Crop season phenology detection method based on multi-year VH polarization signal time sequence

Abstract

The invention belongs to the technical field of remote sensing monitoring, particularly provides a crop season phenology detection method based on a multi-year VH polarization signal time sequence, and aims to solve the problems that most existing phenology detection methods based on SAR (Synthetic Aperture Radar) data are limited to retrospective post-season analysis and lack of near-real-time prediction capability. And meanwhile, the technical problem that the SAR data is easily interfered by noise in near-real-time phenological detection is solved. According to the method, the correlation of a multi-year VH time sequence and the advantages of intra-season and post-season detection are fully considered, and a target pixel optimal prediction VH time sequence is obtained according to multi-year average data of a plurality of sampling points in a space window; and through translation matching of a local space window, selecting a plurality of optimal prediction VH time sequences and corresponding translation amounts thereof, obtaining post-season phenological detection results by adopting an SMF-S method, adding the corresponding translation amounts to obtain in-season phenological detection results, and obtaining a near-real-time phenological detection result after taking a median value from the plurality of in-season phenological detection results.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing monitoring, and specifically provides a method for detecting crop mid-season phenology based on multi-year Sentinel-1 VH polarization signal time series. Background Art

[0002] The Sentinel-1 satellite is a series of advanced Earth observation platforms launched by the European Space Agency (ESA). Its core equipment is a C-band synthetic aperture radar (SAR) with a wavelength of approximately 5.6 cm. This satellite can break through the limitations of clouds and day and night and provide high-resolution surface radar image data under any weather conditions. Based on this unique advantage, Sentinel-1 data is widely used in surface dynamic monitoring, such as land cover classification, inversion of surface fine parameters, ground subsidence analysis, and vegetation monitoring. Especially in crop phenology monitoring, the frequent imaging ability of Sentinel-1 provides continuous and reliable data support for capturing the key growth stages of crops from sowing to harvesting. However, its scattering intensity signal is easily interfered by external environmental factors and changes in crop growth status in time series analysis, resulting in abnormal noise in the data. This problem is more prominent during phenology detection, and local abnormal noise has a great negative impact on the final near-real-time phenology detection results. At the same time, in existing research, phenology detection methods based on Sentinel-1 SAR data are often post-season detection algorithms, that is, after the end of a growing season, the data of the entire season is analyzed to extract complete phenology information (such as sowing date, green-up date, maturity date, etc.). Due to the instability of SAR time series data, mid-season phenology detection is more challenging.

[0003] Currently, researchers have proposed various methods for phenology detection methods based on SAR data. By analyzing the polarization characteristics, backscattering intensity, and covariance matrix characteristics of radar data, indices related to vegetation growth are extracted to capture changes in different growth stages of crops. Combining cluster analysis, the data range of vegetation indices is associated with crop phenological periods to qualitatively identify crop growth stages. These methods deeply explore the potential of radar data through physical and mathematical models and effectively monitor crop status. However, these methods also face many challenges, such as complex index calculations, and the results of cluster analysis are mostly used for qualitative judgments and are difficult to provide accurate phenological period time information. More research detects phenological periods through the geometric characteristics of the time series of polarization signals. For example, geometric features such as the mutation points or local extrema of the slope are analyzed, and the relationship between geometric feature points and crop phenological periods is studied. After smoothing the time series, these key geometric feature points can be directly extracted and accurately matched with known crop phenological stages.

[0004] Most existing studies directly smooth time series data or generate parameters related only to vegetation growth through complex mathematical and physical calculations, and then use these data to obtain the specific phenological periods of crops through post-season phenological detection methods. However, these methods fail to fully restore the regular relationship between radar scattering signals and the growth cycle of target crops, and do not effectively suppress the interference of non-crop factors (such as soil, climate, field debris, etc.) in complex growth environments on time series data. In addition, traditional SAR time series-based phenological detection methods mostly perform post-season analysis based on data from the entire growing season, which is only applicable to retrospective studies and difficult to achieve near-real-time prediction. Therefore, the algorithm research for obtaining phenological information in real-time or near-real-time at key stages of the crop growing season based on SAR time series data still faces great challenges. Summary of the Invention

[0005] The purpose of the present invention is to provide a mid-season phenological detection method based on the Sentinel-1 VH polarization signal time series, which is used to solve the problem that existing SAR data-based phenological detection methods are mostly limited to retrospective post-season analysis and lack the ability of near-real-time prediction, and at the same time overcome the technical problem that SAR data is vulnerable to noise interference in near-real-time phenological detection. The present invention fully considers the correlation of multi-year VH time series and the advantages of mid-season and post-season detection, obtains the annual time series after multi-year averaging of multiple sampling points within the spatial window, and through macro stretching and translation, establishing regression relationships, and local stretching, obtains the preferred predicted VH time series for the current year; then through a local spatial window, translates the preferred predicted VH time series for the current year, calculates the correlation coefficient with the true VH time series of the current year, selects multiple preferred predicted VH time series with the optimal correlation coefficient and their corresponding translation amounts, performs phenological detection on each preferred predicted VH time series based on the SMF-S algorithm, obtains the post-season phenological detection results, adds their corresponding translation amounts to obtain the mid-season phenological detection results, and takes the median of multiple mid-season phenological detection results to obtain the near-real-time phenological detection results.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for detecting the mid-season phenology of crops based on the multi-year Sentinel-1 VH polarization signal time series, characterized by comprising the following steps:

[0008] S1. Obtain the Sentinel-1 VH polarization signal time series VH(t) of the detected crop at the target pixel from the beginning of the current year to the end time of mid-season detection; set a spatial window centered on the target pixel, and obtain the multi-year annual Sentinel-1 VH polarization signal time series VH of the sampling pixels of multiple detected crops within the spatial window target (t); i(t), where i is the sampling pixel number; calculate the average value of the Sentinel-1 VH polarization signal time series for each sampling pixel over multiple previous years throughout the year to obtain the average Sentinel-1 VH polarization signal time series of the sampling pixel over multiple previous years For the average Sentinel-1 VH polarization signal time series of each sampling pixel over multiple previous years and the Sentinel-1 VH polarization signal time series VH of the target pixel target (t) perform mean filtering;

[0009] S2. For the average Sentinel-1 VH polarization signal time series of each sampling pixel over multiple previous years Intercept the partial VH polarization signal time series from the beginning of the year to the detection cut-off time in the middle of the season and perform translation and stretching matching with the Sentinel-1 VH polarization signal time series VH of the target pixel target (t) to generate the preferred predicted VH polarization signal time series of the detected crop in the target pixel in the current year

[0010] S3. For the preferred predicted VH polarization signal time series of the target pixel in the current year generated by each sampling pixel, perform translation matching between the Sentinel-1 VH polarization signal time series of the target pixel and the preferred predicted VH polarization signal time series of the target pixel in the current year generated by the sampling pixel within a preset time window before the detection cut-off time, calculate the correlation coefficient, and select multiple groups of preferred translation amounts and their corresponding preferred predicted VH polarization signal time series according to the correlation coefficient;

[0011] S4. For each group of preferred translation amounts and their corresponding preferred predicted VH polarization signal time series of the target pixel in the current year, use the SMF-S method to perform post-season phenology detection on the preferred predicted VH planned signal time series of the target pixel in the current year to obtain the post-season phenology detection result, and add the post-season phenology detection result to the preferred translation amount to obtain the corresponding mid-season phenology detection result;

[0012] S5. Take the median of the calculated multiple groups of mid-season phenology detection results to obtain the near-real-time mid-season phenology detection result.

[0013] Furthermore, the specific process of step S2 is as follows:

[0014] For the average Sentinel-1 VH polarization signal time series of each sampling pixel over multiple previous years Intercept the partial VH polarization signal time series from the beginning of the year to the detection cut-off time in the middle of the season And combine it with the Sentinel-1 VH polarization signal time series VH of the target pixel target (t) for overall translation and stretching processing, and correspondingly obtain the partial multi-year average Sentinel-1 VH polarization signal time series of the sampled pixel and the Sentinel-1 VH polarization signal time series VH of the target pixel target (T);

[0015] Calculate the stretched and translated VH polarization signal time series of the detected crop in the multi-year average of each sampled pixel in previous years and the Sentinel-1 VH polarization signal time series VH of the target pixel target (T) to obtain the R correlation coefficient value;

[0016] Select the predicted VH time series with the largest correlation coefficient value as the preferred one to obtain the preferred predicted VH polarization signal time series of the detected crop in the target pixel in the current year

[0017] Use the GMFR algorithm to calculate the partial multi-year average optimal Sentinel-1 VH polarization signal time series of the sampled pixel and the Sentinel-1 VH polarization signal time series of the target pixel to obtain the regression coefficients α and β, and use the regression coefficients to generate the predicted VH polarization signal time series of the detected crop in the target pixel in the current year Specifically expressed as:

[0018]

[0019] Obtain the preferred predicted time series of the Sentinel-1 VH polarization signal of the target pixel to obtain the time T at which the maximum VH polarization signal value is located max , when the time T at which the maximum VH polarization signal value is located max is different from the mid-season detection cut-off time t cut_off , centered on the time T at which the maximum VH polarization signal value is located max , perform local stretching on the time series on the left and right sides of the center, and calculate the MSD value of the stretched result and the Sentinel-1 VH polarization signal time series of the target pixel in the corresponding time period. Take the stretched result with the minimum MSD value as the preferred one, and recombine the left and right stretched results with the center to generate the preferred predicted VH time series of the target pixel in the current year

[0020] Furthermore, in step 2, the overall stretching and translation processing is specifically expressed as:

[0021] T = xscale × (t + tshift)

[0022] Among them, T represents the time variable after translation and stretching processing, t represents the original time variable; tshift represents the time scale translation parameter, and xscale represents the time scale stretching parameter.

[0023] Furthermore, in step S2, local stretching is expressed as:

[0024] T' = xscale × |T - T max |

[0025] Among them, T' represents the time variable after local stretching processing, T represents the time variable after translation and stretching processing, and xscale represents the time scale stretching parameter.

[0026] Further, the specific process of step S3 is as follows:

[0027] Set the length of the translation matching time window to T match , and set the maximum translation range and translation step of the translation matching;

[0028] For the preferred predicted VH polarization signal time series of the target pixel of the i-th sampled pixel in the current year With [T end - T match , T end as the starting time window for translation, T end is the detection cut-off time; for the j-th translation matching, intercept the preferred predicted VH polarization signal time series data within the corresponding time window (T match ), and match it with the VH polarization signal time series VG target (T) within the time window [T end - T match , T end to calculate the correlation coefficient R i,j ;

[0029] Set the quantity threshold N, and based on the correlation coefficient R i,j , select the top N maximum values, record the translation amount corresponding to each maximum value as the preferred translation amount and record the corresponding preferred predicted VH polarization signal time series of the target pixel in the current year n = 1, 2,..., N.

[0030] Based on the above technical solutions, the beneficial effects of the present invention are as follows:

[0031] The present invention provides a method for detecting mid-season phenology based on multi-year Sentinel-1 VH polarization signal time series. By sampling the same crop pixels around the target pixel, a multi-year average VH polarization signal time series is generated, ensuring the diversity and consistency of the Sentinel-1 VH polarization time series data. Moreover, through the overall curve stretching and translation of the multi-year average VH polarization signal time series, the macroscopic consistency between the multi-year average VH polarization signal time series and the known part of the VH polarization time series data of the target pixel in the current year is ensured, avoiding large systematic differences between the data. By using the GMFR algorithm to establish the regression relationship between the multi-year average VH polarization signal time series and the known part of the VH polarization time series data of the target pixel in the current year, the VH polarization time series data of the target pixel in the current year is predicted through the regression relationship, improving the prediction effect of the abnormal outliers in the data on the VH polarization time series data of the target pixel in the current year. At the same time, through the local curve stretching for predicting the VH polarization time series data of the target pixel in the current year, the local environmental noise will not cause further negative impacts on the known part of the VH polarization time series data in the current year, restoring the regularity between the VH polarization time series data and crop growth. Through the local spatial window translation matching of multiple optimal predicted VH polarization time series data of the target pixel in the current year and the true known part of the VH polarization time series data in the current year, multiple most similar predicted VH polarization time series data of the target pixel in the current year and their optimal translation steps are obtained from the mid-season detection part, making the multiple most similar predicted VH polarization time series data of the target pixel in the current year closer to the actual situation in the current year before and after the phenological period to be predicted, better meeting the requirements of near-real-time mid-season phenology detection. Through the post-season phenology detection of the SMF-S algorithm for multiple most similar predicted VH polarization time series data of the target pixel in the current year and combining the results with the optimal translation steps, the phenology detection result of the present algorithm is finally obtained, making up for the instability of mid-season phenology detection, combining the advantages of the two phenology detection methods, and further improving the accuracy of phenology detection. Finally, the present invention has the advantages of suppressing the influence of noise in the complex growth environment on the VH polarization time series data, restoring the regularity between the VH polarization time series data and the true crop growth, and at the same time combining the results of mid-season time window translation and SMF-S post-season detection, making up for the instability of simple mid-season phenology detection, and finally obtaining a high-precision mid-season phenology detection result. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 FIG. is a schematic flowchart of the method for detecting mid-season phenology based on multi-year Sentinel-1 VH polarization signal time series in the present invention.

[0033] Figure 2 FIG. is a distribution map of the research area in the embodiment of the present invention, including the distribution of corn and soybean planting in the corn belt states of the United States and the Phenocam sites.

[0034] Figure 3 This is a box plot of all samples between the Phenocam ground phenological observation values and the phenological estimation values of corn and soybeans in the embodiments and comparative examples of the present invention.

[0035] Figure 4 This is a scatter plot comparing the phenological detection results and phenological labels of corn and soybeans in each state of the corn belt in the embodiments and comparative examples of the present invention.

[0036] Figure 5 This is a curve showing the change of the proportion of the corn phenological period over time in the embodiments and comparative examples of the present invention in Kansas.

[0037] Figure 6 This is a curve showing the change of the proportion of the soybean phenological period over time in the embodiments and comparative examples of the present invention in Kansas. Detailed implementation mode

[0038] To make the objectives, technical solutions and beneficial effects of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0039] This embodiment provides a method for detecting the mid-season phenology of crops based on the time series of Sentinel-1 VH polarization signals over multiple years. Using the VH polarization time series data after multi-year averaging, through macroscopic and local stretching and translation calculations, a processed time series is generated to reduce the noise generated by environmental factors; through a local time window translation mechanism, the correlation between multiple processed time series and the Sentinel-1 VH polarization signal time series of the target pixel from the beginning of the year to the end time of mid-season detection is calculated, combining the SMF-S post-season phenology detection method with the optimal translation step length to obtain the mid-season phenology detection result, and finally taking the median of multiple mid-season phenology detection results to improve the stability of the final result; the specific process is as Figure 1 shown, including the following steps:

[0040] S1. Obtain the Sentinel-1 VH polarization signal time series VH of the target pixel of the crop to be detected (the crop under study) from the beginning of the current year (January 1st) to the end time of mid-season detection; set a spatial window centered on the target pixel, with the side length of the window set to 3 km, and obtain the Sentinel-1 VH polarization signal time series VH of multiple sampling pixels of the same crop under study within the spatial window for multiple previous years throughout the year target (t), where i is the sampling pixel number; calculate the average value of the Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years throughout the year to obtain the average Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years i (t), where i is the sampling pixel number; calculate the average value of the Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years throughout the year to obtain the average Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years The average Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years and the Sentinel-1 VH polarization signal time series VH of the target pixel target (t) is subjected to mean filtering;

[0041] S2. For each sampled pixel, the average Sentinel-1 VH polarization signal time series over multiple previous years Intercept the partial VH polarization signal time series from the beginning of the year to the detection cut-off time in the middle of the season and match it with the Sentinel-1 VH polarization signal time series VH of the target pixel target (t) for overall stretching and translation matching calculation;

[0042] Specifically, the translation and stretching processing is expressed as:

[0043] T = xscale×(t + tshift)

[0044] where T represents the time variable after translation and stretching processing, t represents the original time variable; tshift represents the time scale translation parameter, and the value range of tshift is [-30, 30]; xscale represents the time scale stretching parameter, and the value range of xscale is [0.9, 1.1];

[0045] Calculate the average overall stretched and translated VH time series over multiple previous years for each sampled pixel of the detected crop and the Sentinel-1 VH polarization signal time series VH of the target pixel target (T) to obtain the correlation coefficient value; select the predicted VH time series with the largest correlation coefficient as the preferred one to obtain the preferred predicted VH time series of the target pixel of the detected crop in the current year

[0046] Use the GMFR algorithm to calculate the partial average Sentinel-1 VH polarization signal time series of the sampled pixels from the beginning of the year to the detection cut-off time in the middle of the season after overall stretching and translation and the Sentinel-1 VH polarization signal time series VH of the target pixel target (T) to obtain the regression coefficients α and β, and use the regression coefficients to generate the predicted VH time series of the target pixel of the detected crop in the current year Specifically expressed as:

[0047]

[0048] Obtain the preferred predicted time series of the Sentinel-1 VH polarization signal of the target pixel The time T at which the maximum value of the VH polarization signal is locatedmax , when the maximum value of VH polarization signal is at time T max and the mid-season detection cut-off time t cut_off If they are the same, no further processing is performed; when the maximum value of the VH polarization signal occurs at time T max and the mid-season detection cut-off time t cut_off At different times, the time T when the VH polarization signal reaches its maximum value max The left and right time series of the center are locally stretched, and the MSD value of the stretched result and the Sentinel-1 VH polarization signal time series of the target pixel in the corresponding time period are calculated. The stretched result with the smallest MSD value is selected as the preferred one, and the left and right stretched results are recombined with the center to generate the preferred predicted VH time series of the target pixel in the current year. The specific local stretching is expressed as:

[0049] T′=xscale×|TT max |

[0050] Where T′ represents the time variable after local stretching processing, and the value range of the time scale stretching parameter xscale is [0.9, 1.1] with a step size of 0.05;

[0051] S3, obtain the optimal predicted VH time series of the target pixel generated by each sampling pixel in S2 for the current year, and convert the Sentinel-1 VH polarization signal time series VH of the target pixel from the beginning of the year to the mid-season detection cutoff time target (t) performing translation matching with the optimal predicted VH time series of the target pixel in the current year generated by each sampling pixel within a preset time window before the detection cutoff time, calculating the correlation coefficient, and selecting the optimal translation amount and its corresponding optimal predicted VH time series of the target pixel in the current year based on the correlation coefficient;

[0052] Specifically, the length of the translation matching time window is set to T match , set the maximum translation range and translation step size of translation matching;

[0053] The optimal prediction VH time series of the target pixel in the current year for the i-th sampling pixel [T end -T match ,T end ] is the starting time window for translation, T end is the detection deadline; for the jth translation matching, the corresponding time window (T match ) and compare it with the target pixel’s Sentinel-1 VH polarization signal time series VH target (T) In the time window [T end -Tmatch ,T end Match the VH timing data within [], and calculate the correlation coefficient R i,j ;

[0054] Set the quantity threshold N, and based on the correlation coefficient R i,j , select the top N maximum values, and record the translation amount corresponding to each maximum value (the product of the translation steps and the translation step size) and the preferred predicted VH time series of the target pixel for the current year n = 1, 2,..., N;

[0055] S4. Obtain the top N groups of preferred translation amounts with the highest correlation coefficients obtained in S3 and their corresponding preferred predicted VH time series of the target pixel for the current year. For each group of data, use the SMF-S method to perform post-season phenology detection on the preferred predicted VH time series of the target pixel for the current year to obtain the post-season phenology detection result and add the post-season phenology detection result to the preferred translation amount to obtain the corresponding in-season phenology detection result

[0056] S5. Take the median of the N in-season phenology detection results calculated in S4 to obtain the near-real-time in-season phenology detection result.

[0057] Next, taking American corn and soybeans as examples, the beneficial effects of the crop in-season phenology detection method based on multi-year Sentinel-1 VH polarization signal time series in the present invention will be described in detail.

[0058] In the planting areas of American corn and soybeans, the phenology detection experiment and test were carried out using the crop in-season phenology detection method based on multi-year Sentinel-1 VH polarization signal time series in the present invention, and the in-season SMF phenology detection algorithm group was used as a comparative example to conduct a comparative analysis with the in-season SMF final phenology detection result; as Figure 2 shown, the two test scenarios have different characteristics:

[0059] Test scenario 1: Based on the PhenoCam phenology observation network data of American corn and soybeans, the Sentinel-1 radar timing data was obtained using the single-site positioning mode, and the phenological period was interpreted in combination with the digital photos provided by the PhenoCam site as high-precision labeled data; this scenario focuses on small-scale and high-precision phenology detection and is suitable for the verification and optimization of the algorithm in local areas.

[0060] Test scenario 2: Taking the corn and soybean planting areas in 12 states in the US Corn Belt as the research scope, a regional phenological detection experiment at a large spatial scale was carried out; this scenario used the crop phenological period survey data (such as the proportion of crop area reaching a specific phenological period) in the Crop Progress Report released by the US Department of Agriculture (USDA) as labeled data, combined with Sentinel-1 radar time series data, to verify the applicability of the algorithm in a large-scale complex environment. The planting periods of corn and soybeans in the US usually start around May and end around November.

[0061] For quantitative evaluation, phenological detection was carried out using two methods: the algorithm proposed in the present invention and the mid-season SMF algorithm. The phenological detection results were compared with the labeled results, and the MAE value was calculated.

[0062] Quantitative evaluation was carried out on the US corn and soybean Phenocam sites. The test method was based on the algorithm proposed in the present invention and the mid-season SMF phenological detection algorithm, and the phenological detection accuracy was compared. As Figure 3 shown, it shows the scatter plot of all samples between the Phenocam ground phenological observation values and the phenological estimation values obtained by two mid-season phenological detection algorithms; the cut-off date was set to September 1st, and phenological detection was carried out on the R6 phenological period of corn and the R7 phenological period of soybeans respectively. Corn and soybeans generally reach this phenological period around mid-October. By averaging the MAE of all phenological periods, for corn, the near-real-time phenological detection MAE by the method of the present invention was 26.67 days for corn and 17.02 days for soybeans, while the corresponding MAE of the mid-season SMF phenological detection algorithm was 42.63 days for corn and 37.36 days.

[0063] From the results, in the phenological detection of the two algorithms, the predicted VH polarization time series data generated by the method of the present invention has stronger stability. Compared with the original data, the algorithm reduces the influence of low-frequency environmental noise and restores the regularity between the VH time series and the crop phenological growth cycle. At the same time, by combining the respective advantages of mid-season and post-season phenological detection, through local window translation matching and the SMF-S phenological detection algorithm, the instability of simple near-real-time detection is reduced. Compared with the simple mid-season SMF phenological detection algorithm, for both corn and soybean crops, the MAE is improved by more than 15 days. It can be seen from the scatter plot that due to the influence of non-growing season noise on mid-season detection, the phenological detection results of many Phenocam sites using the SMF mid-season detection algorithm have the same error, that is, the non-growing season noise causes the shape model to tend to stretch and translate to the left, and this result is finally obtained due to the influence of noise. However, this situation does not occur in the phenological detection results of the method of the present invention. This is because of two reasons. On the one hand, the predicted target pixel VH polarization time series data generated by the method of the present invention suppresses the non-growing season noise. On the other hand, the combination of post-season detection and mid-season detection improves the stability of the phenological detection results. At the same time, when Phenocam sites are spread throughout the United States, within a large and complex spatial range, it is relatively stable in the mid-season for all sample time series, suppressing the obstacles caused by multiple environmental factors to VH time series analysis and correcting some abnormal results of the SMF mid-season phenological detection.

[0064] For the phenological detection of corn and soybean in a large spatial range of 12 states in the US Corn Belt, based on the Cropland Data Layer data released by the US Department of Agriculture, 200 crop pixels are selected within the scope of each state as sample data; the phenological detection of the two algorithms, namely the SMF mid-season phenological detection algorithm and the method of the present invention, is carried out on all sample data. Based on the Crop Progress Report (CPR) provided by the US Department of Agriculture, taking 50% area ratio as the threshold, the date when the phenological area reaches 50% in the agricultural survey of the state is used as the label of the crop phenological period in the state, and the median of the final results of the phenological detection of each state is taken for comparison with the phenological period results. As Figure 4As shown, a scatter plot is presented, which shows the results of phenological detection of corn and soybeans in each state compared with the phenological labels provided by CPR. It can be seen from the figure that the mid-season phenological detection performance of the method of the present invention is better than that of the SMF mid-season phenological detection algorithm in both crops. In addition, the estimation of the method of the present invention for all samples is relatively stable, without outliers, that is, there are no obvious large errors, indicating that the present invention has robustness. Moreover, some points that were significantly outliers in the original data returned to the normal range after being processed by the method of the present invention. There is a significant improvement in the MAE of phenological detection. In terms of accuracy, comparing the results of the method of the present invention and the SMF mid-season phenological detection algorithm, the average MAE of phenological detection of soybeans and corn has decreased significantly (soybeans: 16.78 days vs. 10.02 days; corn: 18.73 days vs. 7.63 days). In terms of specific phenological periods, the MAE of phenological detection of the Mature period and Harvested period of corn in the method of the present invention is 8.25 days and 7.00 days, while the MAE in the SMF mid-season phenological detection is 18.58 days and 18.88 days; the MAE of phenological detection of the DroppingLeaves period and Mature period of soybeans in the method of the present invention is 9.88 days and 10.15 days, while the MAE of the SMF mid-season phenological detection is 16.13 days and 17.42 days. There is an obvious improvement in accuracy for both crops, and the abnormal outliers are corrected back to the normal range.

[0065] As Figure 5 , Figure 6 shown, the curves of the proportion of the crop area reaching this phenological period recorded in the CPR data of corn and soybean crops in Kansas over time, the curve of the proportion of the sample data detected by the algorithm of the present invention reaching this phenological period over time, and the curve of the proportion of the sample data of the SMF mid-season phenological detection reaching this phenological period over time are compared. By comparing the changing trends of the three curves, it is found that the changing trend of the proportion of the sample data detected by the algorithm of the present invention reaching this phenological period is more consistent with the proportion of the corresponding phenological period area surveyed by the CPR data, and can reflect the growth phenological cycle of the crops in Kansas.

[0066] As described above, the above is only the specific implementation manner of the present invention. Any feature disclosed in this specification, unless specifically stated, can be replaced by other equivalent or similar-purpose alternative features; all the features disclosed, or all the steps in any method or process, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A method for detecting crop mid-season phenology based on multi-year VH polarization signal time series, characterized in that, Including the following steps: S1. Obtain the Sentinel-1 VH polarization signal time series VH(t) of the detected crop at the target pixel from the beginning of the current year to the detection cut-off time in the middle of the season; set a spatial window centered on the target pixel, and obtain the Sentinel-1 VH polarization signal time series VH(t) of multiple sampling pixels of the detected crop within the spatial window for multiple previous years throughout the year, where i is the sampling pixel number; calculate the average value of the Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years throughout the year to obtain the average Sentinel-1 VH polarization signal time series of the sampling pixel for multiple previous years target (t); set a spatial window centered on the target pixel, and obtain the Sentinel-1 VH polarization signal time series VH(t) of multiple sampling pixels of the detected crop within the spatial window for multiple previous years throughout the year, where i is the sampling pixel number i (t), where i is the sampling pixel number; calculate the average value of the Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years throughout the year to obtain the average Sentinel-1 VH polarization signal time series of the sampling pixel for multiple previous years The average Sentinel-1 VH polarization signal time series of each sampling pixel for multiple previous years and the Sentinel-1 VH polarization signal time series VH(t) of the target pixel target (t) are subjected to mean filtering S2. For the multi-year average Sentinel-1 VH polarization signal time series of each sampling pixel in previous years Intercept the partial VH polarization signal time series from the beginning of the year to the detection cut-off time in the middle of the season And translate and stretch-match it with the Sentinel-1 VH polarization signal time series VH of the target pixel target (t) to generate the preferred predicted VH polarization signal time series of the detected crop in the current year of the target pixel S3. For the time series of the preferred predicted VH polarization signals of the target pixel generated for each sampled pixel, perform translational matching on the Sentinel-1 VH polarization signal time series of the target pixel and the time series of the preferred predicted VH polarization signals of the target pixel generated for the sampled pixel within a preset time window before the detection cut-off time, calculate the correlation coefficient, and select multiple groups of preferred translation amounts and their corresponding time series of preferred predicted VH polarization signals according to the correlation coefficient; S4. For each group of preferred translation amounts and their corresponding time series of the preferred predicted VH polarization signals of the target pixel in the current year, use the SMF-S method to perform post-season phenological detection on the time series of the preferred predicted VH planned signal of the target pixel in the current year to obtain the post-season phenological detection result, and add the post-season phenological detection result to the preferred translation amount to obtain the corresponding in-season phenological detection result; S5. Take the median of the calculated multiple groups of in-season phenological detection results to obtain the near-real-time in-season phenological detection result.

2. The method for detecting crop mid-season phenology based on multi-year VH polarization signal time series according to claim 1, wherein The specific process of step S2 is as follows: The time series of the Sentinel-1 VH polarization signals averaged over multiple previous years for each sampled pixel Intercept the partial VH polarization signal time series from the beginning of the year to the detection cut-off time in the middle of the season And perform overall translation and stretching processing on it and the Sentinel-1 VH polarization signal time series VH of the target pixel target (t), and correspondingly obtain the partial time series of the Sentinel-1 VH polarization signals averaged over multiple previous years of the sampled pixel And the Sentinel-1 VH polarization signal time series VH of the target pixel target (T), and obtain it according to the maximum value of the correlation coefficient Calculating the multi-year average Sentinel-1 VH polarization signal time series of some previous years for the sampled pixel using the GMFR algorithm and the Sentinel-1 VH polarization signal time series VH target (T) of the target pixel to obtain the regression coefficients α and β, and generating the predicted VH polarization signal time series for the current year of the target pixel for detecting crops using the regression coefficients Specifically expressed as: Obtain the preferred predicted VH polarization signal time series of the detected crop in the current year of the target pixel Obtain the optimal predicted time series of the Sentinel-1 VH polarization signal for the target pixel The time T corresponding to the maximum value of the VH polarization signal max , when the time T corresponding to the maximum value of the VH polarization signal max is different from the mid-season detection cut-off time t cut_off , centered on the time T corresponding to the maximum value of the VH polarization signal max , perform local stretching on the time series to the left and right of the center, and calculate the MSD value between the stretched result and the Sentinel-1 VH polarization signal time series of the target pixel in the corresponding time period. Select the stretched result with the minimum MSD value as the optimal one, and recombine the left and right stretched results with the center to generate the optimal predicted VH time series of the target pixel for the current year 3. The method for detecting crop mid-season phenology based on multi-year VH polarization signal time series according to claim 2, wherein In step 2, the overall stretching and translation processing is specifically expressed as: T = xscale × (t + tshift) where T represents the time variable after translation and stretching, t represents the original time variable; tshift represents the time scale translation parameter, and xscale represents the time scale stretching parameter.

4. The method for detecting crop mid-season phenology based on multi-year VH polarization signal time series according to claim 2, wherein In step S2, the local stretching is expressed as: T′ = xscale × |T - T max | where T′ represents the time variable after local stretching, T represents the time variable after translation and stretching, and xscale represents the time scale stretching parameter.

5. The method for detecting crop mid-season phenology based on multi-year VH polarization signal time series according to claim 1, wherein The specific process of step S3 is as follows: Set the length of the translation matching time window to T match , and set the maximum translation range and translation step size for translation matching; Optimal predicted VH polarization signal time series for the target pixel of the current year for the i-th sampled pixel With [T end -T match , T end as the starting time window for translation, where T end is the detection cut-off time; For the j-th translational matching, the time series data of the preferably predicted VH polarization signal within the corresponding time window (T match ) is intercepted and matched with the VH polarization signal time series VH target (T) of the target pixel within the time window [T end -T match , T end , and the correlation coefficient R i,j is calculated; Set a quantity threshold N and select the top N maximum values according to the correlation coefficient R i,j , and record the translation amount corresponding to each maximum value as the preferred translation amount and record the preferred predicted VH polarization signal time series of the corresponding target pixel for the current year

Images