Construction method of nitrogen application rate classification model based on crop canopy multispectral
By constructing a nitrogen application rate classification model based on canopy multispectral data and utilizing multi-temporal monitoring and principal component analysis of vegetation indices, the initial background differences are reduced, solving the problems of long monitoring cycles and severe background interference in existing technologies, and achieving rapid and accurate nitrogen fertilizer management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2026-04-30
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies for monitoring crop canopy nitrogen supply suffer from long monitoring cycles, low timeliness, severe background color interference, difficulty in eliminating varietal differences, and insufficient physiological feedback characterization, resulting in poor model universality and low timeliness.
By acquiring canopy multispectral image data under different fertilization gradients in the same area, and after preprocessing, a nitrogen application rate classification model is constructed using support vector machine and Bayesian optimization algorithms to reduce the initial background differences. Combined with multi-temporal monitoring and vegetation index principal component analysis, a multi-temporal feature matrix is constructed to achieve rapid capture of nitrogen fertilizer response.
It significantly improved the model's identification accuracy and universality across different fields, shortened the evaluation cycle, ensured the model's hierarchical robustness in complex field environments, and enabled rapid and precise nitrogen fertilizer management.
Smart Images
Figure CN122385501A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural remote sensing monitoring technology, and in particular to a method for constructing a nitrogen application rate classification model based on crop canopy multispectral data. Background Technology
[0002] In intensive agricultural production systems, the application of chemical nitrogen fertilizers has become a key measure to ensure food security and improve land productivity. However, the long-standing concept of "ensuring yield through quantity" in fertilization has led to widespread excessive nitrogen fertilizer input, far exceeding the actual needs of crops and the carrying capacity of the soil, easily causing agricultural non-point source pollution. Crop growth and development are highly sensitive to nitrogen supply levels. As a core element regulating the construction of photosynthetic organs and chlorophyll synthesis, nitrogen supply significantly affects crop canopy structure and its optical properties. Therefore, using near-ground remote sensing technology for real-time, non-destructive nitrogen nutrition diagnosis of crop canopies has become a core approach to achieving fertilizer reduction and efficiency improvement in the field of precision agriculture.
[0003] Currently, methods for guiding nitrogen application using crop canopy multispectral imagery mostly focus on feature mining of static nodes, making it difficult to eliminate interference caused by varietal genetic characteristics and differences in initial background. For example, Chinese patent CN112903600B discloses a method for recommending nitrogen fertilizer for rice based on multispectral imagery from a fixed-wing UAV. This method obtains sensitive spectral indicators of key crop growth stages, fits a time-series spectral dynamic curve of the entire growth period, and uses a sufficiency index algorithm combined with monitoring results to give a recommended amount of nitrogen fertilizer.
[0004] However, the aforementioned existing technologies still have the following limitations in practical applications: (1) Long monitoring cycle and low timeliness: The aforementioned patent (CN112903600B) requires long-term continuous monitoring to depict the dynamic curve of the entire growth period, and cannot provide rapid feedback and identification for specific fertilization behaviors.
[0005] (2) Severe interference from background color: Due to the inherent differences in initial leaf color and growth potential among different rice varieties, single-phase or single-trend monitoring is difficult to distinguish between "initial background difference" and "nitrogen contribution", resulting in poor universality of the model when applied across varieties.
[0006] (3) Insufficient physiological feedback characterization: Single-phase or conventional growth period monitoring cannot effectively characterize the instantaneous dynamic physiological feedback of crops to specific fertilization behaviors, which is essentially a fragmentation of information in the high-dimensional feature space.
[0007] Therefore, there is an urgent need for an identification method that can accurately capture the dynamic response patterns of crop nitrogen by comparing the time sequence before and after fertilization, effectively eliminating background noise. Summary of the Invention
[0008] Purpose of the invention: The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing a nitrogen application rate classification model based on crop canopy multispectral density, including: S1: Acquire canopy multispectral image data of crops under different fertilization gradient treatments in the same area before and after fertilization, and perform preprocessing such as stitching, cropping, geometric correction, radiometric calibration and atmospheric correction on the multispectral image data; S2: Based on the preprocessed multispectral image data, extract the canopy multispectral reflectance of no less than 15 spatial sampling points before and after fertilization for each processing; S3: Based on the canopy response characteristics of each treatment spatial sampling point before fertilization, taking the canopy response characteristics of one treatment as a reference, and based on the initial background differences of each treatment before fertilization, the canopy response characteristics of the other treatments before and after fertilization are corrected to weaken the influence of the initial background differences on nitrogen application identification. The corrected canopy multispectral reflectance is used to calculate multiple vegetation indices and extract the principal components of the vegetation indices. S4: Using the modified canopy multispectral reflectance or vegetation index principal components as input data, and the grading labels of the fertilization gradient treatment corresponding to the field where the spatial sampling point is located as output data, a nitrogen application rate grading model sample set is constructed. S5: The nitrogen application rate grading model sample set is trained based on the machine learning classification algorithm to obtain the nitrogen application rate grading model. The hyperparameters involved in the grading model are optimized using the optimization algorithm.
