Vegetation index space-time fusion method considering farmland space-time heterogeneity

Abstract

The invention provides a vegetation index spatio-temporal fusion method considering farmland spatio-temporal heterogeneity, which comprises the following steps: simultaneously acquiring high-resolution (fine-resolution) and low-resolution (coarse-resolution) satellite remote sensing data of a target area, preprocessing the satellite remote sensing data, carrying out field parcel segmentation on the high-resolution remote sensing data, and carrying out field parcel segmentation on the low-resolution remote sensing data; optimizing a selection strategy of similar pixels by taking the field parcel as a unit; vegetation index time series data of the high-quality coarse-resolution image are constructed through data reconstruction; fitting a crop vegetation index growth curve of the fine-resolution image by adopting a Double Logistic function, and generating a fine-resolution pixel initial value at a prediction moment; constructing an optimization model and optimizing a fine pixel vegetation index value in combination with coarse and fine resolution pixel data; weighted distribution is carried out based on the spatial consistency in the field, and a prediction image is optimized in combination with spatial filtering. According to the method, the problems of unreasonable assumption of a traditional space-time fusion model on crop spectrum linear growth in a high-heterogeneity farmland scene and inaccurate similar pixel matching are effectively solved, the fusion precision and robustness are improved, and the method has a wide application prospect.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing and spatiotemporal data processing technology, and more specifically to a spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland. Background Technology

[0002] High spatial and temporal resolution data are crucial for crop growth monitoring and precision agriculture management. However, due to limitations in satellite sensor imaging systems and the influence of cloud and rain weather, remote sensing images are mutually restrictive in terms of temporal and spatial resolution, making it impossible to achieve both simultaneously. Spatiotemporal fusion algorithms combine the advantages of these two types of satellite sensors to generate time-series remote sensing data with both high temporal and spatial resolution, providing an effective means to solve this problem. Currently, most spatiotemporal fusion algorithms lack consideration for the heterogeneity of spectral spatiotemporal variations and the nonlinear changes in spectral reflectance under the same land cover type. Farmland areas exhibit significant spatiotemporal heterogeneity due to differences in crop planting types and phenological stages. Existing methods are difficult to adapt to spatiotemporal data fusion in complex and heterogeneous farmland scenarios. Traditional methods rely on the assumption that the spectral reflectance of land cover changes linearly over time, which may lead to large errors during nonlinear crop growth periods, such as inflection points during key phenological transitions. In addition, current methods often rely on fixed windows to calculate similar pixels, making it difficult to accurately represent the consistency characteristics of farmland boundaries and plots, resulting in increased similar pixel matching errors and reduced fusion accuracy. Therefore, there is an urgent need to construct a spatiotemporal data fusion method that can take into account the spatiotemporal heterogeneity of farmland in order to improve the quality of spatiotemporal fusion in farmland scenarios. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland, and to improve upon the problems of poor adaptability, inaccurate selection of similar pixels, and low fusion accuracy of traditional methods in complex farmland scenarios. This invention proposes the Fit-STVIF (Fitting Spatial and Temporal Vegetation Index Fusion) model, which, based on farmland plot segmentation, uses the segmented plots as window units to optimize the selection strategy of similar pixels. Simultaneously, it innovatively introduces the idea of ​​fitting the trajectory of crop vegetation index growth curves, fully considering the nonlinear variation characteristics of vegetation index growth curves in the solution of fine pixels, thereby improving the accuracy of spatiotemporal data fusion.

[0004] Firstly, this invention provides a spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland, including:

[0005] 1) Preprocessing, field segmentation and similar pixel selection of coarse resolution (low resolution, high temporal sequence, coarse pixels) and fine resolution (high resolution, low temporal sequence, fine pixels) image data. Based on the fine resolution image, the farmland is segmented. Based on the segmented fields, each field is used as a window to search for pixels similar to the center pixel.

[0006] 2) Reconstruction of coarse-resolution time-series data: Time-series reconstruction of coarse-resolution remote sensing images is performed to remove outliers and fill in missing data, generating high-quality time-series vegetation index data to fill in missing values ​​and remove outliers, ensuring data integrity and accuracy.

[0007] 3) Reconstruction of fine-resolution (medium-high resolution, low time series, fine pixel) vegetation index time series data: Double Logistic fitting is used to fit all available fine-resolution data, describe its time series change trend, and generate initial values ​​of fine pixels for the prediction date.

