Rice variable fertilization method, device, electronic equipment and storage medium

Abstract

The application provides a rice variable fertilization method and device, electronic equipment and storage medium, and relates to the technical field of agricultural intelligent information processing. The method comprises the following steps: removing non-rice plant interference in a multispectral remote sensing image to generate a rice plant canopy mask; based on the rice plant canopy mask, extracting multi-dimensional features of the rice plant canopy, and fusing to obtain a fused growth vigor feature vector; through a growth vigor grade prediction model, classifying the growth vigor of rice according to the fused growth vigor feature vector to obtain a predicted growth vigor grade; standardizing the predicted growth vigor grade to obtain a standardized growth vigor grade; based on the standardized growth vigor grade, determining a nitrogen fertilizer adjustment factor; and adjusting the standard nitrogen application amount according to the nitrogen fertilizer adjustment factor to obtain a recommended nitrogen application amount. The application can accurately recommend the nitrogen application amount in the rice planting process, thereby realizing precision fertilization.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent information processing technology, and in particular to a method, apparatus, electronic device and storage medium for variable fertilization of rice. Background Technology

[0002] Traditional nitrogen management for rice often relies on static and uniform regional recommendations, which often fail to take into account crop growth, weather conditions and soil differences in real time, resulting in low nitrogen fertilizer utilization and increased environmental burden.

[0003] In recent years, the development of UAV remote sensing technology has provided new possibilities for variable-rate fertilization of rice. However, most existing studies use inversion models that combine a single vegetation index (such as the Differential Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), and Normalized Difference Red Edge Index (NDRE)) with a specific growth indicator (such as Leaf Area Index (LAI), leaf nitrogen concentration, canopy cover, or dry matter). These methods are easily affected by environmental and background factors, resulting in a lack of robustness in diagnostic results. They also fail to comprehensively reflect the overall growth status determined by factors such as leaf area index, leaf nitrogen content, canopy height, biomass, and texture characteristics. Furthermore, rice nitrogen requirements exhibit significant temporal characteristics, while existing methods often rely on single temporal images or static indicators, neglecting the influence of crop growth dynamics and accumulated temperature processes. Additionally, the lack of integration with agronomic mechanisms leads to model results that are biased towards spectral fitting and lack explanatory power regarding nitrogen nutrition, often resulting in biased fertilization decisions. This shows that existing UAV remote sensing nitrogen management methods still lack a systematic, dynamic, and agronomically constrained diagnostic framework, which limits their application value in large-scale promotion.

[0004] Therefore, how to achieve precise fertilization of rice based on UAV multispectral images is a technical problem that urgently needs to be solved in the field of agricultural technology. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for variable fertilization of rice, which can accurately recommend the amount of nitrogen applied during the rice planting process, thereby achieving precision fertilization.

[0006] This invention provides a method for variable fertilization of rice, comprising: Non-rice plant interference objects are removed from multispectral remote sensing images to generate rice canopy masks; Based on the rice canopy mask, multidimensional features of the rice canopy are extracted and fused to obtain a fused growth feature vector. The rice growth is classified according to the fused growth feature vector using the growth level prediction model to obtain the predicted growth level. The predicted growth level is standardized to obtain the standardized growth level; Based on the standardized growth level, the nitrogen fertilizer adjustment factor is determined; Based on the nitrogen fertilizer adjustment factor, the standard nitrogen application rate is adjusted to obtain the recommended nitrogen application rate.

[0007] According to the present invention, a method for variable fertilization of rice, wherein removing non-rice plant interference in a multispectral remote sensing image to generate a rice canopy mask includes: Based on the first reflectance of each pixel in the multispectral remote sensing image on multiple first preset spectral bands, the Normalized Differential Water Index (NDWI), the Supergreen Index (ExG), and the Algae Index (Algae Index) are calculated respectively. The water body area is determined based on the NDWI value and the preset NDWI threshold; the area with missing seedlings and broken rows is determined based on the ExG value and the preset ExG threshold; and the algal patch area is determined based on the Algae Index value, the preset Algae Index threshold, the reflectance of the green light band, and the preset green light band reflectance threshold. In the multispectral remote sensing image, the water body area, the area with missing seedlings and broken rows, and the algal patch area are masked to obtain a first intermediate image; A gray-level co-occurrence matrix is ​​calculated on the first intermediate image to extract the first texture features, and the first texture features are fused with the spectral features of the first intermediate image to generate a second intermediate image; The second intermediate image is input into the pixel segmentation model to perform pixel-level segmentation of the rice canopy, and the initial rice canopy mask output by the pixel segmentation model is obtained. The initial rice canopy mask was optimized using morphological operations to obtain the rice canopy mask.

[0008] According to a variable fertilization method for rice provided by the present invention, the step of extracting multidimensional features of the rice canopy based on the rice canopy mask and fusing them to obtain a fused growth feature vector includes: Based on the rice canopy mask, multidimensional features of the rice canopy are extracted; wherein, the feature extraction dimensions include at least two of the following: spectral dimension, texture dimension, structural dimension, and temporal dimension. The multidimensional features are combined to generate an initial feature vector; Principal component analysis is used to reduce the dimensionality of the initial eigenvectors to obtain principal component eigenvectors. Based on the principal component feature vectors, the importance scores of each principal component feature are evaluated. The fused growth feature vector is determined based on the importance score.

[0009] According to the present invention, a method for variable fertilization of rice, wherein the extraction of multidimensional features of the rice canopy based on the rice canopy mask includes: Based on the rice canopy mask, the second reflectance of the pixels of the rice canopy in the multispectral remote sensing image is obtained in multiple second preset spectral bands. Based on the second reflectance, a vegetation index is calculated, the vegetation index including at least one of the Normalized Difference Vegetation Index (NDVI), the Normalized Green Difference Vegetation Index (GNDVI), and the Normalized Red Edge Index (NDRE); and / or, Based on the rice canopy mask, calculate the supergreen index ExG and the visible light band atmospheric impedance index VARI of the rice canopy in the RGB image; and / or, Based on the rice canopy mask, the RGB image is converted to grayscale, and the grayscale co-occurrence matrix is ​​calculated on the converted grayscale image to obtain the second texture feature; and / or, The RGB image is processed by Structure for Motion Recovery (SFM) to generate a Digital Surface Model (DSM). Subtracting the preset digital elevation model (DTM) from the DSM yields the canopy height model (CHM), and the rice canopy height is obtained based on the CHM and the rice canopy mask; and / or, Obtain the cumulative effective accumulated temperature, and determine the reproductive period correction function based on the cumulative effective accumulated temperature; The growth index is calculated based on the NDVI, historical NDVI, and the reproductive period correction function; The multidimensional features include multiple features such as NDVI, GNDVI, NDRE, ExG, VARI, the second texture feature, CHM, and the growth index.

[0010] According to the variable fertilization method for rice provided by the present invention, the growth level prediction model is trained in the following manner: Physiological indicators of multiple rice samples were obtained, and a comprehensive growth index of each rice sample was constructed based on the physiological indicators. Calculate the population mean and population standard deviation based on the comprehensive growth index; Based on the overall mean and overall standard deviation, the comprehensive growth index of each rice sample is mapped to the sample growth level; Obtain the sample fusion growth feature vector corresponding to each rice sample, and construct a training sample set based on the sample fusion growth feature vector and the sample growth level; The preset growth level prediction model is trained using the training sample set to obtain the growth level prediction model.