[0009] Furthermore, the multispectral reflectance in step S2 corresponds to at least three of the following bands: green band G, red band R, red-edge band RE, and near-infrared band NIR.
[0010] Furthermore, in step S2, the canopy multispectral image data after fertilization corresponds to at least one sampling time point.
[0011] Furthermore, in step S3, the extracted reflectance of each treatment band before and after fertilization is corrected using the following formula: ; ; In the formula: for Processing Band reflectivity correction value, Before fertilization ( The processed λ-band reflectivity, Before fertilization Processing Band reflectivity, for Processing Second sampling Band reflectivity, for The correction value is n, which is an integer greater than or equal to 1, representing the number of fertilization treatments; m is an integer greater than or equal to 1, representing the number of multispectral image acquisitions after fertilization.
[0012] Preferably, when using multispectral band reflectance as input for modeling, m is greater than or equal to 2.
[0013] Preferably, when using vegetation index principal components as input for modeling, m is greater than or equal to 1.
[0014] Preferably, when using canopy spectral reflectance as input for modeling, the corrected spectral features are selected as input.
[0015] Furthermore, the multiple vegetation indices in step S3 include the Normalized Difference Vegetation Index (NDVI), Renormalized Difference Vegetation Index (RDVI), Normalized Difference Red Edge Index (NDRE), Difference Vegetation Index (DVI), Optimized Soil-Adjusted Vegetation Index (OSAVI), Normalized Difference Greenness Index (NDGI), Ratio Vegetation Index 1 (RVI1), Ratio Vegetation Index 2 (RVI2), Improved Nonlinear Vegetation Index (MNVI), and Vegetation Near-Infrared Reflectance Vegetation Index (NIR). V The vegetation index principal components are selected from at least 10 of the following: triangular vegetation index (TVI), improved triangular vegetation index (MTVI), green normalized vegetation index (GNDVI), improved ratio vegetation index (MSR), and green index (GI); the principal components of the vegetation index are extracted by extracting all principal components in descending order of eigenvalues and retaining principal components with a cumulative variance explained rate of not less than 90%.
[0016] Furthermore, the input data in step S4 must include spectral data before and after fertilization.
[0017] Furthermore, the machine learning classification algorithm in step S5 is the support vector machine algorithm, and the Bayesian optimization algorithm is used to optimize the hyperparameters during the model training process.
[0018] Beneficial effects: Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention eliminates background interference by introducing the pre-fertilization spectrum as a "self-reference benchmark" for differential correction, effectively eliminating the interference of differences in soil fertility, initial seedling size and environmental background, significantly improving the model's identification accuracy in different fields and making it highly universal.
[0019] (2) This invention introduces multi-temporal monitoring, identifies the key time window for nitrogen application after fertilization and constructs a multi-temporal feature matrix accordingly, which can quickly capture the delayed physiological response of crops to nitrogen fertilizer. Compared with monitoring throughout the entire growth period, it significantly shortens the evaluation cycle and has strong timeliness.
[0020] (3) By exploring different temporal combination strategies and using Bayesian algorithm to optimize SVM hyperparameters, this invention effectively avoids the dilution of core discrimination signals by redundant information, realizes feature space optimization, and ensures the hierarchical robustness of the model in complex field environments. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the method flow of an embodiment of the present invention. Detailed Implementation
[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings. Example 1
[0023] This embodiment discloses the specific process and effect of constructing a nitrogen application rate grading model using three-phase multispectral reflectance as an input parameter, as detailed below: S1: Data Collection and Preprocessing Paddy fields were selected as the experimental area, and four nitrogen application levels were set up: no nitrogen fertilizer (N1), reduced nitrogen fertilizer (N2), normal nitrogen fertilizer (N3), and excessive nitrogen fertilizer (N4). Crop canopy images were acquired using a multispectral camera at three time points: 1 day before fertilization, 3 days after fertilization, and 6 days after fertilization. Preprocessing of the images, including image stitching, cropping, geometric correction, radiometric calibration, and atmospheric correction, was performed to ensure data quality.