[0008] 4) Combining the initial values ​​of the fitted fine pixels with their relationship with the coarse pixels, optimize the calculation of the fine pixel vegetation index value for the prediction time.

[0009] 5) Assign fusion weights based on the spectral similarity and spatial distance of pixels within the field area, and perform spatial filtering to improve image quality and spatial continuity;

[0010] 6) With R 2 The root mean square error (RMSE) was used as an evaluation metric to verify the progress of the Fit-STVIF model and to compare its accuracy with other classic models.

[0011] Furthermore, the similar pixel selection based on field segmentation includes: segmenting the study area into fields using an image segmentation algorithm, and further refining and optimizing the boundaries to ensure the accuracy and consistency of the segmentation results; using the segmented fields as a window for searching similar pixels. Similar pixels within the window provide specific temporal and spatial information to calculate the WDRVI value of the center pixel.

[0012] Furthermore, the coarse pixel time-series data reconstruction includes: correcting time deviations and selecting images from high-quality observation points through multi-temporal data complementation; using SG filtering to smooth the time-series data to generate high-quality image data with high temporal resolution; and effectively filling data gaps and ensuring temporal continuity during cloud pollution periods through synthetic data trend constraints and interpolation algorithms. The reconstructed WDRVI time-series images combine high temporal resolution, low noise, and temporal consistency, providing a more reliable data foundation for vegetation dynamic monitoring.

[0013] Furthermore, preliminary fitting of fine-pixel time-series data includes: introducing the Double Logistic function for fitting, combining all available fine-resolution time-series vegetation indices, and fitting the initial values ​​of the average vegetation growth curves for similar pixels.

[0014] Furthermore, the optimization of the fine pixel growth curve includes: combining the fitted initial vegetation index curve of the fine pixel, using the relationship between coarse pixels and fine pixels as a constraint, and optimizing the vegetation index growth curve of the fine pixel by changing the fitting parameters of the fine pixel curve.

[0015] Furthermore, pixel weighting and spatial filtering are employed. This involves calculating the proportion of each similar pixel in the preceding and following reference images, and then weighting these proportions based on the difference between the coarse pixels of the predicted date and the coarse pixels of the preceding and following reference dates. Additionally, the spectral difference between the center pixel and its neighboring pixels is calculated, and the top n pixels with the smallest spectral differences are selected from the local window. These n pixels are then weighted according to spatial distance to generate the final predicted value. Pixels that are closer together have higher weights to preserve local spatial structure and improve prediction accuracy.

[0016] Furthermore, the Fit-STVIF model was validated, including: using R... 2 The root mean square error (RMSE) is used as an evaluation metric to compare the actual image with the predicted image.

[0017] The beneficial effects of adopting the above scheme are as follows:

[0018] 1) It performs better in terms of spatial continuity of predicted images, preservation of field boundaries and control of outliers, demonstrating strong generalization ability and stability. It performs better in field boundary preservation and spatial detail restoration, and achieves excellent prediction accuracy in different heterogeneous farmland scenarios.

[0019] 2) Fully consider the heterogeneity of spectral spatiotemporal changes in cultivated land, and express the spatiotemporal changes of plot data for different crop types or different growth stages in a differentiated manner to improve the progress of data spatiotemporal fusion.

[0020] 3) The Fit-STVIF model innovatively introduces the idea of ​​fitting crop vegetation index growth curve trajectories, fully considering the nonlinear variation characteristics of vegetation index growth curves in the solution of fine pixels, thereby improving the accuracy of spatiotemporal fusion of crop growth curve inflection point data. By introducing the idea of ​​fitting crop vegetation index growth curves and a field-based similar pixel selection mechanism, the Fit-STVIF model better reconstructs the nonlinear growth trajectory of crop vegetation index changes.

[0021] 4) Based on the segmentation of farmland plots, the segmented plots are used as window units to optimize the selection strategy of similar pixels, improve the progress of data fusion and the ability to restore the details of plot boundaries. Attached Figure Description

[0022] Figure 1 This is a basic flowchart of the present invention;

[0023] Figure 2 This is a detailed flowchart of the present invention;

[0024] Figure 3 This is a graph showing the accuracy comparison results. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings.

[0026] According to an embodiment of the present invention, an example of a spatiotemporal fusion method for vegetation index that takes into account the spatiotemporal heterogeneity of farmland is provided.

[0027] like Figure 1 and Figure 2 As shown, the method includes the following steps:

[0028] Step 1: Data preprocessing. Calculate the WDRVI of the target area in the remote sensing image and perform field segmentation. Based on the segmented fields, use each field as a window to search for pixels similar to the center pixel within the field area.