[0011] According to the present invention, a variable fertilization method for rice is provided, wherein the growth level prediction model includes multiple components, and the step of classifying rice growth according to the fused growth feature vector through the growth level prediction model to obtain the predicted growth level includes: Multiple growth level prediction models are used to predict growth level based on the fused growth feature vector, resulting in multiple classification probability vectors. The multiple classification probability vectors are weighted and summed to obtain the fusion probability vector; The growth level corresponding to the maximum probability value in the fusion probability vector is determined as the predicted growth level.

[0012] According to the present invention, a variable fertilization method for rice is provided, wherein determining a nitrogen fertilizer adjustment factor based on the standardized growth level includes: The standard canopy index is determined based on the standardized growth level. Obtain the measured canopy index, and calculate the relative growth index based on the measured canopy index and the standard canopy index; The nitrogen fertilizer adjustment factor is calculated based on the relative growth index and the preset adjustment function.

[0013] The present invention also provides a variable fertilization device for rice, comprising: The mask generation module is used to remove non-rice plant interference in multispectral remote sensing images and generate rice canopy masks. The feature fusion module is used to extract multidimensional features of the rice canopy based on the rice canopy mask and fuse them to obtain a fused growth feature vector. The grade prediction module is used to grade the rice growth based on the fused growth feature vector using the growth grade prediction model, and obtain the predicted growth grade. The grade standardization module is used to standardize the predicted growth grade to obtain a standardized growth grade; The factor determination module is used to determine the nitrogen fertilizer adjustment factor based on the standardized growth level. The nitrogen application rate adjustment module is used to adjust the standard nitrogen application rate according to the nitrogen fertilizer adjustment factor to obtain the recommended nitrogen application rate.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the variable fertilization method for rice as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the variable fertilization method for rice as described in any of the preceding claims.

[0016] The variable fertilization method, apparatus, electronic device, and storage medium for rice provided by this invention first remove non-rice plant interference from multispectral remote sensing images to generate a pure rice canopy mask, avoiding misjudgments caused by background environmental interference. Then, based on the rice canopy mask, multidimensional features of the rice canopy are extracted and fused to obtain a fused growth feature vector. This fusion of multidimensional features provides a more comprehensive reflection of the true growth status of the rice canopy. Next, a growth level prediction model is used to classify the rice growth according to the fused growth feature vector, obtaining a predicted growth level. This predicted growth level is then standardized to obtain a standardized growth level. Through the classification prediction and standardization processing of the growth level prediction model, the complex fused growth feature vector can be converted into a unified standardized growth level, thus ensuring absolute comparability of growth levels across regions and time periods. Finally, based on standardized growth levels, a nitrogen fertilizer adjustment factor is determined. Then, according to this factor, the standard nitrogen application rate is adjusted to obtain the recommended nitrogen application rate. Through this method, the nitrogen application rate can be dynamically adjusted according to the real-time status of the rice, achieving accurate recommendation of nitrogen application during rice cultivation. In summary, this invention, through the aforementioned interconnected technical steps, solves a series of technical problems, forming an accurate, robust, systematic, dynamic, and agronomically integrated variable fertilization method for rice. This method enables accurate recommendation of nitrogen application during rice cultivation, thereby achieving precision fertilization. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts of the variable fertilization method for rice provided by the present invention; Figure 2 This is the second flowchart of the variable fertilization method for rice provided by the present invention; Figure 3 This is the third flowchart of the variable fertilization method for rice provided by the present invention; Figure 4 This is the fourth flowchart of the variable fertilization method for rice provided by the present invention; Figure 5 This is a schematic diagram of the structure of the variable fertilization device for rice provided by the present invention; Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Traditional nitrogen management for rice often relies on static and uniform regional recommendations, which often fail to take into account crop growth, weather conditions and soil differences in real time, resulting in low nitrogen fertilizer utilization and increased environmental burden.

[0021] In recent years, the development of UAV remote sensing technology has provided new possibilities for variable-rate fertilization of rice. However, most existing studies use inversion models that combine a single vegetation index (such as the Differential Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), and Normalized Difference Red Edge Index (NDRE)) with a specific growth indicator (such as Leaf Area Index (LAI), leaf nitrogen concentration, canopy cover, or dry matter). These methods are easily affected by environmental and background factors, resulting in a lack of robustness in diagnostic results. They also fail to comprehensively reflect the overall growth status determined by factors such as leaf area index, leaf nitrogen content, canopy height, biomass, and texture characteristics. Furthermore, rice nitrogen requirements exhibit significant temporal characteristics, while existing methods often rely on single temporal images or static indicators, neglecting the influence of crop growth dynamics and accumulated temperature processes. Additionally, the lack of integration with agronomic mechanisms leads to model results that are biased towards spectral fitting and lack explanatory power regarding nitrogen nutrition, often resulting in biased fertilization decisions. This shows that existing UAV remote sensing nitrogen management methods still lack a systematic, dynamic, and agronomically constrained diagnostic framework, which limits their application value in large-scale promotion.

[0022] Therefore, how to achieve precise fertilization of rice based on UAV multispectral images is a technical problem that urgently needs to be solved in the field of agricultural technology.

[0023] Analysis shows that the following factors can affect the precision fertilization of rice.

[0024] First, image preprocessing and background interference removal methods are needed to address the unique growth environment of rice. Rice paddies are often flooded, leading to water reflection, algae cover, and uneven sowing resulting in missing seedlings or broken rows. These factors interfere with spectral and texture features, increasing the deviation of vegetation indices and canopy parameters. Existing studies mostly rely on empirical methods or simple thresholds for vegetation extraction, lacking systematic background removal strategies and multi-source feature alignment processes. This results in mixed vegetation pixels, high noise, and poor model reproducibility, directly affecting the accuracy of growth inversion and the robustness of fertilization decisions.

[0025] Second, a systematic growth grading model is needed. Crop growth is determined by multiple dimensions, including leaf area index, aboveground biomass, canopy height, leaf nitrogen concentration, and canopy texture. However, existing studies mostly rely on single vegetation indices or empirical thresholds to achieve growth inversion or grading, rarely fusing phenotypic texture features derived from RGB images with multispectral indices. These methods may be effective in small areas or specific growth stages, but their diagnosis is biased towards chlorophyll-like features, greatly affected by background, light, and canopy structure, and lacks stability. Furthermore, there is a lack of systematic algorithms for comprehensively fusing, reducing dimensionality, and robustly modeling multiple dimensions using statistical learning or machine learning, causing grading results to easily fail in complex field environments.

[0026] Third, dynamic adaptability and temporal diagnostic capability. Rice exhibits significant differences in spectral sensitivity at different growth stages, but most methods do not dynamically select / switch diagnostic indicators based on accumulated effective temperature or multi-temporal data. Instead, they primarily rely on single-temporal or static thresholds, neglecting the temporal characteristics of crop growth and development. This makes it difficult to reflect the dynamic nitrogen requirements of rice and guide variable fertilization.

[0027] Fourth, the dynamic correlation between growth grading, nitrogen diagnosis, and precision fertilization. Most existing methods remain at the nitrogen diagnosis level, with model results biased towards spectral fitting and lacking constraints from agronomic mechanisms. They are difficult to couple with target yield methods, staged nitrogen allocation, or yield-nitrogen response curves, and it is difficult to quantify the grading results into specific nitrogen application amounts for each growth stage. As a result, a closed loop of "growth diagnosis → quantitative fertilization → execution" cannot be formed, limiting the algorithmic and generalization capabilities of prescription maps.

[0028] Based on the above, this invention proposes a method, apparatus, electronic device, and storage medium for variable fertilization of rice. The following is a detailed description... Figures 1-6 Describe it.

[0029] Figure 1 This is one of the flowcharts illustrating the variable fertilization method for rice provided by this invention, such as... Figure 1 As shown, the variable fertilization method for rice includes steps S110, S120, S130, S140, S150 and S160.