[0024] S2: Band reflectance extraction Seventeen spatial sampling points were set up in each treatment field. Reflectance data of green light (G), red light (R), red edge (RE), and near-infrared (NIR) bands were extracted from each sampling point at the three time points mentioned above.
[0025] S3: Correction for band reflectivity The extracted reflectance values for each treatment band before and after fertilization were corrected using the following formula: ; ; In the formula: for Processing Band reflectivity correction value, Before fertilization ( The processed λ-band reflectivity, Before fertilization Processing Band reflectivity, for Processing Second sampling Band reflectivity, for The correction value is n, which is an integer greater than or equal to 1, representing the number of fertilization treatments; m is an integer greater than or equal to 2, representing the number of multispectral image acquisitions after fertilization.
[0026] The reflectance of each band before and after fertilization was corrected according to the formula to eliminate the influence of background noise and environmental factors, and the corrected reflectance sequence was obtained.
[0027] S4: Constructing a hierarchical model sample set A training sample set was constructed by using the multispectral reflectance of a single spatial sampling point at three time nodes as input data and the nitrogen application treatment label (N1~N4) of the field where the sampling point is located as output.
[0028] S5: Model Building and Training A hierarchical model was constructed using the Support Vector Machine (SVM) algorithm, and the hyperparameters of the model were optimized using the Bayesian optimization algorithm.
[0029] Model Performance Comparison Analysis To investigate the impact of different band combinations on model accuracy, this embodiment also compared the model performance under different band combinations during the rice tillering and heading stages, as shown in Table 1.
[0030] Table 1. Accuracy Comparison of Hierarchical Models Based on Three-Phase Multispectral Reflectance Combinations (I-Tillering Stage) (Single-band reflectivity before correction) Results Analysis The nitrogen application rate classification model constructed using any three or more combinations of G, R, RE, and NIR as inputs during the tillering and heading stages of rice exhibits good classification accuracy. Furthermore, the model constructed using corrected reflectance as input generally demonstrates higher accuracy than the model constructed using uncorrected reflectance as input. Therefore, the classification model constructed using three or more corrected single-band reflectances as input performs well, and optimizing the SVM hyperparameters using the Bayesian optimization algorithm ensures good robustness of the model constructed using any three single-band reflectances. Example 2
[0031] This embodiment discloses the specific process and effect of constructing a nitrogen application rate classification model using multi-temporal vegetation index principal component (PCA) as input parameters.
[0032] S1: Data Collection and Preprocessing Similar to Example 1, field crop canopy multispectral image data were collected at three time points before and after the fertilization date (1 day before fertilization, 3 days after fertilization, and 6 days after fertilization) under different fertilization gradient treatments in the same area, and preprocessed.
[0033] S2: Band reflectance extraction Seventeen spatial sampling points were set up in each treatment field. Reflectance data of green light (G), red light (R), red edge (RE), and near-infrared (NIR) bands were extracted from each sampling point at the three time points mentioned above.
[0034] S3: Vegetation Index Calculation and Principal Component Extraction When using the principal component of vegetation index as the input feature, the extracted reflectance of each treatment band before and after fertilization is first corrected according to the following formula: ; ; In the formula: for Processing Band reflectivity correction value, Before fertilization ( The processed λ-band reflectivity, Before fertilization Processing Band reflectivity, for Processing Second sampling Band reflectivity, for The correction value is n, which is an integer greater than or equal to 1, representing the number of fertilization treatments; m is an integer greater than or equal to 2, representing the number of multispectral image acquisitions after fertilization.
[0035] The reflectance of each treatment before and after fertilization was corrected according to the formula to eliminate the influence of background noise and environmental factors, resulting in a corrected reflectance sequence. Then, the vegetation indices were calculated using the corrected canopy multispectral reflectance according to the calculation formulas for the 15 vegetation indices in Table 2. Principal component analysis (PCA) was used to extract the principal components of the vegetation indices in descending order of eigenvalues. In this embodiment, the first three principal components (PCA1) were extracted. VI PCA2 VI PCA3 VI Its cumulative variance explained is no less than 90%.
[0036] Table 2. Vegetation indices selected in Example 2 Table 3. Explanation of Total Variance of Principal Components of Vegetation Index S4: Constructing a hierarchical model sample set The input data for each sample consists of the principal component analysis (PCA1) values of the vegetation index at a single spatial sampling point before and after fertilization. VI PCA2 VI PCA3 VI The output is the grading label of the field where the spatial sampling point is located.
