High-temporal-spatial-resolution vegetation index fusion method based on multi-source optical satellite image

Abstract

A high temporal-spatial resolution vegetation index fusion method based on a multi-source optical satellite image comprises the following steps: firstly, performing radiometric calibration, atmospheric correction and geometric fine correction on Landsat, Sentinel-2 and MOD09A1 data, and unifying temporal-spatial resolution to 10m / 8 days; pixel-level fusion is carried out by adopting an improved continuous correction method, a correction coefficient K is introduced to compensate Sentinel-2 critical period data defect influence, and fusion precision is improved through dynamic weight adjustment; and finally, a continuous and smooth EVI time sequence is constructed by using cubic spline interpolation and Savitzky-Golay filtering. According to the method, single-source data space-time limitation is broken through, after fusion, the vegetation index spatial resolution reaches 10 m, the time resolution reaches 8 days, the key phenological period extraction error is smaller than or equal to 3 days, the crop classification precision is larger than or equal to 90%, the accuracy and continuity of farmland-scale vegetation monitoring can be remarkably improved, high-precision data support is provided for agricultural application such as crop growth assessment and water resource management, and the method is suitable for popularization and application. The method is suitable for cloudy and rainy areas and various crop types.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing technology and agricultural monitoring, and in particular relates to a high temporal and spatial resolution vegetation index fusion method based on multi-source optical satellite images. Background Art

[0002] In agricultural remote sensing monitoring, the spatiotemporal resolution of vegetation indices directly impacts the accuracy of crop phenological period extraction, planting structure classification, and water resource management. Existing single-source optical satellite data have significant limitations: high-spatial-resolution data (e.g., Sentinel-2, 10m) is frequently affected by cloud cover and has low temporal resolution (5-day revisit period), making it difficult to capture dynamic changes during the crop growth period. High-temporal-resolution data (e.g., MODIS, 8-day composite) has low spatial resolution (500m) and mixed pixel artifacts, leading to distortion in field-scale monitoring.

[0003] Although traditional multi-source fusion methods attempt to combine the two types of data, they face the following technical bottlenecks: Inconsistent spatiotemporal benchmarks: Differences in coordinate systems and resolutions of multi-source data lead to large geometric registration errors, and spectral information is easily distorted after interpolation; Insufficient compensation for data loss during key growth periods: High-resolution imagery during key crop growth periods (such as heading) is often missing due to cloud cover, and existing models do not quantify the impact of the degree of loss on fusion accuracy. Poor dynamic adaptability: The fixed-weight fusion model cannot adaptively adjust according to the crop growth stage (such as seedling stage and maturity stage), resulting in deviations in the extraction of vegetation index peak / trough features.

[0004] The technical problems that need to be solved urgently in this invention are: how to achieve high-precision spatiotemporal alignment in multi-source data preprocessing, how to quantify the impact of data missing in the critical period on the fusion results and dynamically correct it, and how to construct a time series smoothing algorithm with both noise suppression and feature retention, which have become the core challenges to improving the accuracy of vegetation index fusion. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a high-temporal-spatial-resolution vegetation index fusion method based on multi-source optical satellite images. The present invention systematically solves the above problems through standardized preprocessing of multi-source data, a pixel-level fusion model with a dynamic correction coefficient, and time-series smoothing optimization, thereby providing high-temporal-spatial-resolution vegetation index data for refined crop monitoring.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A high temporal and spatial resolution vegetation index fusion method based on multi-source optical satellite images, the steps are as follows: S1. Preprocess multi-source remote sensing data and unify spatiotemporal benchmarks: This is used to eliminate geometric distortion, radiometric differences, and spatiotemporal misalignment in multi-source data, and to construct a compatible basic dataset. S2. Build and dynamically optimize a pixel-level vegetation index fusion model: Combining the high temporal resolution of MODIS with the high spatial resolution of Sentinel-2, we will construct an adaptive fusion model to improve the spatiotemporal continuity of vegetation indices. S3. Construct and optimize high-temporal and spatial resolution EVI time series: Generate a continuous, smooth 10m / 8-day EVI time series to provide a basis for crop phenological extraction and growth monitoring; S4. Verify the fusion results and adapt them to agricultural applications: Verify the accuracy and reliability of the fusion method and connect it with crop monitoring services.

[0007] Preferably, S1.1 Perform radiometric calibration and atmospheric physics correction: First, perform radiometric calibration on the Landsat series data. Use ENVI software to call the radiometric conversion parameters in the metadata to convert the pixel brightness value into the apparent reflectance of the atmosphere without repeating the geometric correction.