[0030] Step S110: Remove non-rice plant interference in the multispectral remote sensing image to generate a rice plant canopy mask.

[0031] Multispectral remote sensing images can be obtained by taking aerial photos of target fields during the rice planting process using drones equipped with multispectral cameras.

[0032] Furthermore, drones can also be equipped with RGB (Red-Green-Blue) cameras to capture RGB images from the air for subsequent analysis.

[0033] Non-rice plant disturbances may include, but are not limited to: water bodies, bare soil, and green algae.

[0034] In one implementation, NDWI (Normalized Difference Water Index) can be used to identify and remove obvious water bodies such as ditches and puddles in the field.

[0035] In another implementation, ExG (Excess Green Index) is used to identify and remove bare soil areas caused by missing seedlings or broken rows.

[0036] In another implementation, a specially constructed Algae Index is used in combination with green light band reflectance to determine algal patch regions under dual conditions in order to eliminate green algae interference.

[0037] Furthermore, after masking water areas, areas with missing seedlings and broken rows, and algal patches in the multispectral remote sensing image to obtain a first intermediate image, a gray-level co-occurrence matrix is ​​calculated on the first intermediate image to extract texture features. These texture features are then fused with the spectral features of the first intermediate image to generate a second intermediate image. The second intermediate image is then input into a pixel segmentation model to perform pixel-level segmentation of the rice canopy, resulting in a rice canopy mask output by the pixel segmentation model. The pixel segmentation model can be a U-Net (U-shaped network) model.

[0038] Furthermore, considering that the rice canopy mask output by the pixel segmentation model may have some small noise, internal voids, or uneven edges, morphological operations can be used to optimize the rice canopy mask output by the pixel segmentation model to obtain the final rice canopy mask.

[0039] Using the above method, the pure rice canopy region can be accurately separated from complex raw multispectral remote sensing images, providing a clean and accurate analysis object for subsequent feature extraction.

[0040] Step S120: Based on the rice canopy mask, extract the multidimensional features of the rice canopy and fuse them to obtain a fused growth feature vector.

[0041] Within the rice canopy region defined by the aforementioned rice canopy mask, multidimensional features of the rice canopy are extracted. The feature extraction dimensions include at least two of the following: spectral, texture, structural, and temporal dimensions. Correspondingly, the multidimensional features include at least two of the following: spectral features, texture features, structural features, and temporal features.

[0042] Spectral features include, but are not limited to: NDVI (Normalized Difference Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), and NDRE (Normalized Difference Red Edge Index).

[0043] Texture features include, but are not limited to: ExG (Excess Green Index), VARI (Visible Atmospherically Resistant Index), contrast, energy, and homogeneity.

[0044] Structural features include rice canopy height, which is obtained based on the CHM (Canopy Height Model).

[0045] The temporal characteristics include the growth index, which is calculated based on the reproductive period correction function constructed from the accumulated effective temperature, and NDVI.

[0046] After extracting the multidimensional features of the rice plant canopy, the multidimensional features are fused to obtain a fused growth feature vector.

[0047] In one implementation, multidimensional features are combined into a high-dimensional initial feature vector. Then, PCA (Principal Component Analysis) is used to reduce the dimensionality and decorrelate the initial feature vector, and the importance scores of each principal component feature are evaluated. Based on the importance, the fused growth feature vector is determined.

[0048] By using the above methods, the growth status of rice can be comprehensively quantified from multiple dimensions, avoiding the one-sidedness of a single indicator and forming a comprehensive growth characteristic.

[0049] Step S130: Using the growth level prediction model, the rice growth is classified according to the fused growth feature vector to obtain the predicted growth level.

[0050] The fused growth feature vector is input into the growth level prediction model, which then classifies the rice growth of each pixel to obtain the predicted growth level. For example, 1 (weak), 2 (slightly weak), 3 (standard), 4 (slightly strong), and 5 (strong). This generates a predicted growth level map covering the entire target field, denoted as L. raw .

[0051] The above method can transform complex remote sensing feature vectors into intuitive and easy-to-understand growth levels.

[0052] Step S140: Standardize the predicted growth level to obtain a standardized growth level.

[0053] To ensure comparability of results across different plots and observation flights, the predicted growth grades were standardized to form a unified growth grade, which is denoted as the standardized growth grade.

[0054] Specifically, the predicted growth level map L generated in the previous step is calculated. raw The mean (denoted as the current mean) and standard deviation (denoted as the current standard deviation) are calculated. Then, based on the current mean, current standard deviation, preset standardized mean, and preset standardized standard deviation, the standardized growth level corresponding to the predicted growth level of each pixel is calculated. The specific calculation formula is as follows: ; in, This represents the normalized growth level of the i-th pixel. This represents the predicted growth level of the i-th pixel. This represents the current mean. Indicates the current standard deviation. This indicates the preset standardized mean. The standard deviation is set to the pre-defined standard deviation.

[0055] By using the above methods, systematic biases caused by different plots and different observation flights can be eliminated, making the growth level absolutely comparable across regions and time.

[0056] Step S150: Determine the nitrogen fertilizer adjustment factor based on the standardized growth level.

[0057] First, the measured canopy index is determined based on the standardized growth level. Then, the measured canopy index is compared with the standard canopy index corresponding to the current growth stage to calculate the relative growth index. Finally, the relative growth index is substituted into the preset adjustment function to calculate the nitrogen fertilizer adjustment factor.

[0058] Step S160: Adjust the standard nitrogen application rate according to the nitrogen fertilizer adjustment factor to obtain the recommended nitrogen application rate.

[0059] The standard nitrogen application rate refers to the nitrogen application rate corresponding to the current growth stage. The specific method for obtaining this rate is as follows: By combining historical data from multi-year, multi-location nitrogen application gradient experiments in the same region, a model relating rice yield to nitrogen application rate is established. Specifically, a univariate quadratic regression model can be used: ; Where Y represents rice yield, N represents nitrogen application rate, and a, b, and c are regression parameters. The optimal nitrogen application rate N can be calculated by differentiating this model. opt The details are as follows: ; This optimal nitrogen application rate is used as the standard nitrogen level, and a baseline fertilization plan is determined in conjunction with the target yield method. Furthermore, a staged nitrogen allocation coefficient is introduced to decompose the total nitrogen amount into key growth stages such as tillering, jointing, and heading stages, in order to calculate the nitrogen application rate for each growth stage. Details are as follows: ; in, This represents the standard nitrogen application rate for the i-th stage.

[0060] After determining the nitrogen fertilizer adjustment factor, the standard nitrogen application rate is adjusted to obtain the recommended nitrogen application rate. Recommended nitrogen application rate = Standard nitrogen application rate × Nitrogen fertilizer adjustment factor.

[0061] Furthermore, the recommended nitrogen application rate is mapped to the field space for each raster cell, generating a variable fertilizer prescription map, and outputting a standard file compatible with drone sprayers or variable fertilizer applicators to achieve precise variable fertilizer application.

