A vegetation index prediction method for crops
Through high-spatial-temporal resolution data and multivariate linear fitting methods, the historical crop growth curve was extracted and analyzed, and the shortcomings of future growth and yield prediction of crops were solved, and the near-real-time vegetation index prediction of winter wheat was achieved throughout the growth period was improved, and precise agriculture and food were improved. Security support.
Patent Information
- Application Number
- CN202410494901.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-24
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-04-24
AI Technical Summary
It is difficult for the existing technology to accurately predict the future growth and yield of crops, and the time series prediction of vegetation index is insufficient in agricultural production, affecting precise agriculture and food security.
High-temporal and spatial resolution data are used to extract historical crop growth curves, and the vegetation index of the whole growth period is obtained through translation and fitting, and the correlation between meteorological data and prediction accuracy is analyzed using multiple linear fitting, providing growth curve selection criteria to achieve near real-time Vegetation index forecast.
The prediction of the near-real-time vegetation index of winter wheat throughout the breeding period has been achieved, the prediction accuracy and application prospects have been improved, and the prediction index prediction with high temporal and spatial resolution can be achieved vegetation index prediction 30 days before the harvest period, supporting precise agricultural decision-making and food security.
Smart Images

Figure CN118411601B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of agricultural remote sensing, and in particular to a vegetation index prediction method for crops. Background Art
[0002] Grasping agricultural trends as early as possible during the crop growth cycle and obtaining information on crop growth status in advance are of great significance for agricultural production and guidance. Precision agriculture can help improve crop productivity and maximize the efficiency of input resources. One of the biggest challenges in precision agriculture is predicting the future growth and yield of crops. Knowing crop growth and yield earlier can make accurate management models, planting decisions and macro-control possible.
[0003] Remote sensing technology uses crop spectral reflectance, radiation and backscatter coefficients measured by sensors to establish models related to crop physiological and biochemical parameters, phenological information, structural parameters and yield, which can achieve large-scale agricultural monitoring and is widely used as an important technical means in precision agriculture. Vegetation index (VI) is a mathematical combination designed and obtained based on the reflectivity of vegetation to solar radiation in different bands. The reflectivity of vegetation in different bands will vary due to changes in its growth status and morphological structure. Therefore, in remote sensing agricultural monitoring, vegetation index is often used as an important indicator of crop growth status. It is used to monitor crop phenology and growth, and is also an important parameter for estimating crop yield based on different crop models. Compared with single-phase vegetation indices, continuous vegetation index time series are more capable of reflecting the daily dynamic changes of vegetation, and have great application potential in real-time monitoring of crops and agricultural management decisions. For example, time series curves are more advantageous in phenological monitoring, and multi-phase vegetation indices can better reflect the growth changes of crops throughout the growth cycle. At the same time, using the vegetation index time series of the entire growing season has higher accuracy in yield estimation and phenological monitoring.
[0004] In fact, the future growth and final yield of crops are often more concerned by producers and have more practical application value. Therefore, the prediction of vegetation index time series is more practical in agricultural production. Accurate VI prediction data can be used to accurately predict crop phenology and yield, provide data support for agricultural production and life and national food policies, and adjust planting strategies and food reserves in time according to the prediction results, so as to ensure national food security. Summary of the invention
[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a crop-oriented vegetation index prediction method which realizes the near real-time prediction of the winter wheat vegetation index throughout the growth period.
[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:
[0007] A method for predicting vegetation index for crops is provided, which comprises the following steps:
[0008] S1. Extract the crop growth curve of pure crop pixels in historical years through high temporal and spatial resolution data, and process the abnormal values and missing values in the curve;
[0009] S2. The crop growth curve is translated and continuously fitted with the vegetation index of the early crop growth period in the predicted year, and the curve whose determination coefficient of the fitted curve and the early vegetation index is greater than the mean is selected for secondary fitting. The fitted curve is the predicted vegetation index of the whole growth period;
[0010] S3, use the later real high spatial resolution data to evaluate the accuracy of the predicted vegetation index;
[0011] S4. Use the multivariate linear fitting method to analyze the correlation between meteorological data and prediction accuracy, and provide the best standard for growth curve selection.