[0008] For Sentinel-2 images, atmospheric correction was performed using the Sen2cor plug-in provided by the European Space Agency. By inputting parameters such as the altitude of the study area and aerosol optical depth, the effects of aerosol scattering and water vapor absorption in the atmosphere were eliminated to obtain the true reflectance of the surface. At the same time, in view of the resolution differences among the various bands of Sentinel-2, the nearest neighbor interpolation method was used to resample the entire band to a resolution of 10m to ensure pixel-level spatial consistency.

[0009] For the MOD09A1 data, the MRT tool was first used for batch mosaicking and coordinate system conversion, converting the sinusoidal projection to the WGS-84 geographic coordinate system. Subsequently, the geometrically corrected Sentinel-2 imagery was used as a benchmark, and an image-to-image geometric correction method was used to eliminate spatial offsets by automatically matching ground control points, with errors controlled to the sub-pixel level.

[0010] S1.2 Temporal and Spatial Resolution Unification and Regional Cropping: To achieve pixel-level fusion of multi-source data, the MOD09A1 data were re-interpolated using the nearest neighbor interpolation method, increasing the spatial resolution from 500 m to 10 m. This avoids spectral blurring caused by traditional interpolation methods. After achieving resolution unification, all images were cropped using the vector boundary of the study area to remove invalid background areas, reduce data redundancy, and improve subsequent processing efficiency. Furthermore, a quality control mask was constructed. Based on the cloud cover band of Sentinel-2 and the quality assessment flag of MOD09A1, pixels with cloud cover >10% were removed to retain high-quality observation data.

[0011] Preferably, S2.1, a continuous correction method is used as a basic model, and the spatial details of the low-resolution image are corrected by using the high-resolution image through linear weighting on the time series; S2.2: To address the data loss problem caused by cloud cover during the critical growth period of Sentinel-2, a correction factor K is introduced to optimize the fusion model. S2.3. Use a dual weight adjustment mechanism to improve fusion accuracy: 1. Prioritizing spatial quality: Based on Sentinel-2 cloud cover assessment results, the weight of pixels with cloud cover <5% is increased by 30% to reduce interference from low-quality data. 2. Temporal phenology sensitivity: During critical crop phenological periods, the Sentinel-2 weight is forced to exceed 60% through a preset phenological calendar, ensuring the effective preservation of high-resolution details during critical periods.

[0012] Preferably, S3.1: splice the fused EVI images in chronological order to form an annual time series. For periods of continuous data loss due to extreme weather, cubic spline interpolation is used to fill in the gaps. This method avoids trend deviations that may be introduced by linear interpolation by fitting the curve shape of adjacent valid data points, ensuring the continuity and physical rationality of the time series.

[0013] S3.2: In order to eliminate high-frequency noise such as sensor errors and atmospheric random disturbances, the Savitzky-Golay filter is used to smooth the EVI time series.

[0014] The specific parameters were set as a 21-day window length and a polynomial order of 2. Least squares fitting was used to fit the data points within the window, removing noise while preserving key features of the EVI curve (such as the sudden drop during the sowing period and the peak plateau during the maturity period). After filtering, the signal-to-noise ratio of the EVI curve improved by approximately 15%-20%, and the positional error of phenological key points (such as the maximum slope point and peak point) was less than 2 days.

[0015] Preferably, S4.1, verification is carried out from two dimensions: spatial consistency and temporal stability: Spatial indicators: Calculate the mean, standard deviation, coefficient of variation (CV), and spatial autocorrelation coefficient (Moran's I) of the EVI before and after fusion. The CV value of the fused data is required to increase by more than 20% (reflecting enhanced spatial heterogeneity) and the Moran's I value is ≥0.7 (indicating significant spatial autocorrelation). Time Indicators: Select cloud-free Sentinel-2 ground-truth images as the benchmark. Calculate the Pearson correlation coefficient (R^2 > 0.9) and root mean square error (RMSE < 0.05) between the fused EVI and the ground-truth image to ensure trend consistency of the time series.

[0016] S4.2 Phenological period extraction and field verification: Based on the smoothed EVI time series, the slope method is used to extract the key phenological periods of crops: Sowing period: the inflection point where the EVI curve changes from decreasing to increasing (the slope changes from negative to positive); Heading / tasseling period: the moment when EVI reaches its peak (the slope is 0 and the sign changes before and after); Mature stage: the starting point of the rapid decline phase of EVI (the minimum point of the first-order derivative).

[0017] The extracted results are compared with the data from ground observation stations, and the phenological period extraction error is required to be ≤3 days.