[0062] The variable fertilization method for rice provided in this invention first removes non-rice plant interference from multispectral remote sensing images to generate a pure rice canopy mask, avoiding misjudgments caused by background environmental interference. Then, based on the rice canopy mask, multidimensional features of the rice canopy are extracted and fused to obtain a fused growth feature vector. This fusion of multidimensional features provides a more comprehensive reflection of the true growth status of the rice canopy. Next, a growth level prediction model is used to classify the rice growth according to the fused growth feature vector, obtaining a predicted growth level. This predicted growth level is then standardized to obtain a standardized growth level. Through the classification prediction and standardization processing of the growth level prediction model, the complex fused growth feature vector can be converted into a unified standardized growth level, thus ensuring absolute comparability of growth levels across regions and time periods. Finally, based on standardized growth levels, a nitrogen fertilizer adjustment factor is determined. Then, according to this factor, the standard nitrogen application rate is adjusted to obtain the recommended nitrogen application rate. Through this method, the nitrogen application rate can be dynamically adjusted according to the real-time status of the rice, achieving accurate recommendation of nitrogen application during rice cultivation. In summary, this invention, through the aforementioned interconnected technical steps, solves a series of technical problems, forming an accurate, robust, systematic, dynamic, and agronomically integrated variable fertilization method for rice. This method can accurately recommend nitrogen application rates during rice cultivation, thereby achieving precision fertilization.

[0063] Based on any of the above embodiments Figure 2 This is the second flowchart of the variable fertilization method for rice provided by the present invention, as shown below. Figure 2 As shown, step S110 includes: step S111, step S112, step S113, step S114, step S115 and step S116.

[0064] Step S111: Based on the first reflectance of each pixel in the multispectral remote sensing image on multiple first preset spectral bands, calculate the Normalized Difference Water Index (NDWI), the Supergreen Index (ExG), and the Algae Index (Algae Index) respectively.

[0065] The first preset spectral bands include the red light band, green light band, blue light band, and near-infrared band.

[0066] Based on the reflectance of each pixel in the multispectral remote sensing image across the four bands (denoted as the first reflectance), the NDWI value, ExG value, and Algae Index value are calculated respectively. The reflectance in the red band is denoted as ρ. Red1 The reflectivity of the green light band is denoted as ρ. Green1 The reflectivity of the blue light band is denoted as ρ. Blue1 Near-infrared reflectance is denoted as ρ. NIR1 .

[0067] The NDWI value is calculated as follows: NDWI = (ρ Green1 -ρ NIR1 ) / (ρ Green1 +ρ NIR1 ); The ExG value is calculated as follows: ExG = 2 × ρ Green1 -ρ Red1 -ρ Blue1 ; The Algae Index value is calculated as follows: .

[0068] Step S112: Determine the water body area based on the NDWI value and the preset NDWI threshold, determine the area with missing seedlings and broken rows based on the ExG value and the preset ExG threshold, and determine the algal patch area based on the Algae Index value, the preset Algae Index threshold, the reflectance of the green light band, and the preset green light band reflectance threshold.

[0069] Set a preset NDWI threshold, such as 0.3, and identify all pixels in the multispectral remote sensing image with NDWI>0.3 as water areas.

[0070] Set a preset ExG threshold, such as 0.1, and identify all pixels in the multispectral remote sensing image with ExG < 0.1 as areas of missing seedlings and broken rows. Areas of missing seedlings and broken rows are bare soil areas caused by missing seedlings and broken rows.

[0071] Set a preset Algae Index threshold, for example, 0.3, and a preset green band reflectance threshold, for example, 0.4. When a pixel simultaneously satisfies Algae Index > 0.3 and ρ... Green1 When both conditions are greater than 0.4, the area is identified as an algal patch. Using these dual criteria, green algae that are easily confused with rice paddies can be accurately identified.

[0072] Step S113: In the multispectral remote sensing image, the water body area, the area with missing seedlings and broken rows, and the algal patch area are masked to obtain a first intermediate image.

[0073] In the original multispectral remote sensing image, all pixels in the water body areas, areas with missing seedlings and broken ridges, and algal patches identified in the above steps are masked, for example, by assigning a value of 0 or marking it as an invalid value. The resulting image is denoted as the first intermediate image.

[0074] Step S114: Calculate the gray-level co-occurrence matrix for the first intermediate image to extract the first texture features, and fuse the first texture features with the spectral features of the first intermediate image to generate the second intermediate image.

[0075] Although the first intermediate image removed most of the non-rice plant interference, blurred boundary areas such as rice plants, shadows, and damp soil still existed. Therefore, further texture feature enhancement processing was performed on the first intermediate image.

[0076] The Gray-Level Co-occurrence Matrix (GLCM) is calculated for the first intermediate image. GLCM is a classic method for analyzing image texture. The GLCM is calculated within a neighborhood window (e.g., 7x7 pixels) centered on each pixel, and multiple texture descriptors, such as contrast, energy, and homogeneity, are derived from this, denoted as the first texture feature.

[0077] The extracted first texture features are used as one or more new feature channels and stacked and fused with the original spectral channels of the first intermediate image to generate an image containing both spectral and texture information, denoted as the second intermediate image.

[0078] By introducing texture information from the image, the separability of the rice canopy and the remaining suspected background in the feature space is enhanced, especially the distinguishability of blurred boundaries, providing richer discrimination criteria for subsequent pixel segmentation models.

[0079] Step S115: Input the second intermediate image into the pixel segmentation model to perform pixel-level segmentation of the rice canopy and obtain the initial rice canopy mask output by the pixel segmentation model.

[0080] The second intermediate image is input into a pre-trained pixel segmentation model. The pixel segmentation model classifies each pixel in the second intermediate image and outputs a probability map, where the value of each pixel represents the probability that it belongs to the rice canopy layer. A threshold is set on the probability map output by the pixel segmentation model, for example, 0.5. Pixels with a probability greater than 0.5 are classified as rice canopy layer, and the rest are background. This results in a binary rice canopy layer mask, denoted as the initial rice canopy layer mask.

[0081] Furthermore, the U-Net model is preferred for pixel segmentation. U-Net is a convolutional neural network designed for image segmentation, whose unique encoder-decoder structure and skip connections make it excellent at accurately recovering object boundary details.

[0082] Step S116: The initial rice canopy mask is optimized through morphological operations to obtain the rice canopy mask.

[0083] Considering that the initial rice canopy mask may contain some small noise points or tiny internal voids, morphological operations were further employed for optimization to address this issue.

[0084] Specifically, a closing operation is performed to fill the small internal voids, and an opening operation is performed to eliminate isolated external noise points, resulting in smoother mask edges and a more complete internal structure. After morphological optimization, the final high-precision rice canopy mask is obtained.

[0085] The variable fertilization method for rice provided in this invention first removes major non-rice plant interference based on multispectral features. Then, it enhances the texture features of the image after removing non-rice plant interference to enhance other suspected background and rice plant boundaries. Finally, it performs fine segmentation using a pixel segmentation model based on rich spectral and texture features to ensure boundary integrity. Morphological operations are then used to optimize the boundaries, forming the final rice canopy mask. This method generates a high-precision rice canopy mask, providing a reliable key data foundation for subsequent growth monitoring, variable fertilization, and other analyses.

[0086] Based on any of the above embodiments Figure 3 This is the third flowchart of the variable fertilization method for rice provided by the present invention, as shown below. Figure 3 As shown, step S120 includes: step S121, step S122, step S123, step S124 and step S125.

[0087] Step S121: Based on the rice canopy mask, extract multidimensional features of the rice canopy; wherein, the feature extraction dimensions include at least two of the following: spectral dimension, texture dimension, structural dimension, and temporal dimension.

[0088] Based on the rice canopy mask, multidimensional features of the rice canopy are extracted; the feature extraction dimensions include at least two of the following: spectral dimension, texture dimension, structural dimension, and temporal dimension.

[0089] Correspondingly, multidimensional features include at least two of the following: spectral features, texture features, structural features, and temporal features.

[0090] Spectral features include, but are not limited to: NDVI, GNDVI, NDRE; textural features include, but are not limited to: ExG, VARI, contrast, energy, and homogeneity; structural features include rice plant canopy height; temporal features include growth index.