[0029] Step 2: Data reconstruction of coarse-resolution images. By complementing multi-temporal data, time deviations are corrected and high-quality observation data are selected. SG filtering is used to smooth the temporal data and generate high-temporal, high-quality image data.

[0030] Step 3: Fitting and optimizing fine-resolution data. Use Double Logistic to fit all available fine-resolution images to initially describe their time series change trends. Generate the average vegetation index curve of similar pixels as the initial value. Based on this, combine the pixel relationship with the coarse-resolution image to optimize the calculation of the fine-resolution vegetation index value at the prediction time.

[0031] Step 4, Pixel weighting and spatial filtering: Weights are assigned to similar pixels within the field area, and spatial filtering is applied to process the generated fine-resolution image to eliminate residual noise.

[0032] This invention is implemented using maize-growing areas in the Guanzhong Plain of China, the middle and lower reaches of the Yangtze River, and Iowa, USA as examples. The coarse spatial resolution data used is data from the Moderate-resolution Imaging Spectroradiometer (MODIS) sensor. Band 1 (620nm-670nm) and Band 2 (841nm-876nm) of the daily and 8-day synthesized surface reflectance products MOD90Q1 from MODIS are used to construct the WDRVI spectral index. Then, the DAVIR-MUTCOP method is used to fuse and reconstruct the 8-day synthesized MODIS data and the daily data to generate high-quality daily MODIS vegetation index time-series data. The fine spatial resolution data used were the B4 and B8 bands acquired from products of two Sentinel-2 satellites (Sentinel-2A and Sentinel-2B), which are the red and near-red light bands, respectively. Multiple scenes of fine spatial resolution data were segmented, and after optimization, the final field segmentation results were obtained. Remote sensing image data from MODIS and Sentinel-2 were obtained by cropping vector files of the study area, and WDRVI was calculated. The images predicted in this invention's example are images of key phenological stages under different heterogeneous farmland scenarios, including the seedling stage, peak growth stage, and harvest stage.

[0033] In this invention, data preprocessing is first performed to calculate the WDRVI of the target area in the remote sensing image, followed by field segmentation, selection of similar pixels, and data reconstruction of the coarse-resolution image. Based on the segmented fields, each field is used as a window to search for pixels similar to the center pixel within the field area. Then, a fitting solution is obtained by using Double Logistic regression to fit all available fine-resolution images, describing their time-series variation trends, generating an initial value of the average vegetation index of similar pixels in the fine-resolution image for the prediction date, and then combining this with the pixel relationship of the coarse-resolution image to optimize the similar pixel values ​​for the prediction time. Finally, pixel allocation is performed, assigning weights to similar pixels within the field area, and spatial filtering is applied to process the generated fine-resolution image to eliminate residual noise.

[0034] In this embodiment of the invention, step 1 mainly includes the following parts:

[0035] 1. Calculate the WDRVI of the target region pixels according to the WDRVI calculation formula; the formula is:

[0036] Scaled_WDRVI={[(α-1)+(α+1)×NDVI] / [(α+1)+(α-1)×NDVI]+(1-α) / (1+α)}×100

[0037] NDVI=(ρ nir -ρ red ) / (ρ nir +ρ red ), α is 0.1.

[0038] 2. Obtain coarse and fine spatial resolution data for the target area;

[0039] 3. Segmenting land parcels using fine spatial resolution images;

[0040] 4. Select similar pixels;

[0041] In this invention, the final field segmentation result is obtained based on a high spatial resolution image, and the segmented fields are used as a window for searching similar pixels. Pixels within the window that have a similar spectrum to the center pixel provide specific temporal and spatial information. A dynamic thresholding method is used to search for similar pixels, calculating the difference between adjacent pixels and the center pixel in the fine-resolution image, and setting a threshold to identify similar pixels. The formula is as follows:

[0042] |I(x i ,y i ,t k )-I(x c ,y c ,t k )|≤σ(I)×2 / N

[0043] I(x i ,y i ) is in the window (x i ,y i The pixel value at position ) I(x c ,y c ) is the value of the center pixel in the window. Similar pixels to the center pixel are obtained by calculation. σ(I) is the standard deviation of I for the entire band. N is the estimated land cover category. The larger the category number is set, the more stringent the conditions for selecting similar pixels from the fine resolution image are.

[0044] Calculate similar pixels in the known fine-resolution images before and after the predicted date, and take the similar pixels from the merged images of the known dates before and after the predicted date as the final similar pixel image.

[0045] In this embodiment of the invention, step 2 mainly includes the reconstruction of coarse pixel time-series vegetation index data:

[0046] This example uses the DAVIR-MUTCOP method, combining MODIS 8-day composite products and daily data for reconstruction, correcting time biases and selecting high-quality images from observation points, fully leveraging the advantages of multi-timescale data to generate high-quality daily MODIS vegetation index time-series data.

[0047] In this embodiment of the invention, step 3 mainly includes the following parts:

[0048] 1. Fitting with a double logistic function;

[0049] 2. Optimize the function solution.

[0050] This example introduces the Double Logistic function to describe the dynamic changes of crops during their rapid growth and decline phases. The function, consisting of two superimposed S-shaped curves, captures the peaks and inflection points of the growth curve, revealing critical growth periods. By combining all available fine-resolution images, the average vegetation growth curve of similar pixels within the target area is fitted to obtain the initial prediction results. The mathematical expression of the Double Logistic function is:

[0051]

[0052] A1 and A2 represent the magnitudes of the two logistic functions, t1 and t2 are the inflection points of the two growth stages, and k1 and k2 control the growth rate. The two fitting stages of Double Logistic can independently control the growth and decay rates.

[0053] This invention uses a spectral linear mixture model to decompose the spectral information of pixels and estimate the composition ratio of various land cover types within them. The spectral linear mixture model is based on the linear mixture assumption, where the value of a coarse pixel is equal to the weighted sum of the values ​​of each land cover type within that coarse pixel, as shown in the formula:

[0054]

[0055] C obs R represents the observed coarse pixel value. i This represents the value of the i-th endmember. Here, similar pixels are considered as similar land cover types, and the endmember value is obtained by calculating the average pixel value of similar pixels; f i Let represent the abundance of the i-th endmember. Abundance represents the proportion of a certain land cover type in the mixed pixels, that is, the proportion of each type of fine pixels contained in each coarse pixel. ε represents the residual. The abundance of each endmember contained in each coarse pixel is added together to 1.

[0056] By combining the initial curves fitted by the Double Logistic function and constraining the spectral linear mixing model relationship between coarse and fine pixels, the vegetation index curves of various similar pixels are optimized by adjusting the fitting parameters of the Double Logistic function.

[0057] In this embodiment of the invention, step 4 mainly includes the following parts:

[0058] 1. Pixel weighted allocation;

[0059] 2. Spatial filtering.

[0060] In this invention, similar pixels are considered to have the same growth trend within a certain time period. Each pixel in the similar pixel set should occupy a certain proportion of the average value of similar pixels, and this proportion should be close to the proportion of the corresponding pixel in the reference images before and after the prediction date. By calculating the proportion of each pixel in the similar pixels in the reference images before and after the prediction date, the two proportions are weighted and combined according to the difference between the coarse pixels on the prediction date and the coarse pixels on the reference dates before and after the prediction date. Pixels in the reference image that are closer to the prediction date should have a larger weight. The formula is as follows:

[0061]

[0062] t m t n These are the base times before and after the predicted date, t p Indicates the predicted date, x i y j For the position of a pixel, T k This is the calculated time weight. The proportion of each pixel in the fine-resolution images of the baseline date before and after the predicted date is calculated among corresponding similar pixels. The value x of each pixel obtained in the nonlinear optimization function solution step is then used. i Multiplying by the corresponding weight yields the weighted pixel allocation value, as shown in the following formula:

[0063]

[0064] x jq Indicates the fine pixel value after weighted allocation; x simi and y simj F(x) represents the position of each pixel in the similar pixel group within the fine-resolution image; simi ,y simj ,t k ) represents each pixel in each class of similar pixels at time t m and t n The pixel value at any given moment.

[0065] The spatial filtering process in this example assigns weights to neighboring pixels selected by spatial distance, with closer pixels receiving higher weights. Spatial filtering addresses the blocky artifacts that appear in the prediction process described above. It selects the n pixels with the smallest spectral differences from a local window and weights them according to spatial distance to generate the final predicted value. Pixels that are closer together have higher weights, thus preserving local spatial structure and improving prediction accuracy. The formula is as follows:

[0066]

[0067] The spatial distance weights W of n pixels to the center pixel are used.i Calculate x jq Multiply by the corresponding W i The final prediction result F is obtained. fin (x,y).

[0068] The prediction accuracy of this example is compared with that of three classic spatiotemporal fusion methods: ESTARFM, Fit-FC, and FSDAF2.0. The results show that the Fit-STVIF model proposed in this invention exhibits superior prediction accuracy and stronger stability in different scenarios. The accuracy comparison results are as follows: Figure 3 As shown.

[0069] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland, characterized in that, Includes the following steps: Data preprocessing, field segmentation and similar pixel selection: Radiometric correction, atmospheric correction and geometric registration were performed on the coarse resolution and fine resolution remote sensing images respectively; the farmland area was divided into multiple field units using an image segmentation algorithm, which served as a spatial window for similar pixel selection. Based on the segmented fields, each field was used as a window to search for pixels similar to the center pixel. Coarse-resolution metadata reconstruction: Temporal reconstruction of coarse-resolution remote sensing images, removal of outliers and filling in data gaps to generate high-quality time-series vegetation index data; Fine pixel metadata fitting: Based on all available fine remote sensing images for the growing season, and on the basis of field segmentation, the Double Logistic function is used to fit the nonlinear growth trajectory of the average vegetation index of similar pixels to generate the initial value of fine pixels at the prediction time. Fine pixel growth curve optimization: Combining the reconstructed coarse resolution image and the fitted initial value, the fine resolution image at the prediction time is optimized by using the mixed pixel decomposition relationship of coarse and fine resolution pixel data as a constraint. Pixel weighted allocation and spatial filtering: Based on the spectral similarity and spatial distance of pixels within the field area, fusion weights are assigned and spatial filtering is performed to improve image quality and spatial continuity.

2. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 1, characterized in that, Similar pixel selection based on field segmentation: The field segmentation adopts an object-based image segmentation method, combined with multi-temporal data of high-resolution remote sensing images for boundary optimization, so as to improve the accuracy of field segmentation.

3. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 2, characterized in that, Specifically, the following steps are included: a. Calculate the WDRVI of the target region pixels according to the WDRVI calculation formula; the formula is: Scaled_WDRVI={[(α-1)+(α+1)×NDVI] / [(α+1)+(α-1)×NDVI]+(1-α) / (1+α)}×100 NDVI=(r nir -r red ) / (ρ nir +r red ), α 0.1; b. Obtain the coarse and fine spatial resolution data of the target area; c. Segmenting land parcels using fine spatial resolution images; d. Select similar pixels; The final field segmentation result is obtained based on the fine-resolution image. The segmented fields are used as a window for searching similar pixels. Pixels within the window that have similar spectra to the center pixel provide specific temporal and spatial information. A dynamic thresholding method is used to search for similar pixels. The difference between adjacent pixels and the center pixel in the fine-resolution image is calculated, and a threshold is set to identify similar pixels. The formula is as follows: |I(x i ,y i ,t k )-I(x c ,y c ,t k )|≤σ(I)×2 / N I(x i ,y i ) is in the window (x i ,y i The pixel value at position ) I(x c ,y c ) is the value of the center pixel in the window. Similar pixels of the center pixel are obtained by calculation. σ(I) is the standard deviation of I for the entire band. N is the estimated land cover category. The larger the category number is set, the more stringent the condition for selecting similar pixels from the fine resolution image is. Calculate similar pixels in the known fine-resolution images before and after the predicted date, and take the similar pixels from the merged images of the known dates before and after the predicted date as the final similar pixel image.

4. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 1, characterized in that, The spatiotemporal fusion model incorporates data reconstruction. This coarse-resolution data reconstruction combines vegetation index data across multiple time scales and employs a sliding time window and trend smoothing filtering algorithm to process time-series data, removing abnormal noise and filling in missing observations. Specifically, this includes: correcting time deviations and selecting images from high-quality observation points through multi-temporal data complementation; using SG filtering to smooth time-series data and generate high-quality image data with high temporal density; and effectively filling data gaps and ensuring temporal continuity during cloud pollution periods through synthetic data trend constraints and interpolation algorithms. The reconstructed WDRVI time-series images possess high temporal resolution, low noise, and temporal consistency, providing a more reliable data foundation for vegetation dynamic monitoring.

5. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 1, characterized in that, The nonlinear fitting model, which uses the Double Logistic function to fit the crop growth process, is a fine-pixel vegetation index growth curve. It can characterize multiple key stages such as the rapid growth period, growth inflection point, and maturity period.

6. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 5, characterized in that, By introducing the Double Logistic function to describe the dynamic changes of crops during their rapid growth and decline phases, the function, consisting of two superimposed S-shaped curves, captures the peaks and inflection points of the growth curve, revealing critical growth periods. By combining all available fine-resolution images and fitting the average vegetation growth curve of similar pixels within the target area, initial prediction results are obtained. The mathematical expression of the Double Logistic function is: A1 and A2 represent the magnitudes of the two segments of the Logistic function, t1 and t2 are the inflection points of the two growth stages, and k1 and k2 control the growth rate; the two fitting stages of Double Logistic can independently control the growth and decay rates. The spectral linear mixture model is used to decompose the spectral information of pixels and estimate the composition ratio of various land cover types within them. Based on the linear mixture assumption, the value of a coarse pixel is equal to the weighted sum of the values ​​of each land cover type within that coarse pixel, as shown in the formula: C obs R represents the observed coarse pixel value. i This represents the value of the i-th endmember. Here, similar pixels are considered as similar land cover types, and the endmember value is obtained by calculating the average pixel value of similar pixels; f i ε represents the abundance of the i-th endmember, where abundance represents the proportion of a certain land cover type in the mixed pixel, that is, the proportion of each type of fine pixel contained in each coarse pixel, and ε represents the residual. The abundance of each endmember contained in each coarse pixel is added together to 1. By combining the initial curves fitted by the Double Logistic function and constraining the spectral linear mixing model relationship between coarse and fine pixels, the vegetation index curves of various similar pixels are optimized by adjusting the fitting parameters of the Double Logistic function.

7. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 1, characterized in that, Considering spatial correlation, the spatial filtering process employs a weighted average filter or an edge-preserving filter to smooth noise in the fused image and improve spatial consistency. By calculating the proportion of each pixel in the similar pixels within the preceding and following reference images, and weighting the proportions based on the difference between the coarse pixels of the predicted date and the coarse pixels of the preceding and following reference dates, the spectral differences between the center pixel and its neighboring pixels are calculated. The top n pixels with the smallest spectral differences are selected from the local window and weighted by spatial distance to generate the final predicted value. Pixels closer to each other have higher weights to preserve local spatial structure and improve prediction accuracy.

8. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 7, characterized in that, The main steps include: a, pixel weighted allocation; b, Spatial filtering; Similar pixels are considered to have the same growth trend within a certain time period. Each pixel in a similar pixel should occupy a certain proportion of the average number of similar pixels, and this proportion should be close to the proportion of the corresponding pixel in the reference images before and after the prediction date. By calculating the proportion of each pixel in the similar pixels in the reference images before and after the prediction date, the two proportions are weighted and combined according to the difference between the coarse pixels on the prediction date and the coarse pixels on the reference dates before and after the prediction date. Pixels in the reference image that are closer to the prediction date should have a larger weight. The formula is as follows: t m t n These are the base times before and after the predicted date, t p Indicates the predicted date, x i y j For the position of a pixel, T k These are the calculated time weights; Calculate the proportion of each pixel in the corresponding similar pixels in the fine-resolution images of the baseline date before and after the prediction date, and then calculate the value x of each pixel obtained in the function optimization step. i Multiplying by the corresponding weight yields the weighted pixel allocation value, as shown in the following formula: x jq Indicates the fine pixel value after weighted allocation; x simi and y simj F(x) represents the position of each pixel in the similar pixel group within the fine-resolution image; simi ,y simj ,t k ) represents each pixel in each class of similar pixels at time t m and t n The pixel value at any given time; The spatial filtering process assigns weights to neighboring pixels selected based on spatial distance, with closer pixels receiving higher weights. Spatial filtering addresses the block artifacts that appear in the prediction process described above. It selects the n pixels with the smallest spectral differences from a local window and combines them using spatial distance weights to generate the final predicted value. Pixels that are closer to each other receive higher weights, thus preserving local spatial structure and improving prediction accuracy. The formula is as follows: The spatial distance weights W of n pixels to the center pixel are used. i Calculate x jq Multiply by the corresponding W i The final prediction result F is obtained. fin (x,y); The prediction accuracy of this example is compared with that of three classic spatiotemporal fusion methods, ESTARFM, Fit-FC, and FSDAF2.

0. The results show that the Fit-STVIF model proposed in this invention exhibits better prediction accuracy and stronger stability in different scenarios.

9. The spatiotemporal fusion method for vegetation indices that takes into account the spatiotemporal heterogeneity of farmland according to claim 1, characterized in that, This also includes Fit-STVIF model validation, specifically: using R... 2 The root mean square error (RMSE) is used as an evaluation metric to compare the actual image with the predicted image.

Images