[0018] For example, the deviation between the remote sensing extraction results and the measured values ​​during the wheat sowing period in Lixin County in 2020 was +2 days, and the deviation during the corn maturity period was -1.5 days, which met the accuracy requirements of agricultural monitoring.

[0019] S4.3 The fused high temporal and spatial resolution EVI data can be directly used for: Crop planting structure classification: Using a decision tree algorithm combined with EVI temporal characteristics (such as growth period length and peak intensity), crops such as wheat and corn can be distinguished, with a classification accuracy of ≥90%; Water productivity estimation: Combined with evapotranspiration inverted from the SEBAL model and yield data estimated using the dry matter-harvest coefficient method, crop water productivity (CWP = yield / evapotranspiration) is calculated to provide a basis for optimizing regional water resources allocation.

[0020] Preferably, S2.1 The pixel-level fusion method used in this paper adopts the continuous correction method proposed by Liu Baoyuan et al. The fusion formula is shown in (3.3).

[0021] (3.3); Where: is the high-resolution pixel fusion value; The MODIS pixel value corresponding to this pixel; This pixel corresponds to the high-resolution data value at a certain time, with a total of n scenes; The day number corresponding to the acquisition of MODIS data; The day number corresponding to the acquisition of high-resolution data; for The weight of the high-resolution image of the time period, the expression is shown in (3.4).

[0022] (3.4); This formula was originally used to fuse HJ-NDVI images with MODIS-NDVI images. Here, it is used to fuse Sentinel-2-EVI images with MOD09A1-EVI images resampled to 10 m.

[0023] Preferably, S2.2 (2) Optimization of pixel-level fusion method The original fusion formula failed to account for the uneven time intervals of available high-resolution imagery and the temporal overlap of the fused images. While Sentinel-2 imagery has a short revisit period and a large data volume, some time periods are completely obscured by cloud and need to be excluded. Directly applying this formula would bias the resulting EVI data, hindering the subsequent extraction of key crop phenological periods.

[0024] Therefore, the correction factor is introduced here K 、 P 、 Q Perform optimization and see formulas (3.5)-(3.7) after optimization.

[0025] (3.5) (3.6) (3.7) Where: P The number of completely missing high-resolution images, i.e., Sentinel-2 images, in April and August of the year; Q is the number of completely missing Sentinel-2 images in May, June, September, and October; 0.3 and 0.1 are proportional coefficients, representing the importance of missing months to the final fusion result; the remaining parameters are the same as those in formulas (3.3) and (3.4).

[0026] Preferably, S3.2: In order to eliminate high-frequency noise such as sensor error and atmospheric random disturbance, the EVI time series is smoothed using Savitzky-Golay (SG) filtering.

[0027] After pixel-level fusion, the EVI data still contains noise, which causes the fused EVI time series curve to show sawtooth-like fluctuations, affecting the accuracy of parameter inversion and phenological analysis. Therefore, to better reflect the true growth characteristics of crops, this paper uses cubic spline interpolation to fill the fused EVI data and then reconstructs the filled EVI time series using the SG filtering method.

[0028] The SG filtering algorithm was proposed by Savitzky and Golay in 1964. It is a convolution algorithm based on the least squares principle and smooth time series, see formula (3.8).

[0029] (3.8) Where: is the EVI value after fitting; For the time series values; is the convolution filter coefficient for the i-th EVI value; a is the window width on both sides of the point to be fitted; is the length of the filter, which is 2 a +1. The key to good fitting effect lies in the filter window a and smoothing multi-pattern times d Generally speaking, a The larger the value, d The smaller the value, the smoother the fitted curve will be. a Too large a value will result in an overly smooth curve. a If the value is too small, the curve will retain more noise values, reducing the fitting effect.

[0030] A high-temporal-spatial-resolution vegetation index fusion system based on multi-source optical satellite images adopts a high-temporal-spatial-resolution vegetation index fusion method based on multi-source optical satellite images. A computer device comprising: one or more processors; The processor is configured to store one or more programs; When the one or more programs are executed by the one or more processors, a high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite images is implemented.

[0031] The present invention can achieve the following beneficial effects: 1. The fused vegetation index has both 10m spatial accuracy and 8-day temporal frequency, solving the problem of "spatial coarseness" or "temporal discontinuity" of single-source data, and clearly depicting the growth details and dynamic changes of farmland crops or vegetation.

[0032] 2. By compensating for the impact of data deficiencies through dynamic correction coefficients, the fusion error of key growth periods (such as heading period) is reduced by 40%, and the error in phenological period extraction is ≤3 days, meeting the needs of precise agricultural monitoring.