[0091] Step S122: Combine the multidimensional features to generate an initial feature vector.

[0092] For any pixel within the rice canopy mask, its multidimensional features are combined into a high-dimensional feature vector X = (x1, x2, ..., x...). n ), where each element x1-x n Each of these represents a quantized value of a feature. To distinguish it from other subsequent feature vectors, the feature vector obtained by combining these multi-dimensional features is denoted as the initial feature vector.

[0093] Step S123: The initial feature vector is reduced in dimensionality using principal component analysis to obtain the principal component feature vector.

[0094] PCA (Principal Component Analysis) is a statistical method that aims to transform a set of potentially correlated variables into a set of linearly uncorrelated variables through linear transformation. These uncorrelated variables are called principal components.

[0095] The matrix formed by the initial feature vectors of all pixels is input into the PCA model. After calculation by the PCA model, a set of new, orthogonal (linearly independent) principal components is output. Each principal component is a representation of the original feature vectors (x1, x2, ..., x...). n A linear combination of ). For example, the first principal component PC1 = a1x1 + a2x2 + ... + a n x nTypically, the first few principal components (such as PC1, PC2, PC3) can explain the vast majority (e.g., over 95%) of the variance in the original data. After the above processing, the initial feature vector of each pixel is transformed from the original X into a principal component feature vector P = (p1, p2, ..., p3) composed of principal component values. k ), where p k This represents the score of the pixel on the k-th principal component, where k ≤ n.

[0096] Step S124: Based on the principal component feature vector, evaluate the importance score of each principal component feature.

[0097] When assessing the importance of features, the Random Forest (RF) model can be used. The Random Forest model constructs multiple decision trees and averages their results for regression or classification, and is often used to evaluate the importance of features.

[0098] The principal component feature vectors are input into a random forest model to evaluate the importance of each principal component, resulting in an importance score for each principal component feature. This importance score represents its contribution to rice growth.

[0099] Furthermore, the training process of the random forest model is as follows: Through multi-dimensional sample features, the sample principal component feature vector is generated, and the ground-measured growth data is obtained as sample label. The initial random forest model is trained using the sample principal component feature vector and the sample label to obtain the trained random forest model.

[0100] Step S125: Determine the fused growth feature vector based on the importance score.

[0101] The importance scores of all principal component features are sorted, and the top m principal component features with the highest scores or those exceeding a preset score threshold are selected. These m features are then used to construct the final fused growth potential feature vector Z = (z1, z2, ..., zm). m ), where m≤k≤n.

[0102] The fusion of growth feature vectors yields the final result that can comprehensively, accurately, and efficiently characterize the overall growth of rice.

[0103] The variable fertilization method for rice provided in this invention first uses PCA to reduce dimensionality, eliminating the collinearity problem among the original multidimensional features and improving the stability of the subsequent analysis model. Furthermore, by combining the importance assessment of RF (Radical Randomization), it ensures that the final selected fused growth feature vector is the core information most relevant to the true growth of rice, discarding noise and secondary factors, making the subsequent growth grading results more accurate and reliable.

[0104] Based on any of the above embodiments, step S121 may include steps S1211-S1212, and / or, steps S1213 and / or, steps S1214, and / or, steps S1215-S1216, and / or, steps S1217-S1218.

[0105] Step S1211: Based on the rice canopy mask, obtain the second reflectance of the pixels of the rice canopy in the multispectral remote sensing image on multiple second preset spectral bands.

[0106] Step S1212: Based on the second reflectance, calculate the vegetation index, which includes at least one of the Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), and Normalized Red Edge Index (NDRE).

[0107] Based on the rice canopy mask, the reflectance of the rice canopy in multiple second preset spectral bands is extracted from the multispectral remote sensing image for only the pixels within the mask, and is denoted as the second reflectance.

[0108] The second preset spectral bands include the red light band, green light band, red edge band, and near-infrared band.

[0109] Based on the aforementioned second reflectance, nitrogen-sensitive vegetation indices are calculated, specifically including at least one of NDVI, GNDVI, and NDRE. The red light band reflectance is denoted as ρ. Red2 The reflectivity of the green light band is denoted as ρ. Green2 The reflectivity of the red-edge band is denoted as ρ. RedEdge2 Near-infrared reflectance is denoted as ρ. NIR2 .

[0110] NDVI is calculated as follows: NDVI = (ρ NIR2 -ρ Red2 ) / (ρ NIR2 +ρ Red2 ); GNDVI is calculated as follows: GNDVI = (ρ NIR2 -ρ Green2 ) / (ρ NIR2 +ρ Green2 ); The calculation method for NDRE is: NDRE = (ρ NIR2 -ρ RedEdge2 ) / (ρ NIR2 +ρ RedEdge2 ).

[0111] Step S1213: Based on the rice canopy mask, calculate the supergreen index ExG and visible light band atmospheric impedance index VARI of the rice canopy in the RGB image.

[0112] Step S1214: Based on the rice canopy mask, the RGB image is converted to grayscale, and the grayscale co-occurrence matrix is ​​calculated on the converted grayscale image to obtain the second texture feature.

[0113] RGB images can be obtained by taking aerial photos of the target field during rice planting using a drone equipped with an RGB camera, or by extracting and synthesizing the red, green, and blue light bands of a multispectral image.

[0114] Based on the rice canopy mask, calculate the ExG and VARI of the rice canopy in the RGB image.

[0115] The calculation method for ExG is: ExG = 2 × ρ Green3 -ρ Red3 -ρ blue3 The calculation method for VARI is: VARI = (ρ Green3 -ρ Red3 ) / (ρ Green3 +ρ Red3 -ρ blue3 ), where ρ Green3 ρ represents the reflectance of the green band in an RGB image. Red3 ρ represents the reflectance of the red band in an RGB image. blue3 This represents the blue light band reflectance of the RGB image.

[0116] Then, based on the rice canopy mask, the RGB image within the mask is converted to grayscale. The grayscale co-occurrence matrix (GLCM) is then calculated on the converted grayscale image, and the texture features are extracted and denoted as the second texture features. The second texture features include, but are not limited to, contrast, energy, homogeneity, etc.

[0117] It's worth noting that multispectral images have the advantage of strong recognition capabilities. They can capture information such as near-infrared light, which is invisible to the naked eye. Using indices such as NDVI and NDRE, they can very accurately distinguish rice from non-rice plant interference, and the masks generated by them are more accurate. RGB images, on the other hand, have the advantage of rich texture information, higher resolution, and clearer details. They can better reflect the spatial texture, color variations, and structural morphology of the rice canopy, and can be used to calculate color indices such as ExG and VARI, which are entirely based on the visible light band.

[0118] Step S1215: Perform structure-of-motion (SFM) processing on the RGB image to generate a digital surface model (DSM); Step S1216: Subtract the preset digital elevation model (DTM) from the DSM to obtain the canopy height model (CHM), and obtain the rice canopy height based on the CHM and the rice canopy mask.

[0119] The RGB image is processed using SFM (Structure from Motion) to generate a high-precision DSM (Digital Surface Model), which represents the elevation model of the tops of all objects on the ground surface. Then, a pre-acquired DTM (Digital Terrain Model) is subtracted from the DSM to obtain the CHM (Canopy Height Model), where the DTM represents the elevation model of the bare ground surface. Finally, based on the rice canopy mask, the pure rice canopy height is extracted from the CHM as a structural feature.

[0120] Step S1217: Obtain the cumulative effective accumulated temperature, and determine the reproductive period correction function based on the cumulative effective accumulated temperature.

