A method for remote sensing estimation of protein content in wheat grains at different development stages

Abstract

The application discloses a kind of multi-development period wheat kernel protein content remote sensing estimation method, belong to agricultural remote sensing monitoring technical field.The method includes: S1, obtain satellite, unmanned aerial vehicle and ground multi-source observation data;S2, with SPAD as dependent variable, fusion vegetation index and texture index, using random forest algorithm to build field scale SPAD remote sensing estimation model;S3, with SPAD as intermediate variable, build "remote sensing spectrum-SPAD-GPC" indirect estimation model;S4, for jointing stage, heading stage, flowering stage and grain filling mid-period are respectively optimized model;S5, through scale conversion to establish regional scale GPC estimation model.The application comprehensively utilizes multi-source remote sensing data, fusion spectrum and texture information, with SPAD as intermediate variable to improve machine interpretability, effectively solve scale mismatch problem, realize multi-development period, multi-scale wheat kernel protein content accurate remote sensing estimation, provide technical support for wheat quality monitoring and agricultural production management.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing monitoring technology, and in particular to a remote sensing estimation method for the protein content of wheat grains at multiple developmental stages. Background Technology

[0002] Wheat is a major food crop, providing 20% ​​of the world's protein source and playing a vital role in food security and protein supply. With socio-economic development and the deepening of agricultural supply-side structural reforms, China's wheat supply and demand situation has changed significantly, with the demand for high-quality wheat constantly rising. However, the supply of high-quality wheat is insufficient and relies on imports. Therefore, while ensuring food security, developing high-quality wheat is a top priority for research.

[0003] Grain protein content (GPC) is a key indicator for evaluating wheat quality. Rapid monitoring of wheat GPC during production can guide farmers in optimizing cultivation, harvesting, and storage, and is of great significance for modern wheat management and achieving high yields and high quality.

[0004] Traditional GPC detection methods mainly include the Kjeldahl nitrogen determination method and the H2SO4-H2O2 colorimetric method, which are time-consuming, labor-intensive, costly, and carry certain risks. Near-infrared and hyperspectral imaging analysis is used to detect grain protein content after wheat maturity and harvest. This method has the advantages of low cost and fast analysis speed. However, like laboratory chemical analysis methods, it involves point sampling after wheat maturity and harvest, which cannot be scaled up to a regional scale and is difficult to guide actual production.

[0005] Remote sensing technology, with its rapid and non-destructive characteristics, can efficiently acquire spectral information of ground objects, enabling a shift from point-based to area-based wheat GPC monitoring, and gradually demonstrating its advantages in wheat GPC monitoring. In recent years, in particular, UAV remote sensing has demonstrated high mobility and flexibility, and has been widely used in monitoring crop nitrogen, GPC, and other parameters.

[0006] Current technologies primarily focus on estimating wheat grain protein content using a single remote sensing platform. Directly utilizing satellite and ground-based observation data can lead to scale effects due to scale mismatch, resulting in mechanistic errors in the model. Furthermore, empirical models based on remote sensing information and GPC (Geophysical Processing) do not address the mechanisms of wheat grain protein formation, leading to significant decreases in stability and prediction accuracy when applied to environments with large variations or across different regions. While some studies have introduced agronomic parameters as intermediate parameters to indirectly retrieve GPC information, improving mechanistic interpretability, the following limitations remain: (1) A single remote sensing data source has limitations and cannot fully leverage the advantages of remote sensing data at different scales; (2) The application of texture information in GPC monitoring is insufficient, and the spectral information and texture information are not effectively integrated to improve the inversion accuracy; (3) There is a scale mismatch between ground observations and satellite data, which affects the accuracy of regional scale GPC estimation; (4) There is a lack of systematic GPC estimation models for multiple key developmental stages of wheat, which cannot meet the monitoring needs of different growth stages.

