Wheat SPAD retrieval and topdressing decision method based on multispectral remote sensing

Abstract

The application provides a wheat SPAD inversion and topdressing decision method based on multispectral remote sensing, combines multispectral remote sensing with machine learning, realizes accurate inversion of the SPAD value of wheat at the jointing stage, and effectively overcomes the problems of time-consuming and low efficiency of the traditional measurement method and insufficient spatial resolution of data. Through multi-model integration and arithmetic average optimization, the inversion accuracy and stability are improved, and the data reliability is ensured. Based on the SPAD value obtained by inversion and the preset threshold, a plurality of models are called to calculate and take the average value, and the topdressing guiding amount is obtained, so that the fertilization is carried out according to the need, and the resource waste and environmental pollution caused by the 'one size fits all' are avoided. Further combined with the spatial unit identifier, the visual prescription map and the partition list are generated, so that the topdressing scheme has the executability and the visual effect, the level and the production efficiency of the wheat accurate management are significantly improved, and the quality and the efficiency are improved, and the green agricultural development is promoted.

Description

Technical Field

[0001] This invention relates to the field of wheat planting decision-making technology, and in particular to a method for wheat SPAD inversion and topdressing decision-making based on multispectral remote sensing. Background Technology

[0002] Currently, wheat is a major grain crop in my country. The jointing stage is a critical growth period, and nitrogen supply directly determines yield. Accurately grasping nitrogen nutrition at this time and making topdressing decisions is the core of high-quality and high-yield production. SPAD values ​​can reflect the chlorophyll content and nitrogen status of wheat. Currently, it mainly relies on field measurements with SPAD instruments, which has the problems of being time-consuming and labor-intensive, having limited coverage, and being unable to accurately reflect the spatial differences in nitrogen in the field.

[0003] Multispectral remote sensing technology has made it possible to retrieve SPAD values ​​over large areas, but existing technologies suffer from non-standard preprocessing and simplistic inversion models, resulting in low accuracy and poor stability. Topdressing decisions often employ a "one-size-fits-all" approach, failing to consider field heterogeneity, leading to nitrogen waste and environmental pollution. Some schemes combining SPAD values ​​suffer from large calculation biases due to single models and a lack of complete prescription results, making it difficult to guide precision operations. Furthermore, the lack of systematic integration between SPAD value retrieval and topdressing calculations results in cumbersome processes that cannot meet the needs of large-scale planting. In summary, current methods for obtaining SPAD values ​​and making topdressing decisions during the wheat jointing stage have many shortcomings. Developing accurate and efficient retrieval and topdressing decision-making methods has become an urgent problem to be solved in agricultural production. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for wheat SPAD inversion and topdressing decision-making based on multispectral remote sensing.

[0005] To achieve the above objectives, the technical solution of this invention is as follows: In a first aspect, the present invention provides a method for wheat SPAD inversion and topdressing decision-making based on multispectral remote sensing, the method comprising: Acquire multispectral raw image data of target fields during the wheat jointing stage, and preprocess the multispectral raw image data to generate four-band reflectance orthophoto image data; Based on four-band reflectance orthophoto image data, the reflectance statistical characteristic values ​​of each band in each spatial statistical unit in the target field are calculated, and a multispectral feature dataset containing the correspondence between spatial unit identifiers and the reflectance statistical characteristic values ​​of each band is constructed. The feature vectors of each spatial statistical unit in the multispectral feature dataset are input in parallel into multiple pre-trained random forest regression models to obtain the SPAD prediction values ​​for the jointing stage output by each random forest regression model. The arithmetic mean ensemble method is used to integrate the multiple SPAD prediction values ​​for the jointing stage to obtain the final inverted SPAD value for the corresponding spatial statistical unit. The final inverted SPAD value of each spatial statistical unit and the preset target flowering period flag leaf SPAD threshold are used as input parameters. The first, second and third recommended topdressing amount calculation models are called respectively to calculate the fertilization amount, and the recommended topdressing amount output by each topdressing amount calculation model is obtained. The arithmetic mean of multiple recommended topdressing amounts is calculated as the final recommended topdressing guidance amount for the corresponding spatial statistical unit. Based on the spatial unit identifiers of each spatial statistical unit and the corresponding final recommended topdressing guidance amount, a precise topdressing prescription result is generated, including a list of topdressing amounts for each zone and a visual prescription map.