[0121] Step S1218: Calculate the growth index based on the NDVI, historical NDVI, and the reproductive period correction function.

[0122] The multidimensional features include multiple features such as NDVI, GNDVI, NDRE, ExG, VARI, the second texture feature, CHM, and the growth index.

[0123] Obtain daily meteorological data from the rice sowing date to the date of multispectral / RGB image observation, and calculate the accumulated effective temperature (AGDD). The specific formula is as follows: ; Where n represents the total number of days from the rice sowing date to the date of multispectral / RGB image observation. This represents the average temperature on the i-th day from the rice sowing date. This indicates the crop's reference temperature. It was determined based on historical agronomic data (including crop sowing date and crop variety).

[0124] Then, the reproductive period correction function is determined based on the accumulated effective temperature. The details are as follows: in, This represents the threshold of cumulative effective temperature from the rice sowing date to the current crop growth stage. This represents the upper limit of the cumulative effective temperature from the rice sowing date to the current crop growth stage, i.e., the cumulative temperature at which the crop's growth potential reaches its maximum. This represents the cumulative effective temperature on day t from the rice sowing date.

[0125] Then, the growth index is calculated based on NDVI, historical NDVI, and the reproductive period correction function. The specific formula is as follows: ; in, This represents the growth index on day t from the rice sowing date. This represents the vegetation index on day t, starting from the rice sowing date. The minimum vegetation index in historical observations. This represents the highest vegetation index observed in history. and It is determined based on historical data or prior knowledge.

[0126] The variable fertilization method for rice provided in this invention extracts multidimensional features from four different dimensions: spectrum, texture, structure, and time sequence, and constructs a comprehensive and three-dimensional description of rice growth. It then organically integrates the multidimensional features, effectively overcoming the bottleneck of insufficient growth recognition ability of a single spectral index, and improving the distinguishability and regional adaptability of different growth gradients.

[0127] Based on any of the above embodiments Figure 4 This is the fourth flowchart of the variable fertilization method for rice provided by the present invention, as shown below. Figure 4 As shown, the growth level prediction model is trained through the following steps: steps S10, S20, S30, S40 and S50.

[0128] Step S10: Obtain physiological indicators of multiple rice samples, and construct a comprehensive growth index for each rice sample based on the physiological indicators.

[0129] In experimental fields with different growth gradients, multiple sampling plots were set up, for example, 30 1m×1m sampling plots. Then, physiological indicators were measured and recorded in each sampling plot. Among them, physiological indicators that can represent the growth of rice were selected, including plant height, LAI (Leaf Area Index), aboveground biomass, and leaf nitrogen concentration.

[0130] Furthermore, after obtaining physiological indicators from multiple rice samples, the physiological indicators can be standardized using Z-scores to eliminate dimensional differences.

[0131] Then, for each rice sample, its Z-score-standardized physiological indicators were weighted and summed to obtain the comprehensive growth index of each rice sample. The specific formula is as follows: ; in, This represents the comprehensive growth index of the i-th rice sample; This represents the standardized value of the i-th rice sample under the j-th physiological indicator; The weight of the j-th physiological indicator can be determined by expert experience, principal component analysis, or random forest importance; n represents the number of physiological indicators.

[0132] Step S20: Calculate the overall mean and overall standard deviation based on the comprehensive growth index.

[0133] Then, based on the comprehensive growth index, the mean (denoted as the population mean) and standard deviation (denoted as the population standard deviation) of all rice samples are calculated.

[0134] Step S30: Based on the overall mean and overall standard deviation, map the comprehensive growth index of each rice sample to the sample growth level.

[0135] Based on the overall mean μ and the overall standard deviation σ, the comprehensive growth index of each rice sample is mapped to the sample growth level.

[0136] The growth level can be divided into 3 to 5 levels. Taking the growth level as 5 levels as an example, the mapping relationship is as follows: Based on the above mapping relationship, the comprehensive growth index of each rice sample is mapped to the corresponding growth level, which is denoted as the sample growth level.

[0137] Step S40: Obtain the sample fusion growth feature vector corresponding to each rice sample, and construct a training sample set based on the sample fusion growth feature vector and the sample growth level.

[0138] Step S50: Train the preset growth level prediction model using the training sample set to obtain the growth level prediction model.

[0139] While obtaining the growth level of the samples, a sample-fused growth feature vector is also obtained for each rice sample. Then, a training sample set is constructed based on the sample-fused growth feature vector and the sample growth level, where the sample growth level serves as the sample label. The preset growth level prediction model is trained using the training sample set to obtain the growth level prediction model.

[0140] The preset growth level prediction model can include one or more, and can be one or more of the following: random forest model, support vector machine (SVM) model, and eXtreme GradientBoosting (XGBoost) model. It can be understood that if multiple models are included, each preset growth level prediction model is trained separately.

[0141] The variable fertilization method for rice provided in this invention determines the growth level of a sample by using a comprehensive growth index fused from multiple measured physiological indicators. Compared to traditional subjective and singular level evaluation methods, this ensures the scientific validity and accuracy of the model training labels. Furthermore, growth grading modeling is completed based on multi-source sample fusion growth feature vectors and pre-defined growth level prediction models such as RF, SVM, and XGBoost, achieving a quantitative mapping from spectral signals to agronomic indicators. The growth level prediction model constructed in this way significantly improves accuracy and transferability, fully utilizing the comprehensive features constructed from multi-dimensional information for training and dynamic updates. This ensures the accuracy and consistency of growth assessment under different growth stages and environmental conditions, enhancing the model's robustness and generalization ability.

[0142] Based on any of the above embodiments, the growth level prediction model includes multiple steps, and step S130 includes: step S131, step S132 and step S133.

[0143] Step S131: Using multiple growth level prediction models, growth level prediction is performed based on the fused growth feature vector to obtain multiple classification probability vectors.

[0144] The growth level prediction model includes multiple models. For any pixel, its fused growth feature vector Z is input into multiple growth level prediction models to obtain the classification probability vector output by each model.

[0145] Furthermore, the growth level prediction models include RF models, SVM models, and XGBoost models. For example, taking a growth level classification of 5 levels, the classification probability vectors output by each model are as follows: P RF =[0.05, 0.1, 0.7, 0.1, 0.05]; P SVM =[0.1, 0.15, 0.6, 0.1, 0.05]; P XGB =[0.0, 0.2, 0.65, 0.1, 0.05]; Among them, P RFThis represents the classification probability vector output by the RF model, which considers level 3 to have the highest probability; P SVM P represents the classification probability vector output by the SVM model, which considers level 3 to have the highest probability; XGB This represents the classification probability vector output by the XGBoost model, which also considers level 3 to have the highest probability.

[0146] Step S132: The multiple classification probability vectors are weighted and summed to obtain the fusion probability vector.

[0147] The fusion probability vector is obtained by weighted summing of multiple classification probability vectors. The specific formula is as follows: ; in, Let w1 represent the fusion probability vector, w2 represent the weight coefficients of the RF model in the weighted fusion, w3 represent the weight coefficients of the SVM model in the weighted fusion, and w1+w2+w3=1. w1, w2, and w3 can be determined based on the accuracy of each model.

[0148] For example, if w1=0.3, w2=0.25, w3=0.45, according to P in the above example... RF P SVM and P XGB It can be calculated =[0.04, 0.1575, 0.6525, 0.1, 0.05].

[0149] Step S133: The growth level corresponding to the maximum probability value in the fusion probability vector is determined as the predicted growth level.

[0150] The growth level corresponding to the maximum probability value in the fusion probability vector is determined as the predicted growth level.

