[0075] The following three experiments are provided to evaluate the effectiveness of the Landsat NDVI time series image reconstruction method based on spatiotemporal fusion of the present invention.
[0076] The quantitative evaluation is measured by the root mean square error (RMSE) and the image structure similarity index (SSIM). The specific RMSE is mainly used to compare the deviation of the values of the real image and the fused NDVI image, while the SSIM is mainly used to quantify the spatial similarity between the real image and the fused NDVI image. The smaller the RMSE, the larger the SSIM, and the higher the accuracy of the method.
[0077] The design and implementation results of each experiment are as follows:
[0078] Experiment 1
[0079] In each test area, a cloudless and clear Landsat NDVI image is randomly selected. It is assumed that some clear pixels on the image are missing (i.e., simulated clouds). Then, the missing pixels are restored using the method of the present invention and the IFSDAF method, and the restoration accuracy is quantitatively evaluated as follows:
[0080] 1. SNJ56: deciduous broad-leaved forest (Shenlongjia area, Hubei Province), using DOY287 (20151014) to predict DOY344 (20151210), with a time interval of 56 days, the simulated cloud cover is 16.9%;
[0081] 2. CIA47: irrigated farmland (New South Wales, Australia), using DOY328 (20011124) to predict DOY011 (20020111), 47 days later, the simulated cloud cover is 28.6%.
[0082] 3. SEA31: Evergreen forests in Southeast Asia, using DOY241 (20190829) to predict DOY273 (20190930), with a time interval of 31 days and a simulated cloud cover of 16.4%.
[0083] 4. SEA64: Evergreen forests in Southeast Asia, using DOY273 (20190930) to predict DOY337 (20191203), with a time interval of 64 days and a simulated cloud cover of 16.4%.
[0084] 5. SEA95: Evergreen forests in Southeast Asia, using DOY241 (20190829) to predict DOY337 (20191203), with a time interval of 95 days and a simulated cloud cover of 16.4%.
[0085] like Figure 3-4 As shown, compared with the IFSDAF method, the new method of the present invention obtains a smaller RMSE value and a larger SSIM value in all tests.
[0086] like Figure 5-7 As shown in the figure, compared with the IFSDAF method, the method of the present invention is more similar to the real image in space and can better preserve the spatial details of the image. For example, in the irrigated agricultural area of Australia, the method of the present invention better preserves the field patches, while IFSDAF has obvious boundary blurring. These results show that the new method has better reconstruction effect for single-view Landsat NDVI images.
[0087] Experiment 2
[0088] In this experiment, we tested the effect of different cloud coverage ratios on the reconstruction effect of the method on some cloud-contaminated images. In the clear image, we simulated cloud coverage ratios of 20%, 40%, 60%, 80%, and 100%, where 100% cloud coverage means that there are no clear pixels available in this image.
[0089] The experimental results are as follows Figure 8-13As shown in the figure, the method of the present invention achieves smaller RMSE values and larger SSIM values under different cloud coverage ratios in all three test areas; when the cloud coverage ratio increases, although the accuracy of the new method decreases to a certain extent, the accuracy is still higher than that of IFSDAF. These experimental results show that the new method has a wide range of applicability and can achieve better reconstruction effects for Landsat NDVI images with different degrees of cloud pollution, including polluted images with no clear pixels available.
[0090] (3) Experiment III:
[0091] In this experiment, we compare the effects of the method of the present invention and the IFSDAF method in reconstructing a one-year continuous Landsat NDVI time series. Similarly, we also assume that a clear image of a certain phase is missing in each region.
[0092] The experimental results are as follows Figure 14-16 As shown, the NDVI time series data restored by the method of the present invention is closer to the true value verification point, while the time series reconstructed by IFSDAF deviates further from the verification point.
[0093] In addition, at the deciduous forest site, the time series curve restored by IFSDAF showed obvious differences at both ends, while the curve restored by the new method was more reasonable.
[0094] In summary, whether from the perspective of visual effect or quantitative evaluation, the Landsat NDVI time series data restored by the method of the present invention is better.