[0012] Furthermore, the specific method for extracting the crop growth curve of the pure crop pixels in the historical years in step S1 is:
[0013] The MODIS MCD43A4 product was called in Google Earth Engine (GEE) to calculate the daily NDVI of winter wheat from January to June in the forecast year of the study area. The calculation method of NDVI is based on the following formula:
[0014]
[0015] B1 and B2 are the reflectances of the red band and near-infrared band of MODIS data, respectively. Due to the abnormal data or missing data caused by clouds in MODIS data, outlier correction and null value filling were performed, and the SG (Savitzky-Golay Filter) with a window of 5 was used to smooth the extracted crop growth curve.
[0016] Furthermore, the specific method of step S2 is:
[0017] It is assumed that there is a linear relationship between the vegetation index growth curve of the same crop throughout the growing season and the vegetation index at high spatial resolution:
[0018] Y=aM(x+Δx)+b
[0019] Where x represents the day of the year, Y represents the growth curve extracted from the pure pixel, M(x) represents the vegetation index of the xth day of a pixel with high spatial resolution, Δx represents the phenological difference caused by different years and geographical locations, and a and b are fitting parameters. At the same time, by continuously translating each growth curve and calculating the fitting parameters of each translation, all curves with a determination coefficient greater than the mean after translation are found, and the optimal growth curve is obtained by averaging these curves. The optimal growth curve is used to fit the pixel, so as to predict the NDVI time series curve of the pixel in the late growing season. However, due to the small number of high-resolution vegetation indices in the early stage of crop growth, fitting errors may occur. Therefore, when fitting, we add adaptive threshold restrictions to a and b: a represents the degree to which the optimal growth curve is stretched up and down, so a is a number greater than 0; b represents the size of the curve being translated up and down; Δx represents the size of the left and right translation. By observing the growth curve, the NDVI of the study area is distributed in the range of [0.2,1], so when NDVI is input, the maximum value NDVI among them max and minimum NDVI min This determines that the curve can only be shifted upward by a maximum of 1-NDVI max Or pan down the NDVI min -0.2 range, so the range of b obtained by fitting each pixel is [0.2-NDVI min ,1-NDVI max ], and because the fitted curve should also be between [0.2,1], and so the range of a is Therefore, if the fitted a and b exceed this range during fitting, we directly assign them to the maximum or minimum value to ensure that the growth curve determination coefficient R is selected when the predicted NDVI is reasonable. 2 maximum.
[0020] Furthermore, the specific method of step S3 is:
[0021] In order to verify the accuracy of NDVI prediction using crop growth curves, we compared the actual NDVI of the later period with the predicted results of the same day, and the root mean square error correlation coefficient was used to evaluate the prediction accuracy:
[0022]
[0023] NDVI in the formula i is the actual high spatial resolution NDVI value in the later period, ndvi i is the predicted NDVI value, and n is the number of all pixels used to test the accuracy. We used all the clear NDVI that could be obtained in the later stage to verify the accuracy, so as to compare the prediction ability of the algorithm in different periods.
[0024] Furthermore, the specific method of step S4 is:
[0025] The accumulated temperature, accumulated precipitation and accumulated net surface radiation of winter wheat in the entire growth stage within the range of the growth curve were counted. At the same time, the cumulative values of the three meteorological elements in the predicted study area were counted, and the differences between the accumulated values of the three meteorological data in different years and regions and the predicted year were calculated, and the differences were normalized. The normalized values were defined as: ΔT, ΔP and ΔR, respectively. Through multiple linear regression, the mean A of all correlation coefficients when using NDVI before the jointing stage to predict the heading and filling period was used as the prediction accuracy of the year in the region, and the following linear model was established; the multiple linear regression model of this study was obtained based on R language.
[0026] A=sΔT+jΔP+kΔR+p
[0027] Using the coefficient of determination R 2 , RE and RMSE are used to evaluate the accuracy of the multivariate linear regression model, where R 2 and RE is defined as:
[0028]
[0029] Among them A i is the actual prediction accuracy, is the prediction accuracy of the model, is the mean prediction accuracy.
[0030] The beneficial effects of the present invention are as follows: based on the historical crop growth curve, the present invention achieves near real-time prediction of the vegetation index in the late growth period for the entire growth period by fitting with the vegetation index with high spatial resolution, and can obtain vegetation index data with high temporal and spatial resolution, evaluate the prediction accuracy, and provide the best selection strategy for crop growth curves based on meteorological differences. The algorithm can be used for near real-time vegetation index prediction over a large range and throughout the growth period, and has good application prospects in actual agricultural forecasting and yield forecasting. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 is a schematic diagram of the process of this method;
[0032] Figure 2 is the spatial distribution of the study area;
[0033] Figure 3 Spatial positions extracted for crop growth curves;
[0034] Figure 4 For the extracted multi-year crop growth curve;
[0035] Figure 5To determine the accuracy of predictions using crop growth curves for the forecast year;
[0036] Figure 6 is the trend of prediction accuracy with the interval date;
[0037] Figure 7 The change of prediction accuracy at different interval dates in Hengshui City;
[0038] Figure 8 This is the growth curve of crops from 2018 to 2021. DETAILED DESCRIPTION
[0039] The present invention is described in detail below in conjunction with specific embodiments.
[0040] like Figure 1 As shown in FIG. 1 , the crop vegetation index prediction method includes the following steps:
[0041] S1. Using high temporal and spatial resolution MODIS MCD43A4 data, crop growth curves of pure crop pixels in the entire growth cycle of historical years (nearly five years) are extracted, and outliers and missing values in the curves are processed using SG filtering;
[0042] S2. The crop growth curve and the vegetation index of the early crop growth in the predicted year are translated and continuously fitted to obtain fitting parameters, and the determination coefficient of the curve obtained by fitting the fitting parameters and the early vegetation index is calculated. The curve whose determination coefficient of the fitted curve and the early vegetation index is greater than the mean is selected for secondary fitting. The fitted curve is the predicted vegetation index for each day of the whole growth period;
[0043] S3, using the later real high spatial resolution data, the determination coefficient and root mean square error are used as evaluation indicators to evaluate the accuracy of the vegetation index of all crop pixels in the prediction area;
[0044] S4. Use the multivariate linear fitting method to analyze the correlation between accumulated temperature, precipitation and radiation in meteorological data and prediction accuracy, establish a multivariate linear model of prediction accuracy and meteorological difference, and give the best standard for selecting growth curves. Based on this standard, select the best crop growth curve for predicting vegetation index when making predictions in the required area.
[0045] The accumulated temperature, accumulated precipitation and accumulated net surface radiation of winter wheat in the entire growth stage within the range of the growth curve were counted. At the same time, the cumulative values of the three meteorological elements in the predicted study area were counted, and the differences between the accumulated values of the three meteorological data in different years and regions and the predicted year were calculated, and they were normalized and defined as: ΔT, ΔP and ΔR, respectively. Through multiple linear regression, the mean A of all correlation coefficients when using NDVI before the jointing stage to predict the heading and filling period was used as the prediction accuracy of the year in the region, and the following linear model was established; the multiple linear regression model of this study was obtained based on R language.
[0046] A=sΔT+jΔP+kΔR+p
[0047] Using the coefficient of determination R 2 , RE and RMSE are used to evaluate the accuracy of the multivariate linear regression model, where R 2 and RE is defined as:
[0048]
[0049] Among them A i is the actual prediction accuracy, is the prediction accuracy of the model, is the mean prediction accuracy.
[0050] In one embodiment of the present invention, three test areas were selected, namely Hengshui City, Huaibei City and Zhoukou City, and 2021 was selected as the prediction year. After balancing the influence of clouds and the size of the study area, 500*500, 1000*1000, and 1000*1000 areas were selected as test areas. The random forest algorithm was used to map winter wheat, and only winter wheat pixels were predicted and verified for accuracy. The spatial distribution is shown in Figure 2. Figure 2 Crop growth curves were randomly extracted in the study area and its surrounding areas. The spatial distribution of the extraction points is shown in Figure 3 The extracted crop growth curve is shown in Figure 4 To test the best prediction ability of the algorithm, the growth curve of the prediction year was used to predict the NDVI of local winter wheat. The prediction accuracy varies with the time of the early NDVI input. Figure 5As shown. It can be seen that this method can achieve near-real-time vegetation index prediction for the entire growth period. The prediction accuracy of the vegetation index in the later period gradually increases with the increase of the input real vegetation index, indicating that as the crop growth period progresses, the remote sensing data obtained by this method increases, and the prediction ability of the vegetation index in the entire growth period continues to increase. In Hengshui area, after using the MODIS growth curve to input the NDVI when DOY=122, the prediction accuracy of DOY=147 is R=0.5602, RMSE=0.1187; in Huaibei area, after using the MODIS growth curve to input the NDVI when DOY=81, the prediction accuracy of DOY=129 is R=0.5705, RMSE=0.0459; in Zhoukou area, after using the MODIS growth curve to input the NDVI when DOY=99, the prediction accuracy of DOY=129 is R=0.4016, RMSE=0.0647. In summary, this prediction method has the ability to predict the vegetation index for about 30 days after the middle and late stages of winter wheat, and can achieve high temporal and spatial resolution vegetation index prediction for the entire growth period of winter wheat about 30 days before the harvest period, which greatly improves the possibility of accurate prediction of yield. At the same time, the comparison results of the actual NDVI and the predicted NDVI in the entire study area are as follows Figure 6 Shown
[0051] The Hengshui study area is not in the overlapping area of satellite image strips, so all high spatial resolution Sentinel-2 vegetation index data intervals are 5 or multiples of 5, which can be used to analyze the change of prediction accuracy with the gap date (Gapday) so that there is sufficient and uniform data averaging. Therefore, taking Hengshui as an example, the prediction accuracy of all the same gap dates is averaged, and the changes in prediction accuracy of different gap dates are compared. The results are as follows: Figure 7 As shown. Starting from the input of only 5 scenes (DOY=32) to the input of 18 scenes (DOY=142), the correlation coefficient R and root mean square error RMSE of all Gap days of 5, 10, 15, 20, 25, 30, 35, and 40 were averaged. The results show that when the Gap day is 5 days, the average R value is 0.8352, and the RMSE is 0.0619 at this time. With the increase of Gap day, when the Gap day is 40, the R value is 0.4742 and the RMSE is 0.1176. This method can achieve NDVI prediction results with high temporal and spatial resolution (10 meters, every day). The NDVI 30 days after the prediction can still maintain a high prediction accuracy (R can be maintained at about 0.6, and RMSE is less than 0.1), which can play a greater role in crop monitoring, especially in areas with complex planting conditions.
[0052] The crop growth curves vary in different years and regions. The growth curves from 2018 to 2021 are compared. Figure 8As shown. From the shape of the curve, it can be seen that compared with the other two regions, Hengshui's phenological period is about 15 days later, and there is a significant decrease in NDVI during the wintering period. This is because winter wheat in Hengshui will show a more obvious yellowing phenomenon in order to successfully overwinter, and Hengshui is located in the northern part of the North China Plain, with more frequent snowfall in winter, resulting in significant differences in NDVI in the three prediction areas during this period. At the same time, even in the same prediction area, there are differences in the shape of the NDVI curve, such as delayed or early phenological periods. This may lead to differences in prediction accuracy, so how to select the crop growth curve (time, location) is a key factor in determining whether the method proposed in this study can be used on a large scale.
[0053] Therefore, the crop growth curves extracted from different study areas in different years were used crosswise, and the accuracy was compared.
[0054] As shown in Table 1.
[0055] Table 1 Comparison of prediction accuracy of different growth curves for the three study areas
[0056]
[0057] From the prediction accuracy, we can see that the local growth curve of the current year shows better results in the prediction, which means that the local growth curve of the current year can better match the shape of the NDVI curve of the predicted year, but this is not possible in actual prediction, so we need to find the growth curve that best matches the predicted year in the historical growth curve. The mean A of all correlation coefficients of the NDVI before the jointing stage to predict the heading and filling stage is used as the prediction accuracy A of the region in that year. In this regard, we analyzed the correlation between individual meteorological elements and prediction accuracy A, and used the multivariate linear regression method to combine the differences of three meteorological elements to perform correlation analysis on prediction accuracy A. The analysis results are shown in Table 2.
[0058] Table 2 Linear relationship and determination coefficient R between prediction accuracy A and meteorological element difference in the three study areas 2
[0059]
[0060]
[0061] When the accumulated temperature difference, precipitation difference and radiation difference are used alone to fit the prediction accuracy, their determination coefficients R 2They are 0.433, 0.5495 and 0.4828 respectively. At this time, the determination coefficient of precipitation difference is the highest, but in general, their differences are relatively less explainable for prediction accuracy. A large number of studies have shown that the impact of meteorological conditions on vegetation index changes is comprehensive, not a single factor independently affected. Therefore, this study uses the method of multiple linear regression to establish the relationship between the difference in meteorological elements and prediction accuracy. The multi-source linear model finally established is:
[0062] A=-0.1324ΔT-0.10958ΔP-0.11957ΔR+0.58764
[0063] Among them, ΔT, ΔP and ΔR represent the differences in accumulated temperature, precipitation and radiation respectively. The significance test P values of the three meteorological elements are 0.0325*, 0.1722 and 0.1302 respectively. The determination coefficient R 2 The influence of the accumulated temperature difference on the prediction accuracy in the model is significant, which shows that the influence of the accumulated temperature difference on the prediction accuracy is independent and significant, which may be because the accumulated temperature is one of the most important meteorological factors affecting crop phenology. Compared with the use of a single meteorological factor, the model combining multiple factors improves the model's ability to explain the prediction accuracy. The established multi-source linear regression model can explain 63.07% of the accuracy changes and has lower RMSE and RE, indicating that the model can be used to refer to the historical years and spatial ranges of the crop growth curve.
[0064] In summary, this method predicts winter wheat NDVI based on crop growth curves. This method uses the characteristics of growth curves that contain information such as crop phenology and growth potential to achieve the prediction of NDVI time series curves during the growing period of winter wheat (the accuracy is 5 days, the R mean is 0.8352, and the RMSE is 0.0619. As the interval increases, when the interval is 40 days, the R value is 0.4742, and the RMSE is 0.1176.). At the same time, the best strategy for selecting growth curves is given: it is recommended to select the historical parameter growth curve with the highest simulation prediction accuracy according to the multivariate linear regression model during the growing period of winter wheat in the weather forecast of the prediction area for NDVI prediction, so as to give full play to the prediction ability of this method and obtain the best prediction accuracy.
[0065] This method can achieve accurate prediction of the near-real-time NDVI time series of the entire growth period at least 30 days before the harvest period, and obtain the NDVI data of each day with a spatial resolution of 10m in the later period. This method can also be extended to predict time series curves of different vegetation indices and different crops. The predicted high-temporal and spatial resolution time series curves can be applied to existing crop phenology monitoring and yield estimation methods, improve the timeliness of existing algorithms or models, and help achieve near-real-time prediction of phenology and yield throughout the growth period.
[0066] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.
