Rice yield prediction method based on unmanned aerial vehicle remote sensing image processing

Abstract

The present invention relates to the technical field of rice yield prediction. Disclosed is a rice yield prediction method based on unmanned aerial vehicle remote sensing image processing. In the method, vegetation indices (VIs), spectral indices, and color indices of rice at different growth stages are collected by an unmanned aerial vehicle, and the rice yield is predicted by using single-date VIs and multi-date VIs based on multispectral (MS) and digital (RGB) images. Both MS and RGB digital sensors on an unmanned aerial vehicle are reliable platforms for rice growth and yield prediction, and it is feasible to estimate the rice yield on the basis of dual-sensor data and temporal accumulation across different growth stages, so that full-growth-stage monitoring of rice can be performed, the influence of spectral indices and color indices at different growth stages on the rice yield can be evaluated, and the pre-harvest grain yield can be predicted in a timely and accurate manner, thereby facilitating rice production management and grain trade.

Description

A rice yield prediction method based on unmanned aerial vehicle remote sensing image processing TECHNICAL FIELD

[0001] The present application relates to the technical field of rice yield prediction, in particular to a rice yield prediction method based on unmanned aerial vehicle remote sensing image processing. BACKGROUND

[0002] The development of remote sensing technology makes it possible to predict crop yield. Satellite remote sensing can be used for large-scale crop growth monitoring and has been widely applied. However, the low resolution, high cost and variable weather conditions limit the ability to obtain crop growth information in a timely manner throughout the growing period. In addition, many prediction models can only provide high-precision crop yield prediction at a large scale such as a country, province or county, and cannot describe detailed changes at a relatively small scale. In recent years, the availability of unmanned aerial vehicles (UAVs) has increased, providing potential for improving the monitoring frequency and spatial resolution of image data. Vegetation indices (VIs) calculated based on UAV images have been proven to be an effective method for crop yield prediction. Du and Noguchi (2017) used VIs calculated based on RGB images from UAVs to assess the yield of wheat from the heading stage to the mature stage. Rapid monitoring of the normalized difference vegetation index (NDVI) of wheat at the grain filling stage using a multispectral UAV platform found that NDVI was highly correlated with grain yield (Hassan et al, 2019). Gong et al. (2018) applied a set of VIs from UAV multispectral images to predict the yield of oilseed rape at the early flowering stage. Combinations of VIs calculated from RGB and multispectral images were also used to assess the yield of barley at the booting stage (Kefauver et al, 2017). These studies confirmed that using appropriate vegetation indices can more effectively predict yield compared to traditional methods.

[0003] In the 1990s, J. Peñuelas applied visible spectrum (RGB) technology to plant phenotyping to identify physiological changes caused by water and nitrogen stress, demonstrating the possibility of using visible spectrum information to evaluate vegetation physiological traits. In the late 1990s, plants were distinguished using remote sensing and multispectral (MS) technology, and the relationship between MS image information and important parameters such as leaf area index and growth rate was explored. However, some studies have shown that using a single sensor to estimate specific phenotypic traits of crops has some limitations. Therefore, researchers have gradually begun to fuse and analyze sensor-based data such as RGB and MS to improve the estimation accuracy of phenotypic traits such as chlorophyll, aboveground biomass and yield. This combined technology has been applied to crops such as wheat, corn, soybeans and barley. However, there are still limited studies on chickpea involving dual-sensor fusion. Therefore, this study aims to fill this research gap and explore the impact of dual-sensor data on chickpea yield prediction models.