[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A Landsat NDVI time series image reconstruction method based on spatiotemporal fusion, characterized by: include: S1. Determine a data source and perform image preprocessing on the data source: the data source includes MODIS NDVI data and Landsat NDVI data, use the part of the Landsat NDVI data with cloud cover ≤75% as auxiliary data, select a completely clear image found in the auxiliary data as basic data, and generate surface category data of the area of the basic data, and obtain MODIS NDVI synthetic data after filtering the MODIS NDVI data through SG; S2, determine the forecast date, then perform spatiotemporal fusion processing on the regional surface category data and MODIS NDVI synthetic data to generate a high spatiotemporal resolution forecast image for each forecast date; S3. Perform Landsat NDVI fusion time series: including: a. When there is no clear image in the predicted image, select the first few MODIS NDVI synthetic data with the highest correlation coefficient with the predicted image, use the corresponding auxiliary data to restore the corresponding clear pixels, and then sort the predicted images by the predicted date to obtain the initial Landsat NDVI fusion time data; b. When there is a clear image in the predicted image, the first few scenes of data most relevant to the clear part of the predicted image are selected from the auxiliary data, and the clear pixels in these scenes are used in turn, together with the basic data as the auxiliary data of the last scene, to restore the cloud pixels on the predicted image, and then the predicted images are sorted by the predicted date to obtain the initial Landsat NDVI fusion time data; S4, iterative SG filtering: Iterative SG filtering is used to smooth the fused time series, wherein the weight value of the clear pixel in the iterative SG filtering algorithm is determined by whether the clear pixel is obtained through step a or step b in step S3. If the clear pixel is obtained through step a, the weight coefficient is halved, so that the result of SG filtering is closer to the clear pixel obtained through step b, and high temporal and spatial resolution NDVI long time series data is obtained.
2. The Landsat NDVI time series image reconstruction method based on time-space fusion as claimed in claim 1, characterized in that: The cloudiness of the basic data in step S1 is ≤10%.
3. The Landsat NDVI time series image reconstruction method based on time-space fusion as claimed in claim 1, characterized in that: In the step S1, the data source is collected in units of years, and the projection coordinates of the collected MODIS NDVI data and Landsat NDVI data are consistent. The Landsat NDVI data needs to collect data for at least one year, and the MODIS NDVI data needs to collect data for at least 23 scenes in that year.
4. The Landsat NDVI time series image reconstruction method based on time-space fusion as claimed in claim 1, characterized in that: The basic data in step S1 generates regional surface category data through the ISODATA clustering algorithm.
5. The Landsat NDVI time series image reconstruction method based on time-space fusion as claimed in claim 1, characterized in that: The spatiotemporal fusion process in step S2 adopts the FSDAF method.
6. The Landsat NDVI time series image reconstruction method based on time-space fusion according to claim 1, characterized in that: In step a and b of step S3, the reference date needs to be determined, and then the increment ΔF of the MODIS NDVI composite data between the predicted date and the reference date needs to be obtained, and then the value F of the MODIS NDVI composite data on the predicted date needs to be obtained. p It is equal to the sum of the value F0 of the MODIS NDVI composite data on the base date and the increment ΔF. The formula is: F p =F0+ΔF, used for pixel restoration in steps a and b; It is also necessary to use the thin plate spline interpolation method to interpolate the MODIS NDVI data of the forecast date and the benchmark date, and obtain the interpolation results of the MODIS NDVI synthetic data of the forecast date and the benchmark date. and The spatial correlation ΔS(x j ,y f ), the formula is: Used for pixel restoration in step a.
7. The Landsat NDVI time series image reconstruction method based on time-space fusion according to claim 6, characterized in that: In the step b, the auxiliary data is searched for the first four scene data most relevant to the predicted image, and the basic data obtained in the step S1 are used to perform a class-by-class linear regression of the regional surface category data and the MODIS NDVI synthetic data.
8. The Landsat NDVI time series image reconstruction method based on time-space fusion according to claim 7, characterized in that: The linear regression method includes the formula: ΔF clear =F p (clear)-F0(clear); where ΔF clear and ΔS TPS clear The relationship is: ΔF clear =a*ΔS TPS clear +b; ΔF closed and ΔS TPS closed The relationship is: ΔF closed =a*ΔS TPS closed +b; a and b are regression coefficients, estimated from clear pixels in some cloud-polluted images; Prediction result F o By: F p =F0+(ΔF closed ∪ΔF closer )get.
9. The Landsat NDVI time series image reconstruction method based on time-space fusion according to claim 6, characterized in that: In step a, the increment of MODIS TPS interpolation is directly used to make corresponding predictions. The prediction result F o By: F p =F0+ΔS is obtained.
