A deep learning-based rice deficiency type determination and variable fertilization method
By using a deep learning-based method for identifying rice nutrient deficiency types and applying variable fertilizers, and by employing aerial survey drones and deep learning models, the types of nutrient deficiencies in rice can be quickly identified and fertilizer prescription maps can be generated. This solves the problem of inefficient nutrient deficiency identification in existing technologies and improves rice growth efficiency and yield.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA AGRICULTURAL UNIVERSITY
- Filing Date
- 2024-05-20
- Publication Date
- 2026-06-12
AI Technical Summary
Existing methods for determining nutrient deficiency in rice and applying variable fertilizers are labor-intensive, resource-intensive, and inefficient. They cannot quickly and accurately determine the amount of nutrient deficiency, which leads to stunted rice growth and reduced yield.
Using a deep learning-based approach, a dataset of rice nutrient deficiency samples was established by capturing images of rice symptoms using aerial survey drones. A model for identifying and predicting rice nutrient deficiency types was constructed. Real-time growth images captured by drones were used to determine the nutrient deficiency and generate fertilization prescription maps. A controlled variable fertilization device was then used for quantitative fertilization.
It enables rapid and accurate nutrient deficiency assessment and variable fertilization, improving rice quality and yield while reducing manpower and material consumption.
Smart Images
Figure CN118521926B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of rice topdressing technology, and in particular to a method for determining the types of nutrient deficiency in rice and applying variable fertilizers based on deep learning. Background Technology
[0002] Throughout the rice cultivation process, factors such as soil fertility, climate conditions, and planting management can all lead to a lack of different types of fertilizers in rice, resulting in various nutrient deficiency symptoms. This can hinder rice growth and development, reduce its resistance to adverse conditions, and consequently reduce yield and quality. Therefore, rapidly identifying nutrient deficiency information in rice and promptly applying variable-rate topdressing as needed is crucial for addressing nutrient deficiencies, improving rice quality, and increasing yield.
[0003] Currently, there are two common methods for determining the types of fertilizer deficiencies and applying variable fertilization: soil nutrient testing and human observation. Soil nutrient testing involves sampling, mixing, drying, and testing the soil from which rice is grown to analyze the element content and thus determine the types of fertilizers the rice is currently lacking. Human observation, on the other hand, is a method primarily used by experienced rice growers to determine the current nutrient deficiencies of rice based on the symptoms of rice growth.
[0004] However, both of these methods are labor-intensive and inefficient, and they cannot directly determine the amount of fertilizer deficiency in rice, making it difficult to quickly and effectively address rice diseases. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a method for determining the types of nutrient deficiency in rice and applying variable fertilization based on deep learning. This method can quickly solve the problem of nutrient deficiency in rice during its growth process, thereby improving rice quality and yield.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] A deep learning-based method for determining the types of nutrient deficiency in rice and applying variable fertilization methods includes the following steps:
[0008] S1: Establish a rice fertilizer deficiency experimental field and plant rice varieties deficient in nitrogen, phosphorus, and potassium respectively;
[0009] S2: Use aerial survey drones to capture images of symptoms of nitrogen, phosphorus, and potassium deficiency in rice, and use these images to create a dataset of fertilizer deficiency samples.
[0010] S3: Perform feature engineering on the nutrient-deficient sample dataset;
[0011] S4: By constructing training sets, test sets, and validation sets, a model for determining and predicting the types of nutrient deficiency in rice is established.
[0012] S5: Establish normal rice planting experimental fields, divide the rice planting experimental fields into regions, mark each region and perform geographical coordinate transformation;
[0013] S6: Use aerial survey drones to capture real-time growth images of rice in normal rice planting experimental fields. Based on the rice nutrient deficiency type identification and prediction model, determine the real-time growth images of rice and obtain the location information of the area corresponding to the real-time growth image of rice. In this way, generate a fertilizer prescription map of the current area, thereby determining the fertilizer type and amount for the area.
[0014] S7: Based on the application rates of various fertilizers generated from the fertilizer prescription diagram, control the variable fertilizer application device to discharge fertilizer;
[0015] S8: After fertilization is completed, continue with the determination of fertilizer type and fertilization operation in the next area.