Claims
1. A vegetation index prediction method for crops, characterized in that: The following steps are involved: S1. Extract the crop growth curve of pure crop pixels in historical years through high temporal and spatial resolution data, and process the abnormal values and missing values in the curve; S2. The crop growth curve is translated and continuously fitted with the vegetation index of the early crop growth period in the predicted year, and the curve whose determination coefficient of the fitted curve and the early vegetation index is greater than the mean is selected for secondary fitting. The fitted curve is the predicted vegetation index of the whole growth period; S3, use the later real high spatial resolution data to evaluate the accuracy of the predicted vegetation index; S4. Use the multivariate linear fitting method to analyze the correlation between meteorological data and prediction accuracy, give the best standard for selecting the growth curve in the study area, and select the growth curve in the prediction year according to this standard. The best standard is: select the historical reference growth curve with the highest simulation prediction accuracy in the winter wheat growth period in the weather forecast of the prediction area in the current year according to the multivariate linear regression model for NDVI prediction; The specific method of step S4 is: The accumulated temperature, accumulated precipitation and accumulated net surface radiation of crops in the entire growth stage within the selected range of the growth curve were statistically analyzed. At the same time, the cumulative values of the three meteorological elements in the predicted study area were statistically analyzed. The differences between the accumulated values of the three meteorological data in different years and regions and the predicted year were calculated, and they were normalized and defined as: ΔT, ΔP and ΔR. The multivariate linear regression was established to use the NDVI before the jointing stage to predict the heading and filling stage. The mean A of all correlation coefficients was used as the prediction accuracy of the year in the region, and the following linear model was established: A=sΔT+jΔP+kΔR+p The multiple linear regression model was obtained based on R language, using the coefficient of determination R 2 , RE and RMSE are used to evaluate the accuracy of the multivariate linear regression model, where R 2 and RE is defined as: Among them A i is the actual prediction accuracy, is the prediction accuracy of the model, is the mean of the prediction accuracy. When making predictions again, the growth curve can be selected based on the multivariate linear regression model.
2. A method for predicting vegetation index for crops according to claim 1, characterized in that: The specific method for extracting the crop growth curve in step S1 is: The MODIS MCD43A4 product was called in Google Earth Engine (GEE) to calculate the daily NDVI of winter wheat from January to June in the forecast year of the study area. The calculation method of NDVI is based on the following formula: B1 and B2 are the reflectances of the red band and near-infrared band of MODIS data, respectively. Due to the data anomalies or missing data caused by clouds in MODIS data, outlier correction and null value filling were performed, and the SG (Savitzky-Golay Filter) with a window of 5 was used to smooth the extracted crop growth curve.
3. The crop vegetation index prediction method according to claim 1, characterized in that: The specific method of step S2 is: It is assumed that there is a linear relationship between the vegetation index growth curve of the same crop throughout the growing season and the vegetation index at high spatial resolution: Y=aM(x+Δx)+b Where x represents a day of the year, Y represents the growth curve extracted from the pure pixel, M(x) represents the vegetation index of a pixel on the xth day with high spatial resolution, Δx represents the phenological difference caused by different years and geographical locations, and a and b are fitting parameters; at the same time, by continuously translating each growth curve and calculating the fitting parameters of each translation, all curves with a determination coefficient greater than the mean after translation are found, and the optimal growth curve is obtained by averaging these curves. The optimal growth curve is used to fit the pixel, so as to predict the NDVI time series curve of the pixel in the late growing season; adaptive threshold restrictions are added to a and b during fitting: a represents the degree to which the optimal growth curve is stretched up and down, so a is a number greater than 0; b represents the size of the curve being translated up and down; Δx represents the size of the left and right translation; by observing the growth curve, the NDVI of the study area is distributed in the range of [0.2,1], so when NDVI is input, the maximum value NDVI among them max and minimum NDVI min This determines that the curve can only be shifted upward by a maximum of 1-NDVI max Or pan down the NDVI min -0.2 range, so the range of b obtained by fitting each pixel is [0.2-NDVI min ,1-NDVI max ], and because the fitted curve should also be between [0.2,1], and so the range of a is If the fitted a and b exceed this range during fitting, they are directly assigned the maximum or minimum value to ensure that the growth curve determination coefficient R is selected when the predicted NDVI is reasonable. 2 maximum.
4. The method for predicting vegetation index for crops according to claim 1, characterized in that: The specific method of step S3 is: In order to verify the accuracy of NDVI prediction using crop growth curves, the actual NDVI at a later time was compared with the predicted results on the same day, and the root mean square error and correlation coefficient were used to evaluate the prediction accuracy: NDVI in the formula i is the actual high spatial resolution NDVI value in the later period, ndvi i is the predicted NDVI value, n is the number of all pixels used to test the accuracy, and all the clear NDVI that can be obtained in the later stage are used for accuracy verification to compare the prediction ability of the algorithm in different periods.