[0037] S5: Model Building and Training A hierarchical model was constructed using the Support Vector Machine (SVM) algorithm, and the hyperparameters of the model were optimized using the Bayesian optimization algorithm.
[0038] Model Performance Comparison Analysis To investigate the impact of different time point combinations on model accuracy, this embodiment compares the model performance with two-phase and three-phase inputs at the tillering and heading stages, and the results are shown in Table 4.
[0039] Table 4. Accuracy Comparison of Hierarchical Models Based on Multi-temporal Multispectral Reflectance Combinations Results Analysis During the tillering and heading stages of rice, the model constructed using the principal components of vegetation indices on day 6 after fertilization as input had higher accuracy than the model constructed using the principal components on day 3 after fertilization. This indicates that the rice canopy responds to fertilization with a delay. Furthermore, using the principal components of vegetation indices from multiple time periods as input to construct the model resulted in a better accuracy for the hierarchical model constructed using spectral indicators within a short fertilization period.
Claims
1. A method for constructing a nitrogen application rate grading model based on crop canopy multispectral density, characterized in that, include: S1: Acquire canopy multispectral image data of crops under different fertilization gradient treatments in the same area before and after fertilization, and preprocess the multispectral image data, including one or more of the following: stitching, cropping, geometric correction, radiometric calibration and atmospheric correction. S2: Based on the preprocessed multispectral image data, extract the canopy multispectral reflectance of no less than 15 spatial sampling points before and after fertilization for each processing; S3: Based on the canopy response characteristics of each treatment spatial sampling point before fertilization, taking the canopy response characteristics of one treatment as a reference, and based on the initial background differences of each treatment before fertilization, the canopy response characteristics of the other treatments before and after fertilization are corrected to weaken the influence of the initial background differences on nitrogen application identification. The corrected canopy multispectral reflectance is used to calculate multiple vegetation indices and extract the principal components of the vegetation indices. S4: Using the modified canopy multispectral reflectance or vegetation index principal components as input data, and the grading labels of the fertilization gradient treatment corresponding to the field where the spatial sampling point is located as output data, a nitrogen application rate grading model sample set is constructed. S5: The nitrogen application rate grading model sample set is trained based on the machine learning classification algorithm to obtain the nitrogen application rate grading model. The hyperparameters involved in the grading model are optimized using the optimization algorithm.
2. The method for constructing a nitrogen application rate grading model based on crop canopy multispectral density according to claim 1, characterized in that, The multispectral reflectance in step S2 corresponds to at least three of the following bands: green band G, red band R, red edge band RE, and near-infrared band NIR.
3. The method for constructing a nitrogen application rate grading model based on crop canopy multispectral density according to claim 1, characterized in that, In step S2, the canopy multispectral image data after fertilization corresponds to at least one sampling time point.
4. The method for constructing a nitrogen application rate grading model based on crop canopy multispectral density according to claim 1, characterized in that, In step S3, the extracted reflectance of each treatment band before and after fertilization is corrected according to the following formula: ； ； In the formula: for Processing Band reflectivity correction value, Before fertilization ( The processed λ-band reflectivity, Before fertilization Processing Band reflectivity, for Processing Second sampling Band reflectivity, for The correction value is n, which is an integer greater than or equal to 1, representing the number of fertilization treatments; m is an integer greater than or equal to 1, representing the number of multispectral image acquisitions after fertilization.
5. The method for constructing a nitrogen application rate grading model based on crop canopy multispectral density according to claim 1, characterized in that, The multiple vegetation indices in step S3 include NDVI, RDVI, NDRE, DVI, OSAVI, NDGI, RVI1, RVI2, MNVI, and NIR. V At least 10 of the following vegetation indices: TVI, MTVI, GNDVI, MSR, and GI; the principal components of the vegetation indices are extracted by retrieving all principal components in descending order of eigenvalues and retaining principal components with a cumulative variance explained rate of not less than 90%.
6. The method for constructing a nitrogen application rate classification model based on crop canopy multispectral density according to claim 1, characterized in that, The input data in step S4 must include spectral data before and after fertilization.
7. The method for constructing a nitrogen application rate grading model based on crop canopy multispectral density according to claim 1, characterized in that, The machine learning classification algorithm in step S5 is the support vector machine algorithm, and the Bayesian optimization algorithm is used to optimize the hyperparameters during the model training process.