[0016] Preferably, in step S1, the step of establishing a rice nutrient deficiency experimental field is as follows:
[0017] S11: Select three square paddy fields of equal area with the same soil moisture content and soil fertility;
[0018] S12: Perform land preparation treatment on the three selected square paddy fields, including but not limited to weeding, tilling the land, and leveling the soil surface;
[0019] S13: Apply compound fertilizers that are "nitrogen-deficient", "phosphorus-deficient", and "potassium-deficient" to the three square paddy fields as base fertilizers for the early stage of rice planting;
[0020] S14: In subsequent planting, topdressing should be carried out according to the principle of "no nitrogen application to nitrogen-deficient plots", "no phosphorus application to phosphorus-deficient plots", and "no potassium application to potassium-deficient plots".
[0021] Preferably, in step S2, the step of establishing the nutrient deficiency sample dataset is as follows:
[0022] S21: Set the flight path of the aerial survey drone. The flight path should ensure that the aerial survey drone can capture images of rice growth in all the rice nutrient deficiency test fields.
[0023] S22: Before the aerial survey drone takes off, the GPS module's geographical location information needs to be updated, and the operation of the IMU needs to be checked to ensure the drone's stability during the shooting process; after takeoff, RTK real-time dynamic measurement technology is used to establish a correspondence between the rice growth photos taken by the drone in each area and the latitude and longitude of that area;
[0024] S23: During aerial surveying, the drone must maintain a predetermined flight altitude to ensure that the drone's rotor can blow away the rice and obtain clear information about the rice leaves.
[0025] S24: Import the captured rice growth images into the computer using the image transfer software AirDroid, perform image stitching and correction, and generate usable rice nutrient deficiency sample images.
[0026] Preferably, in step S3, the feature engineering processing includes image preprocessing and numerical preprocessing, wherein,
[0027] The image preprocessing includes the following steps:
[0028] S31: By contrast stretching and adaptive histogram equalization, the contrast and clarity of rice nutrient deficiency sample images in the nutrient deficiency sample dataset are improved, making the leaf color in the rice nutrient deficiency sample images in the nutrient deficiency sample dataset more prominent, and eliminating the influence of light and shadow in rice growth images.
[0029] S32: All collected images of rice samples showing normal growth and those showing nutrient deficiency are uniformly cropped and resized to 224×224 pixels, and then converted to ToTensor format;
[0030] The numerical preprocessing specifically involves: channel normalization. For each pixel in the rice nutrient deficiency sample image after image preprocessing, the ratio of its R, G, and B channel values to the sum of these three channels is calculated. The calculation expression is as follows:
[0031]
[0032] Preferably, in step S4, the steps for constructing the rice nutrient deficiency type determination and prediction model are as follows:
[0033] S41: The rice nutrient deficiency sample images after feature engineering are divided into training set and validation set in a 7:3 ratio;
[0034] S42: Create separate folders for rice images of nitrogen-deficient, phosphorus-deficient, and potassium-deficient fertilizers, as subfolders for the dataset and test set, and place them under the main folder;
[0035] S43: Construct a convolutional neural network image classification model and iteratively train it on rice images in the training set that show normal rice growth, rice deficient in nitrogen, phosphorus, and potassium fertilizer.
[0036] S44: Evaluate the convolutional neural network image classification model. By calculating three metrics—accuracy, recall, and F1 score—optimize the structure, algorithm, and learning rate parameters of the convolutional neural network image classification model to improve its performance.
[0037] S45: Use the convolutional neural network image classification model with the required accuracy as the model for determining and predicting the types of rice lacking fertilizer.
[0038] Preferably, in step S43, the iterative training step of using rice images of normal growth, nitrogen deficiency, phosphorus deficiency, and potassium deficiency in the training set is as follows:
[0039] S431: The preprocessed images of normally growing rice and rice samples lacking various fertilizers are fed into the input layer of the convolutional neural network image classification model. The output results are obtained by passing the data through convolutional layers, pooling layers, and fully connected layers.
[0040] S432: Calculate the loss function to evaluate the accuracy of the prediction results of the convolutional neural network image classification model;
[0041] S433: Calculate the gradient of the loss function with respect to each parameter, update the model parameters based on the gradient, select the stochastic gradient descent (SGD) optimization algorithm to minimize the loss function, and make the convolutional neural network image classification model more accurate in the next iteration.
[0042] S434: Perform iterative training by classifying rice sample images of rice growing normally and rice lacking various fertilizers multiple times, so that the classification accuracy of the convolutional neural network image classification model meets the requirements.