[0006] In some embodiments, the multispectral raw image data is preprocessed, including image registration, radiometric correction, orthorectification and stitching, so that the generated four-band reflectance orthophoto image data has a uniform reflectance scale and spatial reference.

[0007] In some embodiments, the multiple pre-trained random forest regression models include a first random forest regression model, a second random forest regression model, a third random forest regression model, and a fourth random forest regression model; the first random forest regression model uses near-infrared band reflectance as the primary splitting variable of the root node; the second random forest regression model uses near-infrared band reflectance as the primary splitting variable of the root node, and its splitting threshold is different from that of the first random forest regression model; the third random forest regression model uses near-infrared band reflectance and red-edge band reflectance as co-judgment variables; and the fourth random forest regression model uses red light band reflectance as the splitting variable of the root node.

[0008] In some embodiments, each random forest regression model consists of multiple regression decision trees. Each decision tree maps the input feature vector to a leaf node and outputs a basic SPAD prediction value through a binary splitting rule based on the statistical feature value of reflectance. The final SPAD prediction value output by the random forest regression model is the average of the prediction values ​​of all decision trees in the corresponding model.

[0009] In some embodiments, the calculation model for the first recommended topdressing amount is a linear function model, expressed as: In the formula, The first recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0010] In some embodiments, the calculation model for the second recommended topdressing amount is a function model containing quadratic terms, expressed as: In the formula, The second recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0011] In some embodiments, the third recommended topdressing amount calculation model is a power function model, and the expression is: In the formula, Let i be the third recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0012] In some embodiments, the method further includes a closed-loop verification step: recording the SPAD value at the jointing stage, the SPAD value at the flag leaf stage, and the final yield under different topdressing treatments through field trials; verifying whether the SPAD value at the flag leaf stage is stably within a preset suitable range after topdressing regulation, and verifying whether there is a significant positive correlation between the SPAD value at the flag leaf stage and wheat yield.

[0013] This invention provides a method for wheat SPAD inversion and topdressing decision-making based on multispectral remote sensing. Combining multispectral remote sensing with machine learning, it achieves accurate inversion of SPAD values ​​at the wheat jointing stage, effectively overcoming the problems of time-consuming inefficiency and insufficient spatial resolution of traditional measurement methods. Through multi-model integration and arithmetic mean optimization, the inversion accuracy and stability are improved, ensuring data reliability. Based on the inverted SPAD values ​​and preset thresholds, multiple models are invoked for calculation, and their average is taken to obtain the guiding amount of topdressing, thereby enabling zoning-based fertilization and avoiding resource waste and environmental pollution caused by a "one-size-fits-all" approach. Furthermore, by combining spatial unit identification, a visual prescription map and zoning list are generated, making the topdressing plan executable and visually appealing, significantly improving the level of precision management and production efficiency of wheat, contributing to quality and efficiency improvement, and promoting the development of green agriculture. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the wheat SPAD inversion and topdressing decision-making method based on multispectral remote sensing provided in this embodiment of the invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] In the following description, references to "some embodiments" refer to a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the invention have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of the invention pertain. The terminology used in the embodiments of the invention is for the purpose of describing the embodiments of the invention only and is not intended to limit the invention.

[0017] The following describes an exemplary application of the wheat SPAD inversion and topdressing decision-making device based on multispectral remote sensing according to embodiments of the present invention. This device can be implemented as a terminal or a server. In one implementation, the device can be implemented as a laptop, tablet, desktop computer, mobile device, or other types of terminal. In another implementation, it can also be implemented as a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present invention. The following will illustrate an exemplary application of the wheat SPAD inversion and topdressing decision-making device based on multispectral remote sensing as a server.

[0018] This invention provides a method for wheat SPAD inversion and topdressing decision-making based on multispectral remote sensing. See [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating a wheat SPAD inversion and topdressing decision-making method based on multispectral remote sensing, provided by an embodiment of the present invention. Figure 1 The steps shown are explained.

[0019] S110: Acquire multispectral raw image data of target fields during the wheat jointing stage, and preprocess the multispectral raw image data to generate four-band reflectance orthophoto image data.