[0033] 3. The time series smoothing algorithm effectively removes noise while retaining the peak / trough characteristics of the vegetation index, improving the reliability of crop growth analysis.

[0034] 4. Supports high-precision crop classification (accuracy ≥ 90%) and water productivity estimation, providing efficient data support for agricultural resource management and disaster monitoring, and is adaptable to various satellite combinations and regional scenarios.

[0035] 5. Radiometric calibration and atmospheric correction were performed on Landsat, Sentinel-2, and MOD09A1 data to eliminate spectral distortion. Spatial alignment of the multi-source data was achieved through geometric precision correction and coordinate system transformation. MOD09A1 (500m) and Sentinel-2 (multi-resolution bands) were uniformly resampled to 10m resolution using nearest neighbor interpolation to ensure pixel-level alignment. Invalid data with excessive cloud cover were removed, and imagery was clipped using the study area boundary to reduce redundancy and preserve the effective observation range. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 This is a comparison of the images before and after fusion on September 5, 2020 in this embodiment; Figure 2 This is the EVI time series curve diagram of corn pixels before and after filtering in 2020 in this embodiment; Figure 3 This is a flow chart of the method for identifying phenological periods based on the EVI time series curve in this embodiment. DETAILED DESCRIPTION

[0037] The preferred solution is Figures 1 to 3 As shown in Figure 1, a high temporal and spatial resolution vegetation index fusion method based on multi-source optical satellite images is described. The steps are as follows: S1. Preprocess multi-source remote sensing data and unify spatiotemporal benchmarks: This is used to eliminate geometric distortion, radiometric differences, and spatiotemporal misalignment in multi-source data, and to construct a compatible basic dataset. S1.1 First, radiometric calibration is performed on Landsat data (L1T-level products). Using the radiometric conversion parameters in the metadata, ENVI software is used to convert pixel brightness values ​​(DN) to atmospheric apparent reflectance. Repeated geometric correction is not required (because L1T-level products already have fine correction). For Sentinel-2 imagery, atmospheric correction is performed using the Sen2cor plug-in provided by the European Space Agency (or the FLAASH tool in ENVI 5.5). By inputting parameters such as the study area's altitude and aerosol optical depth, the effects of atmospheric aerosol scattering and water vapor absorption are eliminated to obtain true surface reflectance. Furthermore, to account for the varying resolutions of Sentinel-2 bands (10m, 20m, and 60m), the entire band is resampled to 10m resolution using nearest neighbor interpolation to ensure pixel-level spatial consistency.

[0038] For the MOD09A1 data (an 8-day synthetic surface reflectance product with a resolution of 500m), the MRT (MODIS Reprojection Tool) was first used for batch mosaicking and coordinate system conversion, converting the sinusoidal projection to the WGS-84 geographic coordinate system. Subsequently, using the geometrically corrected Sentinel-2 imagery as a benchmark, an image-to-image geometric correction method was employed to eliminate spatial offsets by automatically matching ground control points (GCPs), keeping errors within the sub-pixel level (<5m).

[0039] S1.2 To achieve pixel-level fusion of multi-source data, the MOD09A1 data were re-interpolated using the nearest neighbor interpolation method, increasing the spatial resolution from 500m to 10m. This avoids spectral blurring caused by traditional interpolation methods (such as bilinear interpolation). After achieving resolution uniformity, all images were cropped using the vector boundary of the study area (e.g., the administrative boundary of Lixin County) to remove invalid background areas (setting the background value to 0). This reduces data redundancy and improves subsequent processing efficiency. Furthermore, a quality control mask (QA Mask) was constructed. Based on the Sentinel-2 cloud cover band (Band 10) and the MOD09A1 quality assessment flag, pixels with cloud cover >10% were removed to retain high-quality observation data.

[0040] S2. Build and dynamically optimize a pixel-level vegetation index fusion model: Combining the high temporal resolution of MODIS with the high spatial resolution of Sentinel-2, we will construct an adaptive fusion model to improve the spatiotemporal continuity of vegetation indices. The optimized product MOD09A1 data based on MODIS source data has a resolution of 8 days and a spatial resolution of 500 meters. It fully utilizes the advantages of the high temporal resolution of MOD09A1 data and the high spatial resolution of Sentinel-2 data, and fuses the EVI images inverted from the two to obtain an EVI time series with a spatial resolution of 10 meters and a temporal resolution of 8 days.