[0043] Preferably, in step S5, a normal rice planting experimental field is established, the rice planting experimental field is divided into regions, each region is marked and its geographic coordinates are transformed; including the following steps:
[0044] S51: The rice planting experimental field is prepared in the same way as in steps S11-S14, and a compound fertilizer containing nitrogen, phosphorus and potassium in a predetermined ratio is applied as the base fertilizer for rice planting.
[0045] S52: Based on the boundary and area information of the rice planting experimental field and the effective sowing width of the fertilizer application equipment, the rice planting experimental field is divided into several square areas of equal area; wherein, the effective sowing width of the fertilizer application equipment is the side length a, and the area of each square area is a. 2 ;
[0046] S53: Use RTK technology to label and locate the square area.
[0047] Preferably, in step S6, a fertilizer prescription map is generated based on the rice nutrient deficiency type determination and prediction model, and the regional fertilizer type and fertilizer amount are determined, including the following steps:
[0048] S61: Take real-time growth images of rice in a normal rice planting experimental field using an aerial surveying drone 10 days and 30 days after rice transplanting. The aerial surveying drone needs to maintain a preset flight altitude during the aerial survey, and the preset flight altitude needs to ensure that the clarity of the rice leaves captured meets the requirements.
[0049] S62: The image transmission software AirDroid is used to transmit the rice growth images taken by the aerial survey drone to the background computer in the control console. The background computer automatically inputs the rice growth images into the established rice nutrient deficiency type identification and prediction model.
[0050] S63: Rice Nutrient Deficiency Type Identification and Prediction Model uses three channels, R, G, and B, to compare the leaf colors of normally growing and nutrient-deficient rice to classify the color range of rice leaves.
[0051] S64: Using the color range of rice leaves as the input signal, the degree of rice nutrient deficiency is determined by the rice nutrient deficiency type identification and prediction model, and then the type and amount of nutrient deficiency are predicted and output.
[0052] S65: Output a grayscale image of the degree of nutrient deficiency in rice, using white as the feature of no nutrient deficiency and black as the feature of severe nutrient deficiency, and use the grayscale image to represent the degree of nutrient deficiency in rice to generate a fertilizer prescription map.
[0053] Preferably, in step S7, the application rates of various fertilizers are generated according to the fertilizer prescription diagram, and the variable fertilizer application device is controlled to apply fertilizer separately, specifically including the following steps:
[0054] S71: Fertilization operation is carried out using a fertilization drone equipped with a variable fertilizer application device; wherein, the three fertilizer tanks in the variable fertilizer application device are respectively filled with nitrogen, phosphorus and potassium fertilizers, and a screw conveyor is arranged at the bottom of the fertilizer tank to apply fertilizers by using a screw conveyor forced fertilizer application method.
[0055] S72: The drone performs fixed-point and quantitative operations based on the fertilizer prescription map. It uses RTK technology to locate the position information of the fertilizer drone in real time and transmits it to the control console to match the fertilizer prescription map of the corresponding area.
[0056] S73: Controls the rotation speed of the auger fertilizer dispensers under the three fertilizer tanks according to the electrical signals transmitted from the control console;
[0057] S74: The control console monitors and controls the rotation time of each auger fertilizer applicator, thereby controlling the application rate of various fertilizers in the corresponding area.
[0058] Preferably, the mathematical expression for the fertilizer discharge rate of the variable fertilization device is as follows:
[0059]
[0060] Where q represents the amount of fertilizer discharged per forced layer; d represents the diameter of the fertilizer discharge wheel; f represents the cross-sectional area of the groove; r represents the fertilizer capacity; λ (n) represents the fertilizer filling rate; x represents the tooth spacing.
[0061] Compared with the prior art, the present invention has the following advantages:
[0062] The deep learning-based method for determining rice nutrient deficiency types and applying variable fertilizers of the present invention constructs a model for determining and predicting rice nutrient deficiency types, and determines the type and amount of fertilizer to be applied to rice in the corresponding area based on the model, and generates a fertilizer prescription map for subsequent regional variable fertilizer application. This can quickly solve the problem of nutrient deficiency in rice growth, thereby improving rice quality and yield. Attached Figure Description
[0063] Figure 1 This is a flowchart illustrating the deep learning-based method for determining the types of fertilizer deficiency in rice and applying variable fertilizers, as described in this invention.