[0151] For example, in the above example, the third element in the fusion probability vector has the largest value of 0.6525, which corresponds to a growth level of "Level 3 (Standard)". Therefore, the final predicted growth level of this pixel is L. raw It was determined to be 3.

[0152] The variable fertilization method for rice provided in this invention integrates the advantages of multiple models through an ensemble learning strategy, effectively reducing the risk of misjudgment caused by defects in a single model or data noise, thereby improving the accuracy and robustness of predicting growth level.

[0153] Based on any of the above embodiments, step S150 includes: step S151, step S152 and step S153.

[0154] Step S151: Determine the standard canopy index based on the standardized growth level.

[0155] Based on the standardized growth level and the mapping relationship between the preset comprehensive growth index and the growth level, the standard canopy index is determined, which corresponds to the canopy development status under the target nitrogen level.

[0156] Step S152: Obtain the measured canopy index, and calculate the relative growth index based on the measured canopy index and the standard canopy index.

[0157] Next, the measured canopy index is obtained, which represents the actual growth condition. Then, the relative growth index is calculated based on the measured canopy index and the standard canopy index. The specific formula is as follows: ; Wherein, DI represents the relative growth index, indicating the degree of deviation of the current growth from the standard growth; CI opt CI represents the measured canopy index. std This indicates the standard canopy index.

[0158] Step S153: Calculate the nitrogen fertilizer adjustment factor based on the relative growth index and the preset adjustment function.

[0159] Then, the relative growth index is substituted into the preset adjustment function to calculate the nitrogen fertilizer adjustment factor. The details are as follows: ; in, represents the nitrogen adjustment factor, which corrects the standard nitrogen application rate based on the degree of deviation from the growth pattern; f() represents the preset adjustment function, which can be linear or piecewise linear, or determined by regression / empirical curves.

[0160] Furthermore, considering that factors such as pests and diseases, water stress, or salinity can appear similar to nitrogen deficiency in UAV multispectral images, and that existing technologies lack multi-source information verification and secondary diagnostic mechanisms, which can easily lead to misjudgments and fertilization errors, this invention automatically triggers a secondary diagnostic mechanism when the relative growth index is below a critical threshold. Specifically, a deviation of 10%-15% can be used as the critical threshold, meaning the critical threshold can be set to any value between 0.85 and 0.9. When the secondary diagnostic mechanism is triggered, a detection command can be sent to the management terminal, enabling managers to conduct manual inspections and obtain ground monitoring indicators, such as SPAD (Soil and Plant Analysis Development) values, NNI (Nitrogen Nutrition Index), and canopy temperature. Based on these ground monitoring indicators, it can then be determined whether the stress is non-nitrogenous. Non-nitrogenous stress types include, but are not limited to, pests and diseases, salinity, or water problems. Specific judgment rules can be set according to actual needs and are not specifically limited here. If the stress is determined to be non-nitrogen stress, the nitrogen fertilizer adjustment factor will be adjusted accordingly; if the stress is determined to be non-nitrogen stress, the nitrogen fertilizer adjustment factor will be adjusted to 1, and the standard nitrogen application rate will not be adjusted.

[0161] The above methods can effectively avoid fertilization errors caused by misjudging nitrogen stress, and improve the accuracy of growth monitoring and decision support capabilities.

[0162] The variable fertilization method for rice provided in this invention calculates the relative growth index and then determines the nitrogen fertilizer adjustment factor to determine the recommended nitrogen application rate. This allows the nitrogen application rate to be dynamically adjusted according to the real-time status of the rice, realizing dynamic and responsive fertilization decisions. Compared with the traditional single quantitative fertilization method, it is more flexible and accurate.

[0163] The variable fertilization device for rice provided by the present invention is described below. The variable fertilization device for rice described below can be referred to in correspondence with the variable fertilization method for rice described above.

[0164] Figure 5 This is a schematic diagram of the structure of the variable fertilization device for rice provided by the present invention, as shown below. Figure 5 As shown, the device includes a mask generation module 510, a feature fusion module 520, a grade prediction module 530, a grade standardization module 540, a factor determination module 550, and a nitrogen application rate adjustment module 560; wherein: The mask generation module 510 is used to remove non-rice plant interference in multispectral remote sensing images and generate rice plant canopy masks. The feature fusion module 520 is used to extract multidimensional features of the rice canopy based on the rice canopy mask and fuse them to obtain a fused growth feature vector. The grade prediction module 530 is used to grade the rice growth according to the fused growth feature vector through the growth grade prediction model to obtain the predicted growth grade. The grade standardization module 540 is used to standardize the predicted growth grade to obtain a standardized growth grade. The factor determination module 550 is used to determine the nitrogen fertilizer adjustment factor based on the standardized growth level. The nitrogen application rate adjustment module 560 is used to adjust the standard nitrogen application rate according to the nitrogen fertilizer adjustment factor to obtain the recommended nitrogen application rate.

[0165] The variable fertilization device for rice provided in this invention first removes non-rice plant interference from multispectral remote sensing images to generate a pure rice canopy mask, avoiding misjudgments caused by background environmental interference. Then, based on the rice canopy mask, multidimensional features of the rice canopy are extracted and fused to obtain a fused growth feature vector. This fusion of multidimensional features provides a more comprehensive reflection of the true growth status of the rice canopy. Next, a growth level prediction model is used to classify the rice growth according to the fused growth feature vector, obtaining a predicted growth level. This predicted growth level is then standardized to obtain a standardized growth level. Through the classification prediction and standardization processing of the growth level prediction model, the complex fused growth feature vector can be converted into a unified standardized growth level, thus ensuring absolute comparability of growth levels across regions and time periods. Finally, based on standardized growth levels, a nitrogen fertilizer adjustment factor is determined. Then, according to this factor, the standard nitrogen application rate is adjusted to obtain the recommended nitrogen application rate. Through this method, the nitrogen application rate can be dynamically adjusted according to the real-time status of the rice, achieving accurate recommendation of nitrogen application during rice cultivation. In summary, this invention, through the aforementioned interconnected technical steps, solves a series of technical problems, forming an accurate, robust, systematic, dynamic, and agronomically integrated variable fertilization method for rice. This method can accurately recommend nitrogen application rates during rice cultivation, thereby achieving precision fertilization.

[0166] It should be noted that the rice variable fertilization device provided in this embodiment of the invention can realize all the method steps implemented in the above-mentioned rice variable fertilization method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0167] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a variable fertilization method for rice. This method includes: removing non-rice plant interference in a multispectral remote sensing image to generate a rice canopy mask; extracting multidimensional features of the rice canopy based on the rice canopy mask and fusing them to obtain a fused growth feature vector; classifying the rice growth according to the fused growth feature vector using a growth level prediction model to obtain a predicted growth level; standardizing the predicted growth level to obtain a standardized growth level; determining a nitrogen fertilizer adjustment factor based on the standardized growth level; and adjusting the standard nitrogen application rate according to the nitrogen fertilizer adjustment factor to obtain a recommended nitrogen application rate.

[0168] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0169] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the variable fertilization method for rice provided in the above embodiments. The method includes: removing non-rice plant interference in a multispectral remote sensing image to generate a rice canopy mask; extracting and fusing multidimensional features of the rice canopy based on the rice canopy mask to obtain a fused growth feature vector; classifying the rice growth according to the fused growth feature vector using a growth level prediction model to obtain a predicted growth level; standardizing the predicted growth level to obtain a standardized growth level; determining a nitrogen fertilizer adjustment factor based on the standardized growth level; and adjusting the standard nitrogen application rate according to the nitrogen fertilizer adjustment factor to obtain a recommended nitrogen application rate.

[0170] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0171] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A variable fertilization method for rice, characterized in that, include: Non-rice plant interference objects are removed from multispectral remote sensing images to generate rice canopy masks; Based on the rice canopy mask, multidimensional features of the rice canopy are extracted and fused to obtain a fused growth feature vector. The rice growth is classified according to the fused growth feature vector using the growth level prediction model to obtain the predicted growth level. The predicted growth level is standardized to obtain the standardized growth level; Based on the standardized growth level, the nitrogen fertilizer adjustment factor is determined; Based on the nitrogen fertilizer adjustment factor, the standard nitrogen application rate is adjusted to obtain the recommended nitrogen application rate; The standardization process for the predicted growth level to obtain a standardized growth level includes: A predicted growth level map is generated based on the predicted growth level; Calculate the current mean and current standard deviation of the predicted growth level map; Based on the current mean, the current standard deviation, the preset standardized mean, and the preset standardized standard deviation, the standardized growth level corresponding to the predicted growth level of each pixel is calculated using the following formula: ； in, This represents the normalized growth level of the i-th pixel. This represents the predicted growth level of the i-th pixel. This represents the current mean. Indicates the current standard deviation. This indicates the preset standardized mean. This indicates the pre-defined standardized standard deviation.

2. The variable fertilization method for rice according to claim 1, characterized in that, The process of removing non-rice plant interference from multispectral remote sensing images to generate a rice canopy mask includes: Based on the first reflectance of each pixel in the multispectral remote sensing image on multiple first preset spectral bands, the Normalized Differential Water Index (NDWI), the Supergreen Index (ExG), and the Algae Index (Algae Index) are calculated respectively. The water body area is determined based on the NDWI value and the preset NDWI threshold; the area with missing seedlings and broken rows is determined based on the ExG value and the preset ExG threshold; and the algal patch area is determined based on the Algae Index value, the preset Algae Index threshold, the reflectance of the green light band, and the preset green light band reflectance threshold. In the multispectral remote sensing image, the water body area, the area with missing seedlings and broken rows, and the algal patch area are masked to obtain a first intermediate image; A gray-level co-occurrence matrix is ​​calculated on the first intermediate image to extract the first texture features, and the first texture features are fused with the spectral features of the first intermediate image to generate a second intermediate image; The second intermediate image is input into the pixel segmentation model to perform pixel-level segmentation of the rice canopy, and the initial rice canopy mask output by the pixel segmentation model is obtained. The initial rice canopy mask was optimized using morphological operations to obtain the rice canopy mask.

3. The variable fertilization method for rice according to claim 1, characterized in that, The process involves extracting multidimensional features of the rice canopy based on the rice canopy mask and fusing them to obtain a fused growth feature vector, including: Based on the rice canopy mask, multidimensional features of the rice canopy are extracted; wherein, the feature extraction dimensions include at least two of the following: spectral dimension, texture dimension, structural dimension, and temporal dimension. The multidimensional features are combined to generate an initial feature vector; Principal component analysis is used to reduce the dimensionality of the initial eigenvectors to obtain principal component eigenvectors. Based on the principal component feature vectors, the importance scores of each principal component feature are evaluated. The fused growth feature vector is determined based on the importance score.

4. The variable fertilization method for rice according to claim 3, characterized in that, The extraction of multidimensional features of the rice canopy based on the rice canopy mask includes: Based on the rice canopy mask, the second reflectance of the pixels of the rice canopy in the multispectral remote sensing image is obtained in multiple second preset spectral bands. Based on the second reflectance, a vegetation index is calculated, the vegetation index including at least one of the Normalized Difference Vegetation Index (NDVI), the Normalized Green Difference Vegetation Index (GNDVI), and the Normalized Red Edge Index (NDRE); and / or, Based on the rice canopy mask, calculate the supergreen index ExG and the visible light band atmospheric impedance index VARI of the rice canopy in the RGB image; and / or, Based on the rice canopy mask, the RGB image is converted to grayscale, and the grayscale co-occurrence matrix is ​​calculated on the converted grayscale image to obtain the second texture feature; and / or, The RGB image is processed by Structure for Motion Recovery (SFM) to generate a Digital Surface Model (DSM). Subtracting the preset digital elevation model (DTM) from the DSM yields the canopy height model (CHM), and the rice canopy height is obtained based on the CHM and the rice canopy mask; and / or, Obtain the cumulative effective accumulated temperature, and determine the reproductive period correction function based on the cumulative effective accumulated temperature; The growth index is calculated based on the NDVI, historical NDVI, and the reproductive period correction function; The multidimensional features include multiple features such as NDVI, GNDVI, NDRE, ExG, VARI, the second texture feature, CHM, and the growth index.

5. The variable fertilization method for rice according to any one of claims 1 to 4, characterized in that, The growth level prediction model was trained in the following manner: Physiological indicators of multiple rice samples were obtained, and a comprehensive growth index of each rice sample was constructed based on the physiological indicators. Calculate the population mean and population standard deviation based on the comprehensive growth index; Based on the overall mean and overall standard deviation, the comprehensive growth index of each rice sample is mapped to the sample growth level; Obtain the sample fusion growth feature vector corresponding to each rice sample, and construct a training sample set based on the sample fusion growth feature vector and the sample growth level; The preset growth level prediction model is trained using the training sample set to obtain the growth level prediction model.

6. The variable fertilization method for rice according to claim 5, characterized in that, The growth level prediction model includes multiple components. The step of using the growth level prediction model to classify rice growth based on the fused growth feature vector to obtain the predicted growth level includes: Multiple growth level prediction models are used to predict growth level based on the fused growth feature vector, resulting in multiple classification probability vectors. The multiple classification probability vectors are weighted and summed to obtain the fusion probability vector; The growth level corresponding to the maximum probability value in the fusion probability vector is determined as the predicted growth level.

7. The variable fertilization method for rice according to any one of claims 1 to 4, characterized in that, The determination of nitrogen fertilizer adjustment factors based on the standardized growth level includes: The standard canopy index is determined based on the standardized growth level. Obtain the measured canopy index, and calculate the relative growth index based on the measured canopy index and the standard canopy index; The nitrogen fertilizer adjustment factor is calculated based on the relative growth index and the preset adjustment function.

8. A variable-rate fertilization device for rice, characterized in that, include: The mask generation module is used to remove non-rice plant interference in multispectral remote sensing images and generate rice canopy masks. The feature fusion module is used to extract multidimensional features of the rice canopy based on the rice canopy mask and fuse them to obtain a fused growth feature vector. The grade prediction module is used to grade the rice growth based on the fused growth feature vector using the growth grade prediction model, and obtain the predicted growth grade. The grade standardization module is used to standardize the predicted growth grade to obtain a standardized growth grade; The factor determination module is used to determine the nitrogen fertilizer adjustment factor based on the standardized growth level. The nitrogen application rate adjustment module is used to adjust the standard nitrogen application rate according to the nitrogen fertilizer adjustment factor to obtain the recommended nitrogen application rate; Specifically, the level standardization module is used for: A predicted growth level map is generated based on the predicted growth level; Calculate the current mean and current standard deviation of the predicted growth level map; Based on the current mean, the current standard deviation, the preset standardized mean, and the preset standardized standard deviation, the standardized growth level corresponding to the predicted growth level of each pixel is calculated using the following formula: ； in, This represents the normalized growth level of the i-th pixel. This represents the predicted growth level of the i-th pixel. This represents the current mean. Indicates the current standard deviation. This indicates the preset standardized mean. This indicates the pre-defined standardized standard deviation.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the variable fertilization method for rice as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the variable fertilization method for rice as described in any one of claims 1 to 7.

Images