[0020] It should be noted that the jointing stage of wheat is one of the key growth and development periods for wheat. Specifically, it refers to the growth stage when the basal internodes of the wheat plant begin to elongate, the stem gradually changes from flat to round, the first internode reaches a length of 1-2 cm, and the young spikes begin to differentiate. This period is when wheat has a high demand for fertilizer, directly affecting tillering, spike formation, the number of grains per spike, and the later yield. This is also the application period of this invention. For example, the jointing stage of winter wheat in North China is usually from late March to early April, while the jointing stage of spring wheat is usually from early to mid-May.

[0021] Here, raw multispectral image data refers to raw image data acquired by multispectral sensors, such as multispectral cameras mounted on drones or satellite multispectral sensors, containing multiple specific bands of the wheat canopy, i.e., the four bands in this invention. This data is acquired without any preprocessing such as radiometric or geometric correction and includes raw information such as grayscale values, band information, and spatial coordinates. For example, raw multispectral image data is obtained by a drone equipped with a multispectral camera, capturing raw images of a target field during the wheat jointing stage in the blue, green, red, and near-infrared bands.

[0022] Here, preprocessing in this invention refers to a series of corrections, adjustments, and optimizations performed on the original multispectral image data to remove interference factors and ensure that the image data accurately reflects the reflectance characteristics of the wheat canopy, providing reliable data for subsequent reflectance calculations and SPAD inversion. Preprocessing mainly includes radiometric correction, which eliminates the effects of sensor errors and differences in lighting conditions; geometric correction, which corrects spatial distortion of the image to ensure that image coordinates correspond to actual field coordinates; image stitching, which stitches multiple images together to form a complete field image if a single image cannot cover the entire target field; and noise reduction, which removes noise points from the image and improves image clarity.

[0023] Here, four-band reflectance orthophoto image data refers to orthophotos containing four specific bands obtained after preprocessing multispectral raw image data, where the grayscale value of each pixel in the image has been converted into the reflectance value of the corresponding band. An orthophoto image, after geometric correction, is an image where the coordinates of any point on the image correspond one-to-one with the actual ground coordinates, without spatial distortion; reflectance refers to the ratio of the electromagnetic wave energy reflected by an object to the incident electromagnetic wave energy, and is one of the parameters reflecting the spectral characteristics of wheat canopy, with a value range of 0-1.

[0024] S120, based on four-band reflectance orthophoto image data, calculates the reflectance statistical characteristic value of each band in each spatial statistical unit in the target field, and constructs a multispectral feature dataset containing the correspondence between spatial unit identifiers and the reflectance statistical characteristic values ​​of each band.

[0025] Here, a spatial statistical unit refers to the smallest unit with independent spatial coordinates and boundaries into which the target field is divided according to preset rules. It is the basic spatial unit for SPAD inversion and topdressing calculation in this invention. The wheat growth status within each spatial statistical unit, such as chlorophyll content and nitrogen level, is considered relatively uniform, and its reflectance statistical characteristic value, SPAD inversion value, and topdressing amount can be calculated independently. For example, the target field is divided into a 5m×5m square grid, and each grid is a spatial statistical unit, with each unit corresponding to a unique spatial unit identifier.

[0026] Here, the reflectance statistical feature value refers to the feature parameter obtained after statistically analyzing the reflectance values ​​of each band in the four-band reflectance orthophoto image data within each spatial statistical unit. It is used to characterize the spectral reflectance characteristics of the wheat canopy within the unit and is the core input feature for constructing a multispectral feature dataset and training a random forest regression model.

[0027] In practical applications, the statistical characteristic value of reflectance can be the average value, i.e., the arithmetic mean of the reflectance of all pixels in a certain band within the unit; it can also be the maximum value, i.e., the maximum reflectance of a certain band within the unit; it can also be the minimum value, i.e., the minimum reflectance of a certain band within the unit; it can also be the standard deviation, i.e., reflecting the degree of dispersion of the reflectance of a certain band within the unit; of course, it can also be the coefficient of variation, i.e., the ratio of the standard deviation to the average value, reflecting the relative dispersion of reflectance, etc. The statistical characteristic value of reflectance can be any value that can reflect the reflectance of a certain band within the unit. This embodiment of the invention does not impose specific limitations on this.

[0028] Here, the multispectral feature dataset is a structured data table used to store the feature information of all spatial statistical units in the target field. Its representation is based on the spatial unit identifier and the statistical feature value of reflectance for each band. Each row in the table corresponds to a spatial statistical unit, and each column corresponds to a feature (including spatial unit identifier, average blue band reflectance, standard deviation of green band reflectance, maximum red band reflectance, etc.). This dataset serves as the foundational data carrier for subsequently inputting the feature vectors into the random forest regression model for SPAD inversion.

[0029] S130, the feature vectors of each spatial statistical unit in the multispectral feature dataset are input in parallel into multiple pre-trained random forest regression models to obtain the SPAD prediction values ​​for the jointing stage output by each random forest regression model; the arithmetic mean ensemble method is used to integrate the multiple SPAD prediction values ​​for the jointing stage to obtain the final inverted SPAD value for the corresponding spatial statistical unit.

[0030] Here, the feature vector refers to the vector composed of all reflectance statistical feature values ​​corresponding to a single spatial statistical unit, extracted from the multispectral feature dataset. It is the input parameter of the random forest regression model. For example, the feature vector of a certain spatial statistical unit is [spatial unit identifier 001, average reflectance in blue band 0.12, average reflectance in green band 0.25, average reflectance in red band 0.18, average reflectance in near-infrared band 0.65, standard deviation of reflectance in red band 0.03], where all values ​​except the spatial unit identifier are reflectance statistical feature values.

[0031] Here, parallel input refers to simultaneously inputting the feature vectors of multiple spatial statistical units into multiple pre-trained random forest regression models for computation. This method can significantly improve the computational efficiency of SPAD predictions, is suitable for rapid inversion of large-area target fields, and ensures the timeliness of topdressing decisions.

[0032] Here, the pre-trained random forest regression model is an ensemble learning regression algorithm model composed of multiple decision trees. The final regression prediction value is obtained by integrating the prediction results of multiple decision trees. In this invention, pre-training refers to training and debugging the model with a large number of training samples before using it for SPAD value inversion, determining the optimal parameters of the model, such as the number and depth of decision trees, so that the model has a stable and accurate SPAD value prediction capability.

[0033] Here, the SPAD prediction value at the jointing stage refers to the estimated SPAD value of wheat at the jointing stage within a pre-trained random forest regression model after inputting the feature vector of a single spatial statistical unit into that model.

[0034] Here, the final inverted SPAD value refers to the SPAD value obtained by integrating the predicted SPAD values ​​of the jointing stage output by multiple random forest regression models through the arithmetic mean ensemble method, which can truly reflect the relative chlorophyll content of wheat at the jointing stage within the corresponding spatial statistical unit.

[0035] S140: The final inverted SPAD value of each spatial statistical unit and the preset target flowering period flag leaf SPAD threshold are used as input parameters. The first, second and third recommended topdressing amount calculation models are called respectively to calculate the fertilization amount, and the recommended topdressing amount output by each topdressing amount calculation model is obtained. The arithmetic mean of multiple recommended topdressing amounts is calculated as the final recommended topdressing guidance amount for the corresponding spatial statistical unit.

[0036] Here, the preset target SPAD threshold for flag leaf at the flowering stage refers to the minimum or suitable SPAD value that the flag leaf should reach during the wheat flowering stage, pre-set based on factors such as wheat variety characteristics, target yield, and soil fertility. It serves as a reference standard for judging wheat nitrogen requirements and calculating topdressing amounts. The flowering stage refers to the wheat's growth period from heading to flowering, and the flag leaf refers to the topmost leaf of the wheat plant, which is the main organ for photosynthesis. If the final SPAD value at the jointing stage is lower than this threshold, it indicates insufficient nitrogen, requiring supplementation through topdressing; if it is higher than this threshold, it indicates sufficient nitrogen, and topdressing can be reduced or omitted. For example, for the winter wheat variety Jimai 44, with a target yield of 600 kg / mu, the preset target SPAD threshold for flag leaf at the flowering stage is 50, meaning that if the SPAD value at the jointing stage is lower than 50, topdressing is required.

[0037] Here, the first, second, and third recommended topdressing amount calculation models refer to pre-constructed mathematical models used to calculate the recommended topdressing amount for wheat based on the final inverted SPAD value at the wheat jointing stage and the target flag leaf SPAD threshold at the flowering stage. The three models employ different calculation logics, such as those based on the nitrogen balance principle, SPAD value difference regression, and field trial fitting, to improve the reliability of the topdressing amount calculation. The input parameters for all three models are the final inverted SPAD value and the target flag leaf SPAD threshold at the flowering stage, and the output parameter is the recommended topdressing amount.

[0038] Here, the recommended topdressing amount refers to a single topdressing amount calculation model, namely the first, second, or third recommended topdressing amount calculation model. Based on the input final inverted SPAD value and the target flowering stage flag leaf SPAD threshold, it calculates the output amount of nitrogen to be applied during the jointing stage of wheat within the corresponding spatial statistical unit. Each spatial statistical unit will correspond to three recommended topdressing amounts, which serve as the basis for subsequent calculations of the final recommended topdressing guidance amount.

[0039] Here, the final recommended topdressing guidance amount refers to the average of the recommended topdressing amounts output by the three recommended topdressing amount calculation models, calculated using the arithmetic mean method. This value serves as the core guidance for actual topdressing during the wheat jointing stage within the corresponding spatial statistical unit. This value integrates the results of three different calculation logic models, avoiding the calculation bias of a single model and ensuring the accuracy and rationality of the topdressing amount.

[0040] S150 generates precise topdressing prescription results, including a list of topdressing amounts for each zone and a visual prescription map, based on the spatial unit identifiers of each spatial statistical unit and the corresponding final recommended topdressing guidance amount.

[0041] Here, the zonal topdressing amount list refers to a structured list compiled based on the spatial unit identifiers of all spatial statistical units and their corresponding final recommended topdressing guidance amounts. This list clarifies the topdressing amount for each spatial statistical unit within the target field. The zonal topdressing amount list typically includes spatial unit identifiers, unit locations such as coordinates, field area descriptions, final recommended topdressing guidance amounts, and suggested topdressing methods, facilitating precise topdressing by growers according to the list.

[0042] Here, a visual prescription map refers to a graphical representation of the final recommended topdressing guidance amount for each spatial statistical unit on a spatial map of the target field, intuitively showing the differences in topdressing amounts in different areas of the field. The map typically uses different colors to represent different ranges of topdressing amounts; for example, light colors indicate low topdressing amounts, and dark colors indicate high topdressing amounts. It also includes spatial unit identifiers, topdressing amount values, and legends. Growers can use this map to quickly determine the topdressing needs of different areas of the field and combine it with agricultural machinery to achieve precise topdressing operations.

[0043] Here, the precise topdressing prescription result is the final output of this invention, consisting of a list of topdressing amounts for different zones and a visual prescription chart.

[0044] This invention provides a method for wheat SPAD inversion and topdressing decision-making based on multispectral remote sensing. Combining multispectral remote sensing with machine learning, it achieves accurate inversion of SPAD values ​​at the wheat jointing stage, effectively overcoming the problems of time-consuming inefficiency and insufficient spatial resolution of traditional measurement methods. Through multi-model integration and arithmetic mean optimization, the inversion accuracy and stability are improved, ensuring data reliability. Based on the inverted SPAD values ​​and preset thresholds, multiple models are invoked for calculation, and their average is taken to obtain the guiding amount of topdressing, thereby enabling zoning-based fertilization and avoiding resource waste and environmental pollution caused by a "one-size-fits-all" approach. Furthermore, by combining spatial unit identification, a visual prescription map and zoning list are generated, making the topdressing plan executable and visually appealing, significantly improving the level of precision management and production efficiency of wheat, contributing to quality and efficiency improvement, and promoting the development of green agriculture.

[0045] In some embodiments, the multispectral raw image data is preprocessed, including image registration, radiometric correction, orthorectification and stitching, so that the generated four-band reflectance orthophoto image data has a uniform reflectance scale and spatial reference.

[0046] In some embodiments, the multiple pre-trained random forest regression models include a first random forest regression model, a second random forest regression model, a third random forest regression model, and a fourth random forest regression model; the first random forest regression model uses near-infrared band reflectance as the primary splitting variable of the root node; the second random forest regression model uses near-infrared band reflectance as the primary splitting variable of the root node, and its splitting threshold is different from that of the first random forest regression model; the third random forest regression model uses near-infrared band reflectance and red-edge band reflectance as co-judgment variables; and the fourth random forest regression model uses red light band reflectance as the splitting variable of the root node.

[0047] In some embodiments, each random forest regression model consists of multiple regression decision trees. Each decision tree maps the input feature vector to a leaf node and outputs a basic SPAD prediction value through a binary splitting rule based on the statistical feature value of reflectance. The final SPAD prediction value output by the random forest regression model is the average of the prediction values ​​of all decision trees in the corresponding model.

[0048] In this invention, a random forest regression model is constructed by multiple regression decision trees. The feature vector is mapped by the binary splitting rule of the reflectance statistical feature value. The arithmetic mean of the basic SPAD prediction values ​​output by all decision tree leaf nodes in the model is used as the final SPAD prediction value. This can improve the accuracy and stability of SPAD prediction, enhance the model's anti-overfitting ability, and reduce the prediction bias of a single decision tree.

[0049] In some embodiments, the calculation model for the first recommended topdressing amount is a linear function model, expressed as: In the formula, The first recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0050] In some embodiments, the calculation model for the second recommended topdressing amount is a function model containing quadratic terms, expressed as: In the formula, The second recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0051] In some embodiments, the third recommended topdressing amount calculation model is a power function model, and the expression is: In the formula, Let i be the third recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0052] In some embodiments, the method further includes a closed-loop verification step: recording the SPAD value at the jointing stage, the SPAD value at the flag leaf stage, and the final yield under different topdressing treatments through field trials; verifying whether the SPAD value at the flag leaf stage is stably within a preset suitable range after topdressing regulation, and verifying whether there is a significant positive correlation between the SPAD value at the flag leaf stage and wheat yield.

[0053] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0054] During the wheat jointing stage, a DJI M3M multispectral drone was first used to conduct aerial data collection over the experimental field, obtaining raw multispectral image data covering the target field. Subsequently, the raw multispectral image data was imported into Pix4D for image preprocessing, including image registration, radiometric correction, and orthorectification stitching, generating a four-band reflectance orthorectified image with a unified spatial reference. Based on this, using the matrix / cell vector boundaries of the target field as spatial statistical units, statistical feature parameters were extracted from the four-band reflectance image for each numbered unit, forming ID-green-red-red. The multispectral feature dataset corresponding to edge-NIR is then used as input. Four random forest regression models are called to invert the SPAD value at the jointing stage, obtaining four sets of SPAD prediction results. The mean of these predictions is used as the final inverted SPAD value at the jointing stage, thereby improving the stability and noise resistance of the inversion results. Finally, the inverted SPAD value at the jointing stage is substituted into three topdressing amount calculation models to calculate the recommended topdressing amount. The mean of the outputs of the three models is taken as the final topdressing guidance amount, and the prescription results for variable-oriented fertilization operations are output (including a topdressing amount list by ID and a visual prescription map). This realizes an integrated precision topdressing technology process that enables UAVs to quickly acquire multispectral information, invert the SPAD at the jointing stage, intelligently calculate the topdressing amount, and output the prescription.

[0055] The raw multispectral images acquired by the UAV are first imported into Pix4D for image preprocessing, mainly including image registration, orthorectification, radiometric correction, and mosaicking, to generate a four-band reflectance orthophoto image with a unified spatial reference and comparable reflectance scale. Subsequently, matrix / plot vector surface features used for field division are used as spatial statistical units, and multiple surface features with the same ID are geometrically fused to ensure one ID corresponds to one spatial region. Based on this, reflectance statistical features (preferably the regional mean) are calculated for the region corresponding to each ID on the four-band reflectance image, thereby obtaining the feature vector of each spatial unit. ,in These represent the mean reflectance values ​​for the G, R, RE, and NIR bands within the region. The final output is a parameter data table characterizing the correspondence between spatial cell identifiers and the statistical characteristic values ​​of reflectance for each band, i.e., a multispectral feature dataset. This provides standardized input for subsequent random forest model inversion of SPAD values ​​at the jointing stage and calculation of topdressing prescriptions.

[0056] The above feature vectors The data is fed in parallel into four pre-trained random forest regression models, denoted as Model1, Model2, Model3, and Model4. Each model runs independently and outputs a SPAD prediction. All four models are based on the random forest algorithm framework, but differ in their training data, splitting rules, and primary decision variables, thus forming complementary predictive perspectives.

[0057] Common model structure: Each model consists of multiple regression decision trees. Each decision tree splits the input feature vector using a series of binary splitting rules based on a reflectivity threshold. This is mapped to a leaf node, which outputs a basic SPAD prediction. The final output of the random forest model is the average of all decision tree predictions within the model.

[0058] Model Differences and Characteristics: Model 1: Uses near-infrared (NIR) reflectance as the primary and core splitting variable. Its decision rule indicates that when NIR reflectance is ≤0.42 and >0.42, the sample enters a distinct prediction branch, and then further refines the judgment by combining thresholds for green light, red edge, and other bands, outputting the SPAD value. This model is extremely sensitive to changes in canopy structure and biomass. Model 2: Also uses near-infrared as the primary splitting variable, but the specific splitting threshold is different from Model 1, and the combination rule with red light and red edge is unique. For example, one decision path is: if NIR ≤0.425 and R ≤0.023, then output the SPAD value based on whether RE is greater than 0.211. Model 3: Its rules show a strong dependence on the near-infrared and red edge bands. In the example, the model effectively distinguishes the low and medium SPAD value range and further subdivides the high value area through the synergistic judgment of NIR and RE. Model 4: Employs a unique decision starting point, using red light reflectance as the splitting variable of the root node. This allows it to start with chlorophyll absorption characteristics and then combine information from the NIR, G, and RE bands to make comprehensive decisions, providing a predictive logic different from other models.

[0059] Integrated decision-making and final SPAD value generation: In order to integrate the advantages of each model and eliminate the local bias that may exist in a single model, this invention uses the arithmetic mean integration method to generate the final SPAD inversion value of each evaluation unit.

[0060] Fertilizer application rate calculation: SPAD values ​​are retrieved at the jointing stage of each matrix unit. Subsequently, the present invention is based on As input for topdressing decision-making, the target flowering period flag leaf SPAD threshold / target value is introduced. (Based on the optimal SPAD level corresponding to the flowering period determined in previous experiments), the recommended topdressing amount for this unit is output using three different topdressing amount calculation models (expressed as TD of pure nitrogen per acre). The three models are: The first recommended fertilizer application rate calculation model is a linear function model, with the following expression: ; In the formula, The first recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0061] The second recommended model for calculating the amount of topdressing fertilizer is a function model containing quadratic terms, expressed as follows: ; In the formula, The second recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0062] The third recommended model for calculating the amount of topdressing fertilizer is a power function model, expressed as follows: ; In the formula, Let i be the third recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

[0063] To reduce the risk of bias from a single model under different years, varieties, or field variations, this invention uses the ensemble of the means of three models as the final recommended amount of topdressing: ; The final output includes a list of topdressing guidance amounts and a prescription chart corresponding to each ID. This can be directly used for zonal topdressing or variable fertilization equipment operations, enabling intelligent and precise topdressing decisions based on SPAD diagnosis during the jointing stage, calculation of topdressing amounts, and prescription output.

[0064] Closed-loop achievement and yield increase verification: To verify the closed-loop effectiveness of this invention's SPAD diagnosis at the jointing stage - topdressing prescription - flag leaf SPAD achievement at flowering stage - yield increase, a field trial was conducted using topdressing as the control method. SPAD at the jointing stage, flag leaf SPAD at flowering stage, and the final yield were recorded. The verification approach was as follows: First, to examine whether the flag leaf SPAD at flowering stage was stably guided to the preset suitable range (i.e., achieving the target) after topdressing control; second, to examine whether there was a significant positive correlation between flag leaf SPAD at flowering stage and yield, thereby proving that the control logic targeting flag leaf SPAD at flowering stage can be converted into yield gains. In practice, regression analysis was performed on the SPAD and yield of each treatment / plot. If the regression relationship was significant and well-fitted, it indicated that improving and stabilizing the flag leaf SPAD during the flowering period could achieve yield gains. At the same time, combined with the distribution changes of the flag leaf SPAD during the flowering period under different topdressing amounts, it can be further proved that the topdressing prescription can avoid the deviation of the flag leaf SPAD during the flowering period caused by excessive or insufficient fertilization, thereby maximizing yield and improving fertilization efficiency while ensuring that the flag leaf SPAD during the flowering period meets the standard.

[0065] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

[0066] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0067] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not performed.

[0068] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for wheat SPAD inversion and topdressing decision-making based on multispectral remote sensing, characterized in that, The method includes: Acquire multispectral raw image data of target fields during the wheat jointing stage, and preprocess the multispectral raw image data to generate four-band reflectance orthophoto image data; Based on four-band reflectance orthophoto image data, the reflectance statistical characteristic values ​​of each band in each spatial statistical unit in the target field are calculated, and a multispectral feature dataset containing the correspondence between spatial unit identifiers and the reflectance statistical characteristic values ​​of each band is constructed. The feature vectors of each spatial statistical unit in the multispectral feature dataset are input in parallel into multiple pre-trained random forest regression models to obtain the SPAD prediction values ​​for the jointing stage output by each random forest regression model. The arithmetic mean ensemble method is used to integrate the multiple SPAD prediction values ​​for the jointing stage to obtain the final inverted SPAD value for the corresponding spatial statistical unit. The final inverted SPAD value of each spatial statistical unit and the preset target flowering period flag leaf SPAD threshold are used as input parameters. The first, second and third recommended topdressing amount calculation models are called respectively to calculate the fertilization amount, and the recommended topdressing amount output by each topdressing amount calculation model is obtained. The arithmetic mean of multiple recommended topdressing amounts is calculated as the final recommended topdressing guidance amount for the corresponding spatial statistical unit. Based on the spatial unit identifiers of each spatial statistical unit and the corresponding final recommended topdressing guidance amount, a precise topdressing prescription result is generated, including a list of topdressing amounts for each zone and a visual prescription map.

2. The method according to claim 1, characterized in that, The multispectral raw image data is preprocessed, including image registration, radiometric correction, orthorectification and mosaicking, so that the generated four-band reflectance orthophoto image data has a unified reflectance scale and spatial reference.

3. The method according to claim 1, characterized in that, Multiple pre-trained random forest regression models include the first random forest regression model, the second random forest regression model, the third random forest regression model, and the fourth random forest regression model; The first random forest regression model uses near-infrared reflectance as the primary splitting variable for the root node. The second random forest regression model uses near-infrared reflectance as the primary splitting variable of the root node, and its splitting threshold is different from that of the first random forest regression model. The third random forest regression model uses near-infrared band reflectance and red-edge band reflectance as co-judgment variables; The fourth random forest regression model uses red light band reflectance as the splitting variable of the root node.

4. The method according to claim 3, characterized in that, Each random forest regression model consists of multiple regression decision trees. Each decision tree maps the input feature vector to a leaf node and outputs a basic SPAD prediction value through a binary splitting rule based on the statistical feature value of reflectance. The final SPAD prediction value output by the random forest regression model is the average of the prediction values ​​of all decision trees in the corresponding model.

5. The method according to claim 1, characterized in that, The first recommended fertilizer application rate calculation model is a linear function model, with the following expression: ； In the formula, The first recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

6. The method according to claim 1, characterized in that, The second recommended model for calculating the amount of topdressing fertilizer is a function model containing quadratic terms, expressed as follows: ； In the formula, The second recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

7. The method according to claim 1, characterized in that, The third recommended model for calculating the amount of topdressing fertilizer is a power function model, expressed as follows: ； In the formula, Let i be the third recommended topdressing amount for the i-th spatial statistical unit. Let be the final inverted SPAD value of the i-th spatial statistical unit. The preset target flowering period flag leaf SPAD threshold.

8. The method according to claim 1, characterized in that, The method also includes a closed-loop verification step: Field trials were conducted to record the SPAD values ​​at the jointing stage, the flag leaf SPAD value at the flowering stage, and the final yield under different topdressing treatments. The results were used to verify whether the flag leaf SPAD value at the flowering stage remained stable within the preset suitable range after topdressing regulation, and to verify whether there was a significant positive correlation between the flag leaf SPAD value at the flowering stage and wheat yield.

Images