[0064] Figure 2 This is a diagram of a fertilizer prescription.
[0065] Figure 3 This is a fertilization path planning diagram. Detailed Implementation
[0066] The present invention will be further described below with reference to embodiments and accompanying drawings, but the implementation of the present invention is not limited thereto.
[0067] See Figures 1-3 The present invention provides a method for determining the types of nutrient deficiency in rice and applying variable fertilization based on deep learning, comprising the following steps:
[0068] S1: Establish a rice fertilizer deficiency experimental field and plant rice varieties deficient in nitrogen, phosphorus, and potassium respectively;
[0069] S2: Use aerial survey drones to capture images of symptoms of nitrogen, phosphorus, and potassium deficiency in rice, and use these images to create a dataset of fertilizer deficiency samples.
[0070] S3: Perform feature engineering on the nutrient-deficient sample dataset;
[0071] S4: By constructing training sets, test sets, and validation sets, a model for determining and predicting the types of nutrient deficiency in rice is established.
[0072] S5: Establish normal rice planting experimental fields, divide the rice planting experimental fields into regions, mark each region and perform geographical coordinate transformation;
[0073] S6: Use aerial survey drones to capture real-time growth images of rice in normal rice planting experimental fields. Based on the rice nutrient deficiency type identification and prediction model, determine the real-time growth images of rice and obtain the location information of the area corresponding to the real-time growth image of rice. In this way, generate a fertilizer prescription map of the current area, thereby determining the fertilizer type and amount for the area.
[0074] S7: Based on the application rates of various fertilizers generated from the fertilizer prescription diagram, control the variable fertilizer application device to discharge fertilizer;
[0075] S8: After fertilization is completed, continue with the determination of fertilizer type and fertilization operation in the next area.
[0076] See Figures 1-3 In step S1, the steps for establishing a rice nutrient deficiency experimental field are as follows:
[0077] S11: Select three square paddy fields of equal area with the same soil moisture content and soil fertility;
[0078] S12: Perform land preparation treatment on the three selected square paddy fields, including but not limited to weeding, tilling the land, and leveling the soil surface;
[0079] S13: Apply compound fertilizers that are "nitrogen-deficient", "phosphorus-deficient", and "potassium-deficient" to the three square paddy fields as base fertilizers for the early stage of rice planting;
[0080] S14: In subsequent planting, topdressing should be carried out according to the principle of "no nitrogen application to nitrogen-deficient plots", "no phosphorus application to phosphorus-deficient plots", and "no potassium application to potassium-deficient plots".
[0081] See Figures 1-3 In step S2, the steps for establishing the nutrient deficiency sample dataset are as follows:
[0082] S21: Set the flight path of the aerial survey drone. The flight path should ensure that the aerial survey drone can capture images of rice growth in all the rice nutrient deficiency test fields.
[0083] S22: Before the aerial survey drone takes off, it is necessary to update the geographical location information of the GPS module in the drone flight control system, and also check the operation of the IMU in the flight control system to ensure the drone is stable during the shooting process. After takeoff, RTK real-time dynamic measurement technology is used to enable the console to match the correspondence between the captured rice growth images and geographical coordinates.
[0084] S23: During aerial surveying, the drone must maintain a predetermined flight altitude to ensure that the drone's rotor can blow away the rice and obtain clear information about the rice leaves.
[0085] S24: Import the captured rice growth images into the computer using the image transfer software AirDroid, perform image stitching and correction, and generate usable rice nutrient deficiency sample images.
[0086] See Figures 1-3 In step S3, the feature engineering process includes image preprocessing and numerical preprocessing, wherein...
[0087] The image preprocessing includes the following steps:
[0088] S31: By contrast stretching and adaptive histogram equalization, the contrast and clarity of rice nutrient deficiency sample images in the nutrient deficiency sample dataset are improved, making the leaf color in the rice nutrient deficiency sample images in the nutrient deficiency sample dataset more prominent, and eliminating the influence of light and shadow in rice growth images.