[0007] Therefore, there is an urgent need to develop a method that comprehensively utilizes multi-source remote sensing data, integrates multiple features, and achieves accurate estimation of wheat GPC at multiple scales and developmental stages. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a remote sensing estimation method for wheat grain protein content at multiple developmental stages. This method comprehensively utilizes multi-source observation data from satellites, UAVs, and ground sources, uses SPAD as an intermediate parameter, and integrates vegetation index and texture index to construct a multi-developmental stage and multi-scale wheat GPC remote sensing estimation model, effectively solving the scale mismatch problem and improving the accuracy and applicability of GPC estimation.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A remote sensing method for estimating the protein content of wheat grains at multiple developmental stages includes the following steps: S1. Acquisition of multi-source remote sensing and ground observation data: Satellite remote sensing data, UAV multispectral image data, and ground observation data were acquired at multiple key developmental stages of wheat; the ground observation data included the measured values ​​of SPAD (spider chlorophyll content) of wheat canopy and GPC (grain protein content). S2. Construction of SPAD remote sensing inversion model: The SPAD value measured on the ground is the dependent variable, and the vegetation index and texture index extracted from UAV multispectral images are the independent variables. The random forest algorithm is used to construct a field-scale wheat canopy SPAD remote sensing estimation model. S3. Construction of GPC Indirect Estimation Model: Using SPAD value as an intermediate parameter, the correlation between SPAD value at different developmental stages and GPC protein content in wheat grains at maturity was analyzed, and an indirect estimation model of "remote sensing spectroscopy-SPAD-GPC" was constructed. S4. Multi-developmental stage model optimization: SPAD inversion models and GPC estimation models were constructed for the wheat jointing stage, heading stage, flowering stage and mid-grain filling stage, respectively, and the optimal feature combination and model parameters for each developmental stage were screened. S5. Scale conversion and regional estimation: The SPAD inversion results at the UAV scale are resampled to the same spatial resolution as the satellite data, a quantitative relationship between satellite remote sensing features and GPC is established, a regional scale wheat GPC estimation model is constructed, and multi-scale wheat GPC remote sensing monitoring from field to region is realized.

[0010] Furthermore, in step S1, the multiple key developmental stages include the wheat jointing stage, heading stage, flowering stage, and mid-grain filling stage; the satellite remote sensing data is Sentinel-2L2A level data with a spatial resolution of 10 meters; the UAV multispectral image data is acquired by a UAV equipped with a multispectral camera, including blue, green, red, red-edge, near-infrared, and thermal infrared bands.

[0011] Further, in step S2, the vegetation index includes one or more combinations of RVI, NDRE, NPCI, NRI, Cirede, GNDVI, and OSAVI.

[0012] Further, in step S2, the method for constructing the texture index includes: Texture features of each band in multispectral image were extracted using a second-order probabilistic statistical filtering method, including mean (MEA), variance (VAR), coherence (HOM), contrast (CON), dissimilarity (DIS), information entropy (ENT), second moment (SEC), and correlation (COR). By combining any two texture features in pairs, three types of texture indices are constructed: Difference Texture Index (DTI), Normalized Texture Index (NDTI), and Ratio Texture Index (RTI). Correlation analysis was used to select the texture index combination with the best correlation to the SPAD value.

[0013] Furthermore, in step S2, when the random forest algorithm constructs the SPAD estimation model, it uses 80% of the samples as the modeling set and 20% of the samples as the validation set, and optimizes key parameters such as the maximum number of iterations and the learning depth.

[0014] Furthermore, in step S2, the SPAD remote sensing monitoring model constructed by integrating vegetation index and texture index has the highest inversion accuracy during the wheat flowering period, with a coefficient of determination R² of 0.91 and a root mean square error RMSE of 1.77.

[0015] Furthermore, in step S3, the "remote sensing spectroscopy-SPAD-GPC" indirect estimation model adopts a polynomial model form, and the formula for the flowering period GPC estimation model is: y = 0.0148x² - 1.3132x + 37.196 Where y is the wheat grain protein content (GPC) and x is the wheat canopy chlorophyll content (SPAD).

[0016] Further, in step S5, the scale conversion method includes: The SPAD inversion results from the UAV were resampled to a spatial resolution of 10 meters, consistent with the resolution of Sentinel-2 satellite imagery. Based on the resampled GPC data and Sentinel-2 multispectral data, the reflectance and sensitive vegetation index of each band were calculated. A stepwise regression algorithm was used to construct a regional GPC linear inversion model. The model was iteratively optimized by gradually introducing independent variables, and the final estimation model was constructed by selecting the most sensitive variables.

[0017] Furthermore, the regional-scale GPC estimation model selects six sensitive variables—B2, B3, B7, NDRE2, NPCI, and NRI—to construct a stepwise regression model. The model formula is as follows: GPC=-424.230373+9125.757669×B2-7179.782355×B3+55.546703×B7+73.630540×NDRE2+3337.563011×NPCI+3289.151817×NRI Among them, B2, B3, and B7 are the blue, green, and red edge reflectances of the Sentinel-2 satellite, respectively; NDRE2 is the normalized red edge vegetation index; NPCI is the normalized pigment chlorophyll index; and NRI is the nitrogen reflectance index.

[0018] Furthermore, in step S1, the ground observation data also includes the leaf area index (LAI); the method also includes a verification step: using the coefficient of determination (R²) and root mean square error (RMSE) to evaluate the estimation effect of the model.

[0019] Compared with existing technologies, this rental vehicle monitoring, control, and intelligent traffic management system has the following advantages: (1) This invention constructs a sky-ground collaborative observation system, which comprehensively utilizes satellite, UAV and ground multi-source data to make up for the shortcomings of a single data source in a specific aspect, give full play to the advantages of each data source, and realize the comprehensiveness, accuracy and timeliness of wheat GPC monitoring.

[0020] (2) This invention uses SPAD as an intermediate parameter to construct an indirect estimation model of "remote sensing spectrum-SPAD-GPC". Compared with the direct empirical model, it has better mechanism interpretation and improves the stability and applicability of the model under different environmental conditions.

[0021] (3) The present invention integrates vegetation index and texture index to construct SPAD inversion model, effectively utilizes texture information as a supplement to spectral information, and significantly improves SPAD inversion accuracy. Among them, the flowering period model has an R² of 0.91 and an RMSE of 1.77.

[0022] (4) This invention constructs estimation models for multiple key developmental stages of wheat (jointing stage, heading stage, flowering stage, and mid-filling stage) to meet the monitoring needs of different growth stages and provide technical support for quality monitoring of the entire wheat production process.

[0023] (5) This invention uses a scale conversion method to upgrade the UAV scale inversion results to the satellite scale, effectively solving the problem of scale mismatch between ground observation and satellite data, and realizing multi-scale wheat GPC remote sensing monitoring from field to region, which has important application value. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 For the distribution of sample points and acquisition of field trial data; Figure 2 Characteristics of SPAD and LAI changes in wheat at different developmental stages; Figure 3 The variation characteristics of SPAD and LAI in wheat at different developmental stages and nitrogen application levels; Figure 4 Correlation analysis of spectral reflectance, vegetation index and SPAD; Figure 5 Heatmap showing the correlation coefficients between SPAD and texture index in wheat at different developmental stages; Figure 6 Scatter plot of SPAD monitoring models for wheat canopy at different developmental stages, constructed based on optimal combinations; Figure 7 Spatial distribution map of SPAD estimated at different times based on UAVs; Figure 8 Spatial distribution map of GPC estimation based on UAVs at different times. Detailed Implementation To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below. Example

[0026] like Figure 1-8 As shown, this invention discloses a remote sensing method for estimating the protein content of wheat grains at multiple developmental stages, comprising the following steps: Step S1: Acquisition of multi-source remote sensing and ground observation data Satellite remote sensing data, UAV multispectral imagery data, and ground observation data were acquired at several key developmental stages of wheat.

[0027] The key developmental stages include the wheat jointing stage, heading stage, flowering stage, and mid-grain filling stage.

[0028] (1) Acquisition of ground observation data On the day of the experiment, the LAI-2200 plant canopy analyzer was used to measure the LAI of winter wheat in the experimental plots, and the latitude and longitude of the sampling points were recorded. During measurement, direct sunlight was avoided as much as possible. On the side facing away from the sun, one skylight measurement was taken first, and then the plant canopy analyzer was placed near the roots of the winter wheat to measure four target values. The lens was kept horizontal throughout the measurement, and the average of the four measurements was taken as the result of this experiment. Three replicates were performed at each sample point, and the mean value represents the LAI value of that sample point.