[0041] Remote sensing image data fusion can be categorized into three levels, from low to high: pixel-level, feature-level, and decision-level. Pixel-level fusion fuses multi-source remote sensing images after geometric registration according to a fixed algorithm or formula. It is primarily applicable to image pixels or time series. Although pixel-level fusion is the lowest level and requires a large amount of image data, it preserves as much of the original image information as possible and offers high fusion accuracy. Liu Jilin et al., to achieve more accurate classification, performed a weighted fusion of TM images with panchromatic aerial photographs, improving image quality. Feature-level fusion is the most widely used. Multi-source remote sensing images are segmented and feature extracted. Feature-level fusion methods (such as statistical models) are then used to obtain feature vectors for comprehensive analysis and image fusion. This fusion method is computationally efficient, but suffers from significant information compression, resulting in some information loss. Leng Yang's feature-level fusion method, based on maximum correlation minimum redundancy and principal component analysis, significantly improved target recognition rates for drones. Decision-level fusion is the highest level of fusion. Each data source first classifies the target and then combines it with other criteria to fuse the best image combination representing the target's characteristics. This method is open and highly resistant to interference, but suffers from low precision. Bai Yu et al. used decision-level fusion of visible light and infrared images to achieve accurate target detection. Analysis of the three fusion methods above shows that higher-level fusion methods are simpler and require less computation, but also suffer from lower accuracy. Due to the abundance of satellite imagery in the study area, pixel-level fusion was selected to ensure the fusion accuracy of Sentinel-2 and MOD09A1 images.

[0042] S2.1. Using the continuous correction method as the basic model, the spatial details of the low-resolution image are corrected with the high-resolution image through linear weighting on the time series. The pixel-level fusion method used in this paper adopts the continuous correction method proposed by Liu Baoyuan et al. The fusion formula is shown in (3.3).

[0043] (3.3) Where: is the high-resolution pixel fusion value; The MODIS pixel value corresponding to this pixel; This pixel corresponds to the high-resolution data value at a certain time, with a total of n scenes; The day number corresponding to the acquisition of MODIS data; The day number corresponding to the acquisition of high-resolution data; for The weight of the high-resolution image of the time period, the expression is shown in (3.4).

[0044] (3.4) This formula was originally used to fuse HJ-NDVI images with MODIS-NDVI images. Here, it is used to fuse Sentinel-2-EVI images with MOD09A1-EVI images resampled to 10 m.

[0045] S2.2: To address the data loss problem caused by cloud cover during the critical growth period of Sentinel-2, a correction factor K is introduced to optimize the fusion model. (2) Optimization of pixel-level fusion method The original fusion formula failed to account for the uneven time intervals of available high-resolution imagery and the temporal overlap of the fused images. While Sentinel-2 imagery has a short revisit period and a large data volume, some time periods are completely obscured by cloud and need to be excluded. Directly applying this formula would bias the resulting EVI data, hindering the subsequent extraction of key crop phenological periods.

[0046] Therefore, the correction factor is introduced here K 、 P 、 Q Perform optimization and see formulas (3.5)-(3.7) after optimization.

[0047] (3.5) (3.6) (3.7) Where: P The number of completely missing high-resolution images, i.e., Sentinel-2 images, in April and August of the year; Q is the number of completely missing Sentinel-2 images in May, June, September, and October; 0.3 and 0.1 are proportional coefficients, representing the importance of missing months to the final fusion result; the remaining parameters are the same as those in formulas (3.3) and (3.4).

[0048] The correction coefficient represents the reliability of the overall Sentinel-2 image used for fusion, because these months are the key periods for extracting crop phenological periods, and the EVI values ​​of these months are at the peak or trough of the EVI time series of the entire crop growth period, which has a greater impact on the fusion effect. In addition, images with higher EVI levels (April and August) have a greater impact on fusion accuracy than images with lower EVI levels (May, June, September, and October), so the proportional coefficient of the correction coefficient P is slightly larger than the proportional coefficient of the correction coefficient Q. Calculate the correction coefficients from 2016 to 2021 according to formula (3.7) K The above formula is used to obtain the EVI data after fusion in each period and combine them to construct the corresponding EVI time series. Taking the MOD09A1-EVI data on September 5, 2020 as an example, the image effects before and after fusion are shown in Figure 1 It can be seen that the spatial resolution of the image before fusion is 500m. The presence of a large number of mixed pixels makes it difficult to distinguish the type of ground objects. The resolution of the image after fusion is 10m. The ground object type is clearer and the corresponding EVI value is closer to the true value.

[0049] Table 3.3 Years K Value selection

[0050] S2.3. Use a dual weight adjustment mechanism to improve fusion accuracy: 1. Prioritizing spatial quality: Based on Sentinel-2 cloud cover assessment results, the weight of pixels with cloud cover <5% is increased by 30% to reduce interference from low-quality data. 2. Temporal phenology sensitivity: During critical crop phenological periods, the Sentinel-2 weight is forced to exceed 60% through a preset phenological calendar, ensuring the effective preservation of high-resolution details during critical periods.

[0051] S3. Construct and optimize high-temporal and spatial resolution EVI time series: Generate a continuous, smooth 10m / 8-day EVI time series to provide a basis for crop phenological extraction and growth monitoring; S3.1: The fused EVI images are spliced ​​chronologically to form an annual time series. For periods of continuous data loss due to extreme weather, cubic spline interpolation is used to fill in the gaps. This method fits the curve shape of adjacent valid data points, avoiding trend deviations that may be introduced by linear interpolation, and ensuring the continuity and physical rationality of the time series.

[0052] S3.2: In order to eliminate high-frequency noise such as sensor errors and atmospheric random disturbances, the Savitzky-Golay filter is used to smooth the EVI time series.

[0053] The specific parameters were set as a 21-day window length and a polynomial order of 2. Least squares fitting was used to fit the data points within the window, removing noise while preserving key features of the EVI curve (such as the sudden drop during the sowing period and the peak plateau during the maturity period). After filtering, the signal-to-noise ratio of the EVI curve improved by approximately 15%-20%, and the positional error of phenological key points (such as the maximum slope point and peak point) was less than 2 days.

[0054] S3.2: In order to eliminate high-frequency noises such as sensor errors and atmospheric random disturbances, the Savitzky-Golay (SG) filter is used to smooth the EVI time series.

[0055] After pixel-level fusion, the EVI data still contains noise, which causes the fused EVI time series curve to show sawtooth-like fluctuations, affecting the accuracy of parameter inversion and phenological analysis. Therefore, to better reflect the true growth characteristics of crops, this paper uses cubic spline interpolation to fill the fused EVI data and then reconstructs the filled EVI time series using the SG filtering method.

[0056] The SG filtering algorithm was proposed by Savitzky and Golay in 1964. It is a convolution algorithm based on the least squares principle and smooth time series, see formula (3.8).

[0057] (3.8) Where: is the EVI value after fitting; For the time series values; is the convolution filter coefficient for the i-th EVI value; a is the window width on both sides of the point to be fitted; is the length of the filter, which is 2 a +1. The key to good fitting effect lies in the filter window a and smoothing multi-pattern times d Generally speaking, a The larger the value, d The smaller the value, the smoother the fitted curve will be. a Too large a value will result in an overly smooth curve. a If the value is too small, the curve will retain more noise values, which will reduce the fitting effect. a This paper avoids these two points by expanding the length of the time series. aThe influence of each point on the fitting accuracy is realized by using IDL language to realize the SG filtering of the time series curve. The function for calculating the convolution factor is SAVGOL. The time series curve is composed of interpolated data. The time series is long. The key parameters of the filter are selected a =20, d =2, and taking the 2020 autumn harvest corn pixel as an example, the effects before and after SG filtering are shown. Figure 2 The peak and trough values ​​of the filtered curve do not deviate too much, which not only effectively removes the noise points in the original curve, but also better maintains the shape of the curve, which is conducive to the extraction of key phenological periods.

[0058] S4. Verify the fusion results and adapt them to agricultural applications: Verify the accuracy and reliability of the fusion method and connect it with crop monitoring services.

[0059] S4.1. Verification from the two dimensions of spatial consistency and temporal stability: Spatial indicators: Calculate the mean, standard deviation, coefficient of variation (CV), and spatial autocorrelation coefficient (Moran's I) of the EVI before and after fusion. The CV value of the fused data is required to increase by more than 20% (reflecting enhanced spatial heterogeneity) and the Moran's I value is ≥0.7 (indicating significant spatial autocorrelation). Time Indicators: Select cloud-free Sentinel-2 ground-truth images as the benchmark. Calculate the Pearson correlation coefficient (R^2 > 0.9) and root mean square error (RMSE < 0.05) between the fused EVI and the ground-truth image to ensure trend consistency of the time series.

[0060] S4.2 Phenological period extraction and field verification: Based on the smoothed EVI time series, the slope method is used to extract the key phenological periods of crops: Sowing period: the inflection point where the EVI curve changes from decreasing to increasing (the slope changes from negative to positive); Heading / tasseling period: the moment when EVI reaches its peak (the slope is 0 and the sign changes before and after); Mature stage: the starting point of the rapid decline phase of EVI (the minimum point of the first-order derivative).

[0061] The extracted results are compared with the data from ground observation stations, and the phenological period extraction error is required to be ≤3 days.

[0062] For example, the deviation between the remote sensing extraction results and the measured values ​​during the wheat sowing period in Lixin County in 2020 was +2 days, and the deviation during the corn maturity period was -1.5 days, which met the accuracy requirements of agricultural monitoring.

[0063] S4.3 The fused high temporal and spatial resolution EVI data can be directly used for: Crop planting structure classification: Using a decision tree algorithm combined with EVI temporal characteristics (such as growth period length and peak intensity), crops such as wheat and corn can be distinguished, with a classification accuracy of ≥90%; Water productivity estimation: Combined with evapotranspiration inverted from the SEBAL model and yield data estimated using the dry matter-harvest coefficient method, crop water productivity (CWP = yield / evapotranspiration) is calculated to provide a basis for optimizing regional water resources allocation.

[0064] A high-temporal-spatial-resolution vegetation index fusion system based on multi-source optical satellite images adopts a high-temporal-spatial-resolution vegetation index fusion method based on multi-source optical satellite images. A computer device comprising: one or more processors; The processor is configured to store one or more programs; When the one or more programs are executed by the one or more processors, a high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite images is implemented.

[0065] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite images, characterized by This includes the following methods: S1. Preprocess multi-source remote sensing data and unify spatiotemporal benchmarks: This is used to eliminate geometric distortion, radiometric differences, and spatiotemporal misalignment in multi-source data, and to construct a compatible basic dataset. S2. Build and dynamically optimize a pixel-level vegetation index fusion model: Combining the high temporal resolution of MODIS with the high spatial resolution of Sentinel-2, we will construct an adaptive fusion model to improve the spatiotemporal continuity of vegetation indices. S3. Construct and optimize high-temporal and spatial resolution EVI time series: Generate a continuous, smooth 10m / 8-day EVI time series to provide data support for crop phenological period extraction and growth monitoring; S4. Verify the fusion results and adapt them to agricultural applications: Verify the accuracy and reliability of the fusion method and connect it with crop monitoring services.

2. The high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite imagery according to claim 1, characterized in that: The sub-steps of S1 are: S1.

1. Perform radiometric calibration and atmospheric physics correction: First, perform radiometric calibration on the Landsat data. Using ENVI software, call the radiometric conversion parameters in the metadata to convert pixel brightness values ​​into atmospheric apparent reflectance, without repeating geometric correction. Sentinel-2 images were atmospherically corrected using the Sen2cor plug-in provided by the European Space Agency. By inputting the altitude of the study area and the aerosol optical depth, the effects of aerosol scattering and water vapor absorption in the atmosphere were eliminated to obtain the true surface reflectance. Furthermore, to address the resolution differences among Sentinel-2 bands, the entire band was resampled to 10m resolution using the nearest neighbor interpolation method to ensure pixel-level spatial consistency. For the MOD09A1 data, we first used the MRT tool to perform batch mosaicking and coordinate system conversion, converting the sinusoidal projection to the WGS-84 geographic coordinate system. Then, using the geometrically corrected Sentinel-2 imagery as a benchmark, we employed an image-to-image geometric correction method to eliminate spatial offsets by automatically matching ground control points, keeping errors to sub-pixel levels. S1.

2. Unification of spatiotemporal resolution and regional cropping: To achieve pixel-level fusion of multi-source data, the nearest neighbor interpolation method was again performed on the MOD09A1 data to increase the spatial resolution from 500m to 10m, avoiding spectral blurring caused by traditional interpolation methods. After completing the resolution unification, all images were cropped using the vector boundary of the study area to eliminate invalid background areas, reduce data redundancy, and improve subsequent processing efficiency. At the same time, a quality control mask was constructed. Based on the cloud cover band of Sentinel-2 and the quality assessment mark of MOD09A1, pixels with cloud cover >10% were eliminated to retain high-quality observation data.

3. The high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite imagery according to claim 1, characterized in that: The sub-steps of S2 are: S2.

1. Using the continuous correction method as the basic model, the spatial details of the low-resolution image are corrected with the high-resolution image through linear weighting on the time series. S2.2: To address the data loss problem caused by cloud cover during the critical growth period of Sentinel-2, a correction factor K is introduced to optimize the fusion model. S2.

3. Use a dual weight adjustment mechanism to improve fusion accuracy: 1) Spatial quality priority: Based on the cloud cover assessment results of Sentinel-2, the weight of pixels with cloud cover <5% is increased by 30% to suppress interference from low-quality data; 2) Temporal phenology sensitivity: During the critical phenological period of crops, the weight of Sentinel-2 is forced to be increased to more than 60% through a preset phenological calendar to ensure the effective retention of high-resolution details during the critical period.

4. The high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite imagery according to claim 1, characterized in that: The sub-steps of S3 are: S3.

1. The fused EVI images are spliced ​​in chronological order to form an annual time series. For consecutive periods of missing data due to extreme weather, cubic spline interpolation is used to fill in the gaps. This method fits the curve shape of adjacent valid data points to avoid trend deviations that may be introduced by linear interpolation, ensuring the continuity and physical rationality of the time series. S3.

2. To eliminate sensor errors and atmospheric random disturbances, Savitzky-Golay filtering is used to smooth the EVI time series.

5. The high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite imagery according to claim 1, characterized in that: The sub-steps of S4 are: S4.

1. Verification from the two dimensions of spatial consistency and temporal stability: Spatial indicators: Calculate the mean, standard deviation, coefficient of variation, and spatial autocorrelation coefficient of EVI before and after fusion; Time Indicators: We selected cloud-free Sentinel-2 ground-truth images as the benchmark and calculated the Pearson correlation coefficient and root mean square error between the fused EVI and the ground-truth images to ensure trend consistency over the time series. S4.

2. Phenological period extraction and field verification: Based on the smoothed EVI time series, the slope method is used to extract the key phenological periods of crops: Sowing period: the inflection point where the EVI curve turns from decreasing to increasing; Heading stage: the moment when EVI reaches its peak; Maturity stage: the starting point of the rapid decline phase of EVI; Compare the extracted results with ground observation station data; S4.3 The fused high temporal and spatial resolution EVI data are directly used for: Crop planting structure classification: Using a decision tree algorithm combined with EVI time series features to distinguish between wheat and corn crops; Water productivity estimation: Combined with evapotranspiration inverted from the SEBAL model and yield data estimated using the dry matter-harvest coefficient method, crop water productivity is calculated to provide a basis for optimal allocation of regional water resources.

6. The high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite imagery according to claim 3, characterized in that: In S2.1, the fusion formula of the continuous correction method is shown in (3.3); (3.3)； Where: is the high-resolution pixel fusion value; The MODIS pixel value corresponding to this pixel; This pixel corresponds to the high-resolution data value at a certain time, with a total of n scenes; The day number corresponding to the acquisition of MODIS data; The day number corresponding to the acquisition of high-resolution data; for The weight of the high-resolution image of the time period, the expression is shown in (3;4); (3.4)； The Sentinel-2-EVI image was fused with the MOD09A1-EVI image resampled to 10m.

7. The high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite imagery according to claim 3, characterized in that: In S2.2, (2) Optimization of pixel-level fusion method Introducing correction factors K 、 P 、 Q Perform optimization, and see formulas (3.5)-(3.7) after optimization; (3.5)； (3.6)； (3.7)； Where: P The number of completely missing high-resolution images, i.e., Sentinel-2 images, in April and August of the year; Q is the number of completely missing Sentinel-2 images in May, June, September and October; 0.3 and 0.1 are proportional coefficients, representing the importance of the missing month to the final fusion result; the remaining parameters are the same as those in formula (3.3) and formula (3.4).

8. The high spatiotemporal resolution vegetation index fusion method based on multi-source optical satellite imagery according to claim 4, characterized in that: S3.2; To eliminate sensor errors and atmospheric random disturbances, Savitzky-Golay filtering is used to smooth the EVI time series; After pixel-level fusion, the EVI data still contains noise, resulting in jagged fluctuations on the fused EVI time series curve, which affects the accuracy of parameter inversion and phenological analysis. Therefore, to better reflect the true growth characteristics of crops, the fused EVI data are padded using cubic spline interpolation, and then the padded EVI time series is reconstructed using the SG filtering method. The convolution algorithm based on the least squares principle and smoothed time series is shown in Formula (3.8). (3.8)； Where: is the EVI value after fitting; For the time series values; is the convolution filter coefficient for the i-th EVI value; a is the window width on both sides of the point to be fitted; is the length of the filter, which is 2 a +1.

9. A high-temporal and spatial resolution vegetation index fusion system based on multi-source optical satellite imagery, characterized by: A high temporal and spatial resolution vegetation index fusion method based on multi-source optical satellite images according to any one of claims 1 to 8 is adopted.

10. A computer device, characterized in that: include: one or more processors; The processor is configured to store one or more programs; When the one or more programs are executed by the one or more processors, a high temporal and spatial resolution vegetation index fusion method based on multi-source optical satellite images as described in any one of claims 1 to 8 is implemented.

Images