[0004] One of the biggest challenges of the 21st century is to increase crop yield to meet the growing demand for agricultural products by the world's population. Rice is an important crop for global food security and the main food source for more than 3 billion people. Timely and accurate rice monitoring and yield prediction before harvest has important value for precision management, policy decision and market marketing. Field investigation and destructive sampling at different field scales are commonly used methods to collect rice yield data, but these methods are usually labor-intensive, high-cost and uncertain. In addition, as a long-cycle crop, the yield of rice is not only related to the growth conditions of the reproductive period (booting stage, heading stage, filling stage, maturation stage), but also related to the growth conditions of the nutrient period (tillering stage, jointing stage), so it is necessary to monitor the growth of rice throughout the whole growth period. Therefore, we propose a rice yield prediction method based on unmanned aerial vehicle remote sensing image processing. SUMMARY

[0005] The purpose of the present application is to solve the problem that, as a long-cycle crop, the yield of rice is not only related to the growth conditions of the reproductive period (booting stage, heading stage, filling stage, maturation stage), but also related to the growth conditions of the nutrient period (tillering stage, jointing stage), so it is necessary to monitor the growth of rice throughout the whole growth period. The present application provides a rice yield prediction method based on unmanned aerial vehicle remote sensing image processing.

[0006] In order to achieve the above purpose, the present application specifically adopts the following technical scheme:

[0007] A rice yield prediction method based on unmanned aerial vehicle remote sensing image processing, comprising the following steps:

[0008] S1, geographic coordinate correction, setting 5 ground control points GCP in the study area, and setting markers thereon, to correct the unmanned aerial vehicle images at different growth stages;

[0009] S2, unmanned aerial vehicle image acquisition, pre-setting an automatic flight plan, acquiring images with a forward overlap of 90% and a lateral overlap of 85%, and using the same flight path and camera settings throughout the growth period;

[0010] S3, unmanned aerial vehicle image processing, the pre-processing process includes noise correction, vignetting correction, lens distortion correction, band-to-band registration and radiation correction, the pre-processing of the RGB and MS orthographic images obtained by the unmanned aerial vehicle includes denoising, radiation calibration, band alignment and orthographic image correction, and the plant index, spectral index and color index corresponding to different growth periods of rice are obtained according to the processed images;

[0011] S4, at the physiological maturity stage of rice, 3 uniform and representative 0.5m double row areas are selected in each plot for sampling, and after the ear is cut, it is taken back to the laboratory for threshing, the grain is dried to a constant weight, the moisture content is measured and weighed, and the average value of the three sampling sub-areas is taken as the final yield of the plot, and the yield is converted to kg / ha;

[0012] S5, correlation analysis of rice yield and different growth period spectral / color index;

[0013] S6, correlation analysis of rice yield and different growth period multi-time spectral index;

[0014] S7, correlation analysis of rice yield and different growth period multi-time color index;

[0015] S8, correlation analysis of cumulative time series spectral index and color index combination and rice yield;

[0016] S9, according to the correlation analysis, the index or index combination with the highest correlation at different growth stages is taken as the independent variable, and the corresponding rice yield is taken as the dependent variable to establish a linear regression model and a multiple stepwise regression model, and the model with higher accuracy is determined as the rice yield prediction model through accuracy test;

[0017] S10, using the rice yield prediction model to predict the rice yield of the to-be-measured area.

[0018] Further, the geographical coordinates of the GCP in S1 are obtained by RTK-GPS (Real-Time Kinematic Global Positioning System, CHC X900 GNSS), and the horizontal and vertical errors are within 1cm and 2cm respectively.

[0019] Further, the automatic flight plan in S2 is pre-set by DJI GS pro software, and the unmanned aerial vehicle obtains images of the test area at a height of 30m in the key growth stage of rice, the images are stored in JPEG and TIFF formats, the flight speed is 2m / s, and 4 standard reflective cloths are placed on the ground for radiation correction of the unmanned aerial vehicle images; each flight activity is carried out under the condition of stable light from 11:00 to 14:00 in local time.

[0020] Further, the whole pretreatment process in S3 adopts ArcMap 10.2, a calibrated radiation measurement surface, which is placed in the flight area and allows the user to make radiation measurement calibration. In order to carry out radiation calibration, four calibration targets with nominal reflectivity of 5%, 10%, 20% and 40% are placed in the flight path of the unmanned aerial vehicle, and captured in the unmanned aerial vehicle image. The RGB and MS orthoimages obtained by the unmanned aerial vehicle are pretreated using Pix4Dmapper software.

[0021] Further, in S5-S8, the vegetation index VIs and DAT are normalized, so that the values of different types of VIs and DAT are in the range of (0, 1). By the dynamic curve of different growth stage spectral index or color index with DAT (days after transplanting), the spectral index and color index are defined as the closed region of the region.

[0022] Further, in S5, the correlation coefficient of most vegetation indices with yield increases first and then decreases with the growth of rice, and reaches the maximum in the booting stage and the heading stage. The correlation of most spectral indices with yield is low in the tillering stage, and the highest correlation spectral index in each period is OSAVI (0.716), NDRE (0.730), NDVIRE (0.751), MNDI (0.737), MNDI (0.705), NDVIRE (0.752), respectively. Among them, NDVIRE and MNDI reach the maximum correlation in the booting stage, the mature stage and the heading stage, and the filling stage, respectively. The color index performs well in the early growth stage, and VARI performs best in the jointing stage to the heading stage. The highest correlation color index in each period is ExGR (0.819), VARI (0.742), VARI (0.783), MGRVI (0.720), ExR (-0.697), and ExG (0.722), respectively.

[0023] Further, in S6, the cumulative time series spectral index correlation of the jointing stage, the booting stage and the heading stage is generally higher than that of other growth stages. The cumulative time series spectral index of the jointing stage to the booting stage has the highest correlation with rice yield, and the R2 value reaches 0.722. Compared with the single-stage prediction result, the modeling R2 of the cumulative time series spectral index is improved by 0.02, and the ∑VI model constructed by multi-stage data is mostly better than the single-stage model.

[0024] Further, in S7, the ∑VI model established in the tillering stage to the filling stage (B-IF) has the highest correlation with yield (R2=0.72), and compared with the single-stage color index, the color index prediction error of the validation data set is reduced by 33%.

[0025] Further, the spectral index from the jointing stage to the booting stage and the color index from the booting stage to the early filling stage in S8 have the best prediction accuracy (R2=0.752), and the RRMSE of the verification data is 0.09. Since ∑(VI&ABD) combines the best expression period of the spectral index before the booting stage with the best expression period of the color index after the booting stage, it is superior to the prediction of the spectral index and the color index in the whole growth period, and compared with the optimal spectral index combination model, the multi-stage and multi-variable combination improves the prediction of the rice yield by 0.03 and 0.025, respectively. The sensor fusion from the jointing stage to the booting stage improves the result based on one sensor by 7.3% and 7.2%, respectively.

[0026] Further, the unmanned aerial vehicle image acquisition in S2 is performed twice, and the top of the rice under the flight path of the unmanned aerial vehicle is coated with reflective paint. The flight height of the second unmanned aerial vehicle image acquisition is 5 m. When the unmanned aerial vehicle flies for the second time, the airflow of the unmanned aerial vehicle blows the rice on the path, and the rice is tilted. By comparing the images collected for the first time, the swing amplitude of the top of the rice is obtained. According to the correlation analysis of the swing amplitude in different remaining periods and the rice yield, the swing amplitude in the growth period with the highest correlation is added to the prediction model in S9 to assist in predicting the rice yield. Beneficial effects

[0027] The present application predicts the rice yield by using single-period vegetation index (VIs) and multi-period vegetation index (VIS) based on multi-spectrum (MS) and digital image (RGB). The color index (ExGR) has the highest correlation (0.819) in the tillering stage, and the spectral index (NDVIRE) has the highest correlation (0.752) in the booting stage. The cumulative time-series spectral index combination from the jointing stage to the booting stage has the highest correlation with the rice yield, and the R2 value reaches 0.722. The cumulative time-series color index model established from the tillering stage to the filling stage has the highest correlation (R2=0.727) with the yield. Compared with the single-variable combination, the multi-stage and multi-variable combination significantly improves the prediction of the rice yield. Among them, the spectral index combination from the jointing stage to the booting stage and the color index combination from the booting stage to the filling stage have the best prediction accuracy (R2=0.752), and the multi-stage and multi-variable combination improves the prediction of the rice yield by 0.03 and 0.025, respectively. The sensor fusion from the jointing stage to the booting stage improves the result based on one sensor by 7.3% and 7.2%, respectively. The digital sensors of MS and RGB on the unmanned aerial vehicle are reliable platforms for rice growth and yield prediction. Based on the data of double sensors and the time accumulation in different growth stages, it is feasible to estimate the rice yield, monitor the rice growth in the whole growth period, evaluate the influence of the spectral index and the color index in different growth stages on the rice yield, accurately predict the pre-harvest grain yield, and facilitate the management of rice production and grain trade. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a workflow diagram of the present application. DETAILED DESCRIPTION

[0029] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0030] Referring to Figure 1, the present application provides a rice yield prediction method based on unmanned aerial vehicle remote sensing image processing, comprising the following steps:

[0031] S1, geographic coordinate correction, 5 ground control points GCPs are set in the study area, and markers are set thereon to correct the unmanned aerial vehicle images at different growth stages;

[0032] S2, unmanned aerial vehicle image acquisition, an automatic flight plan is set in advance, images with a forward overlap of 90% and a lateral overlap of 85% are acquired, and the same flight path and camera setting are used throughout the growth period;

[0033] S3, unmanned aerial vehicle image processing, the pre-processing process includes noise correction, vignetting correction, lens distortion correction, band-to-band registration and radiation correction, the RGB and MS orthographic images acquired by the unmanned aerial vehicle are pre-processed, the main process includes denoising, radiation calibration, band alignment and orthographic image correction, and the plant index, spectral index and color index corresponding to different growth periods of rice are obtained according to the processed images;

[0034] S4, at the physiological maturity stage of rice, 3 uniform and representative 0.5m double-row areas are selected for sampling in each plot, the heads are cut off after sampling and taken back to the laboratory for threshing, the grain is dried to a constant weight, the moisture content is measured and weighed, the average value of the three sampling sub-areas is taken as the final yield of the plot, and the yield is converted to kg / ha uniformly;

[0035] S5, correlation analysis of the relationship between rice yield and different growth period spectral / color index;

[0036] S6, correlation analysis of the relationship between rice yield and different growth period multi-temporal spectral index;

[0037] S7, correlation analysis of the relationship between rice yield and different growth period multi-temporal color index;

[0038] S8, correlation analysis of the relationship between the combination of cumulative time series spectral index and color index and rice yield;

[0039] S9, according to the correlation analysis, the indexes or index groups with the highest correlation in different growth periods are combined as independent variables, and the corresponding rice yield is taken as the dependent variable to establish a linear regression model and a multiple stepwise regression model, and the model with higher precision in the linear regression model and the multiple stepwise regression model is determined as the rice yield prediction model through precision test;

[0040] S10, the rice yield of the to-be-measured area is predicted by using the rice yield prediction model.

[0041] In this embodiment, preferably, the geographic coordinates of the GCP in S1 are obtained by RTK-GPS (Real-Time Kinematic Global Positioning System, CHC X900 GNSS), and the horizontal and vertical errors are within 1 cm and 2 cm, respectively.

[0042] In this embodiment, preferably, the image data in S2 is collected by DJI Mavic 2 unmanned aerial vehicle equipped with RGB lens and DJI Phantom 4 RTK unmanned aerial vehicle equipped with multispectral camera with central wavelength of 450, 560, 650, 730 and 840 nm, the automatic flight plan is set in advance by DJI GS pro software, the unmanned aerial vehicle obtains the image of the test area at a height of 30 m in the key growth stage of rice, the image is stored in JPEG and TIFF format, the flight speed is 2 m / s, and four standard reflective cloths are placed on the ground for radiometric correction of the unmanned aerial vehicle image; each flight activity is carried out under stable light conditions from 11:00 to 14:00 in local time.

[0043] In this embodiment, preferably, the entire preprocessing process in S3 uses ArcMap 10.2, a calibrated radiometric surface is placed in the flight area, and the user is allowed to make radiometric calibration, in order to perform radiometric calibration, four calibration targets with nominal reflectivity of 5%, 10%, 20% and 40% are placed in the flight path of the unmanned aerial vehicle and captured in the unmanned aerial vehicle image, and Pix4Dmapper software is used to preprocess the RGB and MS orthoimages obtained by the unmanned aerial vehicle.

[0044] In this embodiment, preferably, the VIs and DAT in S5-S8 are normalized, so that the values of different types of VIs and DAT are in the range of (0, 1), and the dynamic curve of the spectral index or color index with DAT (days after transplanting) is defined as a closed region of the region, and the data set is divided into two parts, of which 70% is used for model training and the remaining 30% is used for model verification. The accuracy of the model is evaluated by the coefficient of determination (R2) and the normalized root mean square error (nRMSE). When R2 is closer to 1 and nRMSE is smaller, the effect of the model is better.

[0045] In this embodiment, preferably, the correlation coefficient between most of the vegetation indices and yield in S5 increases first and then decreases with the growth of rice, and reaches the maximum value in the booting stage and the heading stage. The correlation between most of the spectral indices and yield in the tillering stage is low. The highest correlation spectral index in each period is OSAVI (0.716), NDRE (0.730), NDVIRE (0.751), MNDI (0.737), MNDI (0.705), NDVIRE (0.752), respectively, among which NDVIRE and MNDI reach the maximum correlation in the booting stage, the maturation stage and the heading stage, and the filling stage, respectively. The color index performs well in the early growth period, and VARI performs best in the jointing stage to the heading stage. The highest correlation color index in each period is ExGR (0.819), VARI (0.742), VARI (0.783), MGRVI (0.720), ExR (-0.697), and ExG (0.722), respectively.

[0046] In this embodiment, preferably, the cumulative time-series spectral index correlation in the jointing stage, the booting stage and the heading stage in S6 is generally higher than that in other growth stages. The cumulative time-series spectral index combined in the jointing stage to the booting stage has the highest correlation with the yield of rice, and the R2 value reaches 0.722. Compared with the single-stage prediction result, the modeling R2 of the cumulative time-series spectral index is improved by 0.02, and the ∑VI model constructed by multi-stage data is mostly better than the single-stage model.

[0047] In this embodiment, preferably, the ∑VI model established in the tillering stage to the filling stage (B-IF) in S7 has the highest correlation with the yield (R2=0.72). Compared with the single-stage color index, the color index prediction error of the verification data set is reduced by 33%.

[0048] In this embodiment, preferably, the combination of the spectral index from the jointing stage to the booting stage and the color index from the heading stage to the early filling stage in S8 has the best prediction accuracy (R2=0.752), and the RRMSE of the validation data is 0.09. Since ∑(VI&ABD) combines the best expression period of the spectral index before the heading stage with the best expression period of the color index after the heading stage, it is superior to the prediction of the spectral index and the color index during the whole growth period. Compared with the optimal spectral index combination model, the multi-stage and multi-variable combination improves the prediction of rice yield by 0.03 and 0.025, respectively. The sensor fusion from the jointing stage to the booting stage improves the results based on one sensor by 7.3% and 7.2%, respectively.

[0049] In this embodiment, preferably, the unmanned aerial vehicle image acquisition in S2 is performed twice, and the top of the rice under the flight path of the unmanned aerial vehicle is coated with reflective paint. The flight height of the second unmanned aerial vehicle image acquisition is 5 m. When the unmanned aerial vehicle flies the second time, the unmanned aerial vehicle flight airflow blows the rice on the path, and the swing amplitude of the top of the rice is obtained by comparing with the first collected image. According to the correlation analysis of the swing amplitude of different remaining periods and the yield of rice, the swing amplitude of the growth period with the highest correlation is added to the prediction model in S9 for calculation to assist in predicting the yield of rice.

[0050] Unmanned aerial vehicles have become a new platform for high-precision agriculture to obtain high spatial and temporal resolution images. Based on MS images and unmanned aerial vehicle digital images, the application of single-stage VI and multi-period VI in rice yield prediction is explored. The obtained MS and digital images are reliable for estimating rice growth and grain yield.

[0051] MS and digital images respectively consider the booting stage and the tillering stage as the best growth stages for single-period VI estimation of rice yield, and the corresponding best VIS is NDVIRE and ExGR.

[0052] The application potential of single-period and multi-period cumulative time series of spectral index and color index in rice yield prediction based on unmanned aerial vehicle images is compared. The results show that the cumulative time series of spectral index from the jointing stage to the booting stage has the highest correlation with rice yield, with an R2 value of 0.722; the cumulative time series of color index model established from the tillering stage to the filling stage has the highest correlation with yield (R2 = 0.727). At the same time, the time series model combining the spectral index at the jointing stage and the booting stage and the color index at the heading stage and the early filling stage has the best estimation effect on rice yield, which can provide a new method for predicting rice yield using unmanned aerial vehicle images.

[0053] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting rice yield based on UAV remote sensing image processing, characterized in that, It includes the following steps: S1. Geographic coordinate correction: Five ground control points (GCPs) were set up in the study area and marked on them to perform geographic correction on UAV images at different growth stages. S2. Unmanned aerial vehicle (UAV) image acquisition: Pre-set automatic flight plan to acquire images with 90% forward overlap and 85% lateral overlap. The same flight path and camera settings are used throughout the entire reproductive period. S3. UAV image processing: The preprocessing process includes noise correction, vignetting correction, lens distortion correction, band-to-band registration, and radiometric correction. The RGB and MS orthophotos acquired by the UAV are preprocessed, and the main processes include denoising, radiometric calibration, band alignment, and orthophoto correction. Based on the processed images, plant indices, spectral indices, and color indices corresponding to different growth stages of rice are obtained. S4. During the physiological maturity period of rice, three uniform and representative 0.5m double-row areas were selected in each plot for sampling. After the panicles were cut, they were brought back to the laboratory for threshing. After the grains were dried to a constant weight, the moisture content was measured and weighed. The average value of the three sampling sub-plots was taken as the final yield of the plot. The yield was uniformly converted to the unit kg / ha. S5. Correlation analysis of the relationship between rice yield and spectral / color index at different growth stages; S6. Correlation analysis of the relationship between rice yield and multi-temporal spectral indices at different growth stages; S7. Correlation analysis of the relationship between rice yield and multi-phase color index at different growth stages; S8. Correlation analysis of the relationship between the cumulative time series spectral index and color index combination and rice yield; S9. Based on correlation analysis, the index or combination of indices with the highest correlation at different growth stages is used as the independent variable, and the corresponding rice yield is used as the dependent variable to establish a univariate linear regression model and a multivariate stepwise regression model. The model with higher accuracy between the univariate linear regression model and the multivariate stepwise regression model is determined as the rice yield prediction model through accuracy testing. S10. The rice yield prediction model is used to predict the rice yield of the area to be measured.

2. The method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: The geographic coordinates of the GCP within S1 are obtained by RTK-GPS (Real-time Dynamic Global Positioning System, CHC X900 GNSS), with horizontal and vertical errors within 1 cm and 2 cm, respectively.

3. The method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: The automatic flight plan in S2 is pre-set using DJI GS Pro software. The UAV acquires images of the test area at an altitude of 30 m during the key growth stages of rice. The images are stored in JPEG and TIFF formats. The flight speed is 2 m / s. Four standard reflective cloths are placed on the ground for radiometric correction of the UAV images. Each flight activity is carried out from 11:00 am to 2:00 pm local time under stable lighting conditions.

4. The method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: The entire preprocessing process in S3 uses ArcMap 10.

2. A calibrated radiometric measurement surface is placed in the flight area, allowing users to perform radiometric calibration. To perform radiometric calibration, four calibration targets with nominal reflectivities of 5%, 10%, 20%, and 40% are placed within the UAV's flight path and captured in the UAV image. Pix4Dmapper software is used to preprocess the RGB and MS orthophotos acquired by the UAV.

5. The method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: In S5-S8, VIs and DAT are normalized so that the values ​​of different types of VIs and DAT are within the range of (0,1). Through the dynamic curves of spectral index or color index with DAT (days after transplanting) at different growth stages, the spectral index and color index are defined as closed regions of the area.

6. The method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: In the S5, the correlation coefficients between most vegetation indices and yield first increased and then decreased with rice growth, reaching their maximum values ​​during the booting and heading stages. Most spectral indices during the tillering stage showed low correlation with yield. The spectral indices with the highest correlation at each stage were OSAVI (0.716), NDRE (0.730), NDVIRE (0.751), MNDI (0.737), MNDI (0.705), and NDVIRE (0.752). Among them, NDVIRE and MNDI reached their maximum correlation at the booting, maturity, heading, and grain-filling stages, respectively. Color indices performed well in the early growth stages, while VARI performed best from the jointing to the heading stage. The color indices with the highest correlation at each stage were ExGR (0.819), VARI (0.742), VARI (0.783), MGRVI (0.720), ExR (-0.697), and ExG (0.722).

7. The method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: The cumulative temporal spectral index correlations of the jointing, booting, and heading stages in S6 were generally higher than those of other growth stages. The cumulative temporal spectral index of the jointing to booting stage combination had the highest correlation with rice yield, with an R² value of 0.

722. Compared with the single-stage prediction results, the R² of the cumulative temporal spectral index modeling improved by 0.

02. The ∑VI model constructed from multi-stage data was mostly superior to the single-stage model.

8. The method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: The ∑VI model established in S7 during the tillering to grain-filling stage (B-IF) showed the highest correlation with yield (R2=0.72), and the color index prediction error of the validation dataset was reduced by 33% compared with the single-stage color index.

9. A method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: The S8 model showed the best prediction accuracy for the combination of spectral indices from the jointing stage to the booting stage and color indices from the heading stage to the early grain-filling stage (R² = 0.752). The RRMSE of the validation data was 0.

09. Since ∑(VI&ABD) combines the optimal performance period of the spectral indices before the heading stage with the optimal expression period of the color indices after the heading stage, it is superior to the prediction of spectral and color indices throughout the entire growth period. Compared with the optimal spectral index combination model, the multi-stage and multi-variable combination improved the rice yield prediction by 0.03 and 0.025, respectively. The sensor fusion ratio from the jointing stage to the booting stage improved the results based on a single sensor by 7.3% and 7.2%, respectively.

10. A method for predicting rice yield based on UAV remote sensing image processing according to claim 1, characterized in that: In S2, the UAV image acquisition is performed twice, and the tops of the rice plants along the UAV flight path are coated with reflective paint. The second UAV image acquisition flight height is 5m. During the second UAV flight, the airflow from the UAV causes the rice plants along the path to tilt. By comparing the images with the first acquisition, the swing amplitude of the rice tops is obtained. Correlation analysis is performed on the swing amplitude and rice yield at different remaining stages. The swing amplitude at the growth stage with the highest correlation is added to the prediction model in S9 to assist in predicting rice yield.

Images