[0029] Chlorophyll content was determined using a SPAD-502 chlorophyll meter. Healthy and typical wheat leaves from each growth stage were selected as the measurement subjects, avoiding leaves with disease spots or damage. Leaf overlap was avoided during measurement. Five replicates were performed for each sample point, and the average value was taken to represent the chlorophyll content (SPAD value) of that sample point.

[0030] During the winter wheat maturity period, wheat in a 1m×1m quadrat was harvested, and after threshing, natural drying, and air drying, the grain protein content (GPC) of winter wheat was determined using a near-infrared grain analyzer and the Kjeldahl nitrogen determination method, respectively.

[0031] (2) Acquisition of multispectral data by UAV The flight test used a DJI M300RTK drone equipped with an Altum-PT multispectral camera, including six bands: blue (475.0±16.0nm), green (560.0±13.5nm), red (668.0±7.0nm), red edge (717.0±6.0nm), near infrared (842.0±28.5nm), and thermal infrared (7.5-13.5μm). It can simultaneously acquire RGB images, heat maps, and high-resolution panchromatic data.

[0032] Wheat canopy spectral data were collected under clear, cloudless, and relatively stable atmospheric conditions from 10:00 to 14:00. The flight path was planned using the DJIPilot application. To ensure image resolution and overlap, the forward and lateral overlap rates were both set to 80%, and the flight altitude was 60m.

[0033] Before each flight, absolute reflectance values ​​were radiometrically corrected using the MicaSense Calibration Reflectance Panel (CRP). Multispectral images were stitched together using Agisoft Metashape Professional software to generate orthorectified images with a pixel resolution of 5cm × 5cm. Finally, based on the preprocessed multispectral orthorectified images, the spectral reflectance values ​​of the corresponding pixels at the ground sampling points were extracted.

[0034] (3) Satellite data acquisition Based on the AIEarth Earth Science Cloud Platform, Sentinel-2 L2A level remote sensing data that had undergone preprocessing including radiometric calibration, orthorectification, and atmospheric correction were acquired. ENVI software was used to resample all bands to a spatial resolution of 10m. Combined with UAV flight tests in Datian, Sentinel-2 data from corresponding periods were selected for satellite-scale SPAD inversion and validation.

[0035] Step S2: Construction of SPAD remote sensing inversion model Using ground-measured SPAD values ​​as the dependent variable and vegetation and texture indices extracted from UAV multispectral images as independent variables, a field-scale wheat canopy SPAD remote sensing estimation model was constructed using the random forest algorithm.

[0036] (1) Calculation of vegetation index Vegetation indices that are highly correlated with chlorophyll content and nitrogen content were selected for model construction, including RVI (ratio vegetation index), NDRE (normalized red-edge vegetation index), NPCI (normalized pigment chlorophyll index), NRI (nitrogen reflectance index), Cirede (red-edge chlorophyll index), GNDVI (normalized greenness vegetation index), and OSAVI (optimized soil-adjusted vegetation index).

[0037] (2) Texture index construction Texture features of images are extracted using a second-order probabilistic statistical filtering method. The step size is set to 1, the window size is 3×3, and the gray level is 64. Eight types of texture features are extracted from the blue, green, red, red edge, and near-infrared bands, including mean (MEA), variance (VAR), coherence (HOM), contrast (CON), dissimilarity (DIS), information entropy (ENT), second moment (SEC), and correlation (COR).

[0038] By combining any two texture features pairwise, three types of texture indices are constructed: Differential Texture Index (DTI): DTI(i,j) = TF(i) - TF(j) Normalized Texture Index (NDTI): NDTI(i,j)=(TF(i)-TF(j)) / (TF(i)+TF(j)) Ratio Texture Index (RTI): RTI(i,j) = TF(i) / TF(j) Where TF(i) and TF(j) are the i-th and j-th texture feature values, respectively.

[0039] Correlation analysis was used to select the combination of texture indices with the best correlation to SPAD values. The correlation between the selected sensitive texture indices and SPAD values ​​improved to varying degrees at each developmental stage, with the largest improvement observed during the flowering stage, where the correlation coefficient exceeded 0.72. In terms of feature types, the texture indices sensitive to SPAD prediction were primarily constructed using the mean and variance of the red-edge and near-infrared bands.

[0040] Table 1. Correlation coefficients between optimal texture index and wheat SPAD

[0041] (3) Construction of Random Forest Model Random Forest (RF) is a classic machine learning algorithm proposed by Breiman et al., belonging to the parallel ensemble learning method. Random Forest Regression achieves high accuracy by integrating the predictions of multiple decision trees, and each tree is interpretable. By randomly sampling features, it reduces dependence on training data and the risk of overfitting, and can learn both linear and nonlinear relationships.

[0042] During modeling, 80% of the samples are selected as the modeling set and the remaining 20% ​​as the validation set. To ensure the model is optimal, key parameters such as the maximum number of iterations and the learning depth are optimized, while other parameters remain at their default settings.

[0043] The root mean square error (RMSE) and coefficient of determination (R²) are used to evaluate the simulation performance of the model.

[0044] The inversion results show that the wheat SPAD model constructed using only spectral reflectance has an R² greater than 0.80 and an RMSE less than 3.7, effectively explaining most of the variation patterns in wheat SPAD values. Compared to spectral reflectance alone, the model constructed based on vegetation indices at the heading and grain-filling stages shows enhanced sensitivity to SPAD spectra. After incorporating texture information into the vegetation indices, the model inversion accuracy is further improved, with increased R² at the jointing, heading, and flowering stages, and decreased RMSE at each developmental stage. The model optimization effect is most prominent at the flowering stage, with an R² increase of 0.08 and a RMSE decrease of 0.46.

[0045] In summary, the SPAD remote sensing monitoring model constructed by integrating vegetation index and texture index has the best inversion effect, with the highest accuracy in simulating the flowering period, achieving an R² of 0.91 and an RMSE of 1.77.

[0046] Spatial inversion of SPAD (spider plant growth rate) at four key developmental stages of winter wheat was performed using a random forest model based on vegetation and texture indices. The SPAD inversion results at the field scale are shown below. Figure 8 As shown in the figure, the average SAPD for the four periods were 52.2, 53.9, 54.6, and 49.0, respectively. The results indicate that SPAD showed an increasing trend from the jointing stage to the flowering stage, but decreased significantly during the mid-grain-filling stage. Spatially, the SPAD level remained high in winter wheat under high nitrogen fertilizer treatment.

[0047] In addition, the inversion model was validated using observational data from the flowering period in 2024. The correlation coefficient was 0.80 and the RMSE was 3.83, which fully verified that the model has good generalization ability and application value.

[0048] Table 2 Comparison of SPAD inversion accuracy based on different feature combinations and random forest regression models.

[0049]

[0050] Step S3: Construction of GPC indirect estimation model Using SPAD value as an intermediate parameter, the correlation between SPAD value at different developmental stages and GPC protein content in wheat grains at maturity was analyzed, and an indirect estimation model of "remote sensing spectroscopy-SPAD-GPC" was constructed.

[0051] Wheat grain nutrient content (GPC) is closely related to nitrogen nutrient status. Changes in nitrogen nutrient status cause changes in leaf spadix density (SPAD), which are further reflected in the canopy reflectance spectrum. Analysis of the relationship between mature grain GPC and SPAD and LAI at different developmental stages revealed a highly significant positive correlation between leaf SPAD and GPC at all key developmental stages of wheat, with the highest correlation coefficient (0.74) at the flowering stage. LAI showed a highly significant positive correlation with GPC at the heading and flowering stages, and a significant positive correlation at the early grain-filling stage, but the correlation was lower than that with SPAD. Therefore, SPAD was used as a linking parameter between spectral characteristics and GPC.

[0052] Table 3. Correlation between GPC and SPAD and LAI in mature wheat grains at different developmental stages

[0053] Based on the SPAD inversion results from UAVs, SPAD values ​​within a 10×10 grid area were extracted using the center positions of two sampling points in each field as a benchmark, and the average value was taken as the SPAD value for the corresponding quadrat. Using SPAD as an intermediate parameter, linear, exponential, and multinomial models were employed to construct UAV-based wheat GPC estimation models.

[0054] The GPC modeling results at various stages show that the model accuracy, from highest to lowest, is the multinomial model, the exponential model, and the linear model, indicating a good nonlinear relationship between GPC and SPAD. Comparing the optimal models for each growth stage reveals that the R² is approximately 0.60 and the RMSE is approximately 1.1 for the jointing, heading, and flowering stages, with the highest model accuracy observed during the flowering stage.

[0055] The formula for the GPC polynomial estimation model during the flowering period is: y = 0.0148x² - 1.3132x + 37.196 Where y is the wheat grain protein content (GPC) (%), and x is the wheat canopy chlorophyll content (SPAD value).

[0056] The flowering period model was validated using field trial data, and the RMSE was 1.01, indicating that the model has high reliability.

[0057] Due to the influence of field irrigation conditions, the sampling data during the grain-filling period showed high dispersion, with the model's R² value being only 0.43, indicating a relatively weak modeling effect. Therefore, when constructing the regional-scale model later, the flowering period model was prioritized. The flowering period model was validated using field trial data from 2024, with an RMSE of 1.01.

[0058] GPC estimation results for each period are as follows Figure 8 As shown, the mean values ​​of the four groups are 9.6, 9.7, 9.9, and 9.7, respectively, with relatively small deviations from the measured value (10.0). Among them, the estimated GPC range during the flowering period is between 8.1 and 12.9, while the corresponding measured value ranges from 7.38 to 14.82. The model as a whole exhibits the bias of overestimating low values ​​and underestimating high values, but it can reflect the average level of GPC relatively well.

[0059] Table 4. SPAD-GPC estimation model based on UAV spectral parameters

[0060] Step S4: Optimization of the multi-developmental stage model SPAD inversion models and GPC estimation models were constructed for the wheat jointing stage, heading stage, flowering stage and mid-grain filling stage, respectively, and the optimal feature combination and model parameters for each development stage were screened.

[0061] Wheat SPAD and LAI showed a trend of first increasing and then decreasing as development progressed, with differences also existing among different treatments at the same stage. At the jointing stage, the average SPAD and LAI of wheat were 54.7 and 3.15, respectively. As development progressed, these two indicators reached their peak at the flowering stage, at 55.9 and 4.6, respectively. Subsequently, wheat gradually transitioned from vegetative growth to reproductive growth, leaves began to stop growing and gradually withered, and SPAD and LAI also decreased accordingly, falling to 53.2 and 3.7, respectively.

[0062] Different nitrogen application levels significantly affected the SPAD and LAI of wheat. The SPAD values ​​for N0, N1, and N2 treatments were 50.5, 56.2, and 57.8, respectively, while the LAI values ​​were 3.2, 3.9, and 4.3, respectively. Wheat at the flowering stage exhibited the highest SPAD and LAI under the high nitrogen application level (N2). During the mid-grain-filling stage, the SPAD decreased by 10.2% under low nitrogen application levels, while the SPAD under high nitrogen application levels remained at a relatively high level, demonstrating strong physiological activity and environmental adaptability.

[0063] The accuracy of the SPAD inversion model for each developmental stage is as follows: Jointing stage: R²=0.88, RMSE=2.36; Heading stage: R²=0.87, RMSE=2.50; Flowering period: R²=0.91, RMSE=1.77; Mid-stage grouting: R²=0.89, RMSE=3.12; Among them, the flowering period model has the highest accuracy and can be given priority for wheat GPC estimation.

[0064] Step S5, Scale Transformation and Region Estimation: By resampling the SPAD inversion results at the UAV scale to the same spatial resolution as the satellite data, a quantitative relationship between satellite remote sensing features and GPC is established, and a regional-scale wheat GPC estimation model is constructed to achieve multi-scale wheat GPC remote sensing monitoring from field to region.

[0065] At the field scale, the SPAD and GPC models for wheat flowering period have the highest accuracy. Data from this period were selected to construct a satellite-scale wheat GPC inversion model.

[0066] Table 5. Correlation between remote sensing features and GPC

[0067] (1) Scale transformation The SPAD inversion results from the wheat flowering stage were resampled to a spatial resolution of 10 meters, consistent with the resolution of Sentinel-2 satellite imagery. The wheat GPC was calculated using the constructed quantitative model, and then matched pixel-by-pixel with the band reflectance of the Sentinel-2 satellite data.

[0068] (2) Satellite-scale feature extraction Based on Sentinel-2 multispectral data, sensitive vegetation indices such as RVI, NDRE, NPCI, NRI, Cirede, GNDVI, OSAVI, and GBNDVI were calculated. Correlation analysis between reflectance in each band, vegetation indices, and GPC showed that reflectance in the blue, green, red, and red-edge 704 bands was negatively correlated with GPC retrieved by UAV, while the remaining bands were positively correlated. Among the vegetation indices, RVI, NDRE1, NDRE2, GNDVI, CI1, CI2, OSAVI, and GBNDVI all showed highly significant correlations with GPC.

[0069] (3) Construction of regional scale model To improve the regional applicability and stability of the inversion model, a stepwise regression algorithm was used to construct a regional GPC linear inversion model. By progressively introducing independent variables to conduct iterative optimization, the results show that the changes in R² and RMSE are most significant when the number of independent variables increases to six; after further increasing the number of variables, the model accuracy indicators tend to level off.

[0070] Considering both model complexity and simulation results, six sensitive variables—B2, B3, B7, NDRE2, NPCI, and NRI—were ultimately selected to construct a stepwise regression model. The model formula is as follows: GPC=-424.230373+9125.757669×B2-7179.782355×B3+55.546703×B7+73.630540×NDRE2+3337.563011×NPCI+3289.151817×NRI Among them, B2, B3, and B7 are the blue, green, and red edge reflectances of the Sentinel-2 satellite, respectively; NDRE2 is the normalized red edge vegetation index; NPCI is the normalized pigment chlorophyll index; and NRI is the nitrogen reflectance index.

[0071] The model has an R² of 0.62 and an RMSE of 0.74, which meets the requirements for regional-scale wheat GPC remote sensing inversion.

[0072] This invention integrates multi-source data from satellites, UAVs, and ground-based sources to construct a GPC remote sensing estimation model applicable to different developmental stages and scales of wheat. Based on field experiments covering multiple wheat varieties and different nitrogen fertilizer treatments, the study confirms that SPAD can serve as a key parameter linking spectral characteristics and wheat grain protein content. In constructing the field-level SPAD remote sensing monitoring model, the optimal inversion effect was achieved by fusing vegetation indices and texture indices. Using SPAD as an intermediate parameter, a quantitative estimation model of "remote sensing spectrum-SPAD-GPC" was constructed, and its high reliability was verified. Furthermore, GPC data obtained through UAV inversion was scaled up to 10m, establishing a regional-scale GPC estimation model.

[0073] The resulting multi-scale wheat GPC remote sensing estimation model effectively solved the scale mismatch problem between ground observation data and satellite data, overcame the limitations of traditional single-point methods, and achieved effective observation from the ground to the regional scale. The established method contributes to wheat quality monitoring and provides a scientific basis for macro-management and decision-making in agricultural production.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A remote sensing method for estimating the protein content of wheat grains at multiple developmental stages, characterized in that, Includes the following steps: S1. Acquisition of multi-source remote sensing and ground observation data: Satellite remote sensing data, UAV multispectral image data, and ground observation data were acquired at multiple key developmental stages of wheat; the ground observation data included the measured values ​​of SPAD (spider chlorophyll content) of wheat canopy and GPC (grain protein content). S2. Construction of SPAD remote sensing inversion model: The SPAD value measured on the ground is the dependent variable, and the vegetation index and texture index extracted from UAV multispectral images are the independent variables. The random forest algorithm is used to construct a field-scale wheat canopy SPAD remote sensing estimation model. S3. Construction of GPC Indirect Estimation Model: Using SPAD value as an intermediate parameter, the correlation between SPAD value at different developmental stages and GPC protein content in wheat grains at maturity was analyzed, and an indirect estimation model of "remote sensing spectroscopy-SPAD-GPC" was constructed. S4. Multi-developmental stage model optimization: SPAD inversion models and GPC estimation models were constructed for the wheat jointing stage, heading stage, flowering stage and mid-grain filling stage, respectively, and the optimal feature combination and model parameters for each developmental stage were screened. S5. Scale conversion and regional estimation: The SPAD inversion results at the UAV scale are resampled to the same spatial resolution as the satellite data, a quantitative relationship between satellite remote sensing features and GPC is established, a regional scale wheat GPC estimation model is constructed, and multi-scale wheat GPC remote sensing monitoring from field to region is realized.

2. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 1, characterized in that, In step S1, the multiple key developmental stages include the wheat jointing stage, heading stage, flowering stage, and mid-grain filling stage; The satellite remote sensing data is Sentinel-2L2A level data with a spatial resolution of 10 meters; The UAV multispectral image data is acquired by a UAV equipped with a multispectral camera, including blue, green, red, red-edge, near-infrared, and thermal infrared bands.

3. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 1, characterized in that, In step S2, the vegetation index includes one or more combinations of RVI, NDRE, NPCI, NRI, Cirede, GNDVI, and OSAVI.

4. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 1, characterized in that, In step S2, the method for constructing the texture index includes: Texture features of each band in multispectral image were extracted using a second-order probabilistic statistical filtering method, including mean (MEA), variance (VAR), coherence (HOM), contrast (CON), dissimilarity (DIS), information entropy (ENT), second moment (SEC), and correlation (COR). By combining any two texture features in pairs, three types of texture indices are constructed: Difference Texture Index (DTI), Normalized Texture Index (NDTI), and Ratio Texture Index (RTI). Correlation analysis was used to select the texture index combination with the best correlation to the SPAD value.

5. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 1, characterized in that, In step S2, when the random forest algorithm constructs the SPAD estimation model, it uses 80% of the samples as the modeling set and 20% of the samples as the validation set, and optimizes key parameters such as the maximum number of iterations and the learning depth.

6. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 1, characterized in that, In step S2, the SPAD remote sensing monitoring model constructed by integrating vegetation index and texture index has the highest inversion accuracy during the wheat flowering period, with a coefficient of determination R² of 0.91 and a root mean square error RMSE of 1.

77.

7. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 1, characterized in that, In step S3, the "remote sensing spectroscopy-SPAD-GPC" indirect estimation model adopts a polynomial model form, and the formula for the flowering period GPC estimation model is: y = 0.0148x² - 1.3132x + 37.196 Where y is the wheat grain protein content (GPC) and x is the wheat canopy chlorophyll content (SPAD).

8. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 1, characterized in that, In step S5, the scale conversion method includes: The SPAD inversion results from the UAV were resampled to a spatial resolution of 10 meters, consistent with the resolution of Sentinel-2 satellite imagery. Based on the resampled GPC data and Sentinel-2 multispectral data, the reflectance and sensitive vegetation index of each band were calculated. A stepwise regression algorithm was used to construct a regional GPC linear inversion model. The model was iteratively optimized by gradually introducing independent variables, and the final estimation model was constructed by selecting the most sensitive variables.

9. The remote sensing estimation method for protein content in wheat grains at multiple developmental stages according to claim 8, characterized in that, The regional-scale GPC estimation model selects six sensitive variables—B2, B3, B7, NDRE2, NPCI, and NRI—to construct a stepwise regression model. The model formula is as follows: GPC=-424.230373+9125.757669×B2-7179.782355×B3+55.546703×B7+73.630540×NDRE2+3337.563011×NPCI+3289.151817×NRI; Among them, B2, B3, and B7 are the blue, green, and red edge reflectances of the Sentinel-2 satellite, respectively; NDRE2 is the normalized red edge vegetation index; NPCI is the normalized pigment chlorophyll index; and NRI is the nitrogen reflectance index.

Images