[0089] S32: All collected images of rice samples showing normal growth and those showing nutrient deficiency are uniformly cropped and resized to 224×224 pixels, and then converted to ToTensor format;
[0090] The numerical preprocessing specifically involves: channel normalization. For each pixel in the rice nutrient deficiency sample image after image preprocessing, the ratio of its R, G, and B channel values to the sum of these three channels is calculated. The calculation expression is as follows:
[0091]
[0092] See Figures 1-3 In step S4, the steps for constructing the rice nutrient deficiency type identification and prediction model are as follows:
[0093] S41: The rice nutrient deficiency sample images after feature engineering are divided into training set and validation set in a 7:3 ratio;
[0094] S42: Create separate folders for rice images of nitrogen-deficient, phosphorus-deficient, and potassium-deficient fertilizers, as subfolders for the dataset and test set, and place them under the main folder;
[0095] S43: Construct a convolutional neural network image classification model and iteratively train it on rice images in the training set that show normal rice growth, rice deficient in nitrogen, phosphorus, and potassium fertilizer.
[0096] S44: Evaluate the convolutional neural network image classification model. By calculating three metrics—accuracy, recall, and F1 score—optimize the structure, algorithm, and learning rate parameters of the convolutional neural network image classification model to improve its performance.
[0097] S45: Use the convolutional neural network image classification model with the required accuracy as the model for determining and predicting the types of rice lacking fertilizer.
[0098] See Figures 1-3 In step S43, the iterative training steps for rice images in the training set showing normal growth, nitrogen deficiency, phosphorus deficiency, and potassium deficiency are as follows:
[0099] S431: The preprocessed images of normally growing rice and rice samples lacking various fertilizers are fed into the input layer of the convolutional neural network image classification model. The output results are obtained by passing the data through convolutional layers, pooling layers, and fully connected layers.
[0100] S432: Calculate the loss function to evaluate the accuracy of the prediction results of the convolutional neural network image classification model;
[0101] S433: Calculate the gradient of the loss function with respect to each parameter, update the model parameters based on the gradient, select the stochastic gradient descent (SGD) optimization algorithm to minimize the loss function, and make the convolutional neural network image classification model more accurate in the next iteration.
[0102] S434: Perform iterative training by classifying rice sample images of rice growing normally and rice lacking various fertilizers multiple times, so that the classification accuracy of the convolutional neural network image classification model meets the requirements.
[0103] See Figures 1-3 In step S5, a normal rice planting experimental field is established, the field is divided into regions, each region is marked, and its geographic coordinates are transformed; this includes the following steps:
[0104] S51: The rice planting experimental field is prepared in the same way as in steps S11-S14, and a compound fertilizer containing nitrogen, phosphorus and potassium in a predetermined ratio is applied as the base fertilizer for rice planting.
[0105] S52: Based on the boundary and area information of the rice planting experimental field and the effective sowing width of the fertilizer application equipment, the rice planting experimental field is divided into several square areas of equal area; wherein, the effective sowing width of the fertilizer application equipment is the side length a, and the area of each square area is a. 2 ;
[0106] S53: Use RTK technology to label and locate the square area.
[0107] See Figures 1-3 In step S6, a fertilizer prescription map is generated based on the rice nutrient deficiency type identification and prediction model, and the regional fertilizer type and amount are determined, including the following steps:
[0108] S61: Take real-time growth images of rice in a normal rice planting experimental field using an aerial surveying drone 10 days and 30 days after rice transplanting. The aerial surveying drone needs to maintain a preset flight altitude during the aerial survey, and the preset flight altitude needs to ensure that the clarity of the rice leaves captured meets the requirements.
[0109] S62: Using the image transmission software AirDroid, the rice growth images captured by the aerial survey drone are transmitted to the back-end computer in the control console. The back-end computer automatically inputs the rice growth images into the established rice nutrient deficiency type identification and prediction model.
[0110] S63: Rice Nutrient Deficiency Type Identification and Prediction Model uses three channels, R, G, and B, to compare the leaf colors of normally growing and nutrient-deficient rice to classify the color range of rice leaves.
[0111] S64: Using the color range of rice leaves as the input signal, the degree of rice nutrient deficiency is determined by the rice nutrient deficiency type identification and prediction model, and then the type and amount of nutrient deficiency are predicted and output.
[0112] S65: Output a grayscale image of the degree of nutrient deficiency in rice, using white as a feature of no nutrient deficiency and black as a feature of severe nutrient deficiency. Represent the degree of nutrient deficiency in rice using a grayscale image and generate a fertilizer prescription map (e.g., ...). Figure 2 (As shown).
[0113] See Figures 1-3 In step S7, the application rates of various fertilizers are generated according to the fertilizer prescription diagram, and the variable fertilizer application device is controlled to apply fertilizer separately. This specifically includes the following steps:
[0114] S71: Fertilization operation is carried out using a fertilization drone equipped with a variable fertilizer application device; wherein, the three fertilizer tanks in the variable fertilizer application device are respectively filled with nitrogen, phosphorus and potassium fertilizers, and a screw conveyor is arranged at the bottom of the fertilizer tank to apply fertilizers by using a screw conveyor forced fertilizer application method.
[0115] S72: The drone performs fixed-point and quantitative operations based on the fertilizer prescription map. It uses RTK technology to locate the position information of the fertilizer drone in real time and transmits it to the control console to match the fertilizer prescription map of the corresponding area.
[0116] S73: Controls the rotation speed of the auger fertilizer dispensers under the three fertilizer tanks according to the electrical signals transmitted from the control console;
[0117] S74: The control console monitors and controls the rotation time of each auger fertilizer applicator, thereby controlling the application rate of various fertilizers in the corresponding area.
[0118] See Figures 1-3 The mathematical expression for the fertilizer discharge rate of the variable fertilization device is as follows:
[0119]
[0120] Where q represents the amount of fertilizer discharged per forced layer; d represents the diameter of the fertilizer discharge wheel; f represents the cross-sectional area of the groove; r represents the fertilizer capacity; λ (n) represents the fertilizer filling rate; x represents the tooth spacing.
[0121] The above are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above content. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A method for determining the types of nutrient deficiency in rice and applying variable fertilization based on deep learning, characterized in that, Includes the following steps: S1: Establish a rice fertilizer deficiency experimental field and plant rice varieties deficient in nitrogen, phosphorus, and potassium respectively; S2: A dataset of fertilizer deficiency samples is created by using aerial survey drones to capture images of nitrogen, phosphorus, and potassium deficiencies in rice. The steps for creating this dataset are as follows: S21: Set the flight path of the aerial survey drone. The flight path should ensure that the aerial survey drone can capture images of rice growth in all the rice nutrient deficiency test fields. S22: Before the aerial survey drone takes off, the geographical location information of the GPS module in the drone's flight control system needs to be updated, and the operation of the IMU needs to be checked to ensure the drone's stability during the shooting process; after takeoff, RTK real-time dynamic measurement technology is used to establish a correspondence between the rice growth photos taken by the drone in each area and the latitude and longitude of that area. S23: During aerial surveying, the drone must maintain a predetermined flight altitude to ensure that the drone's rotor can blow away the rice and obtain clear information about the rice leaves. S24: Import the captured rice growth images into the computer using the image transfer software AirDroid, perform image stitching and correction, and generate usable rice nutrient deficiency sample images; S3: Perform feature engineering on the nutrient-deficient sample dataset; S4: By constructing training sets, test sets, and validation sets, a model for determining and predicting the types of nutrient deficiency in rice is established. S5: Establish normal rice planting experimental fields, divide the rice planting experimental fields into regions, mark each region and perform geographical coordinate transformation; S6: Using an aerial survey drone, real-time growth images of rice in a normal rice planting experimental field are captured. Based on a rice nutrient deficiency type identification and prediction model, the real-time growth images are analyzed, and the location information of the corresponding area is obtained. This information is used to generate a fertilizer prescription map for the current area, thereby determining the type and amount of fertilizer to be applied. The process of generating the fertilizer prescription map based on the rice nutrient deficiency type identification and prediction model and determining the type and amount of fertilizer to be applied in the area includes the following steps: S61: Take real-time growth images of rice in a normal rice planting experimental field using an aerial surveying drone 10 days and 30 days after rice transplanting. The aerial surveying drone needs to maintain a preset flight altitude during the aerial survey, and the preset flight altitude needs to ensure that the clarity of the rice leaves captured meets the requirements. S62: The image transmission software AirDroid is used to transmit the rice growth images taken by the aerial survey drone to the background computer in the control console. The background computer automatically inputs the rice growth images into the established rice nutrient deficiency type identification and prediction model. S63: Rice Nutrient Deficiency Type Identification and Prediction Model uses three channels, R, G, and B, to compare the leaf colors of normally growing and nutrient-deficient rice to classify the color range of rice leaves. S64: Using the color range of rice leaves as the input signal, the degree of rice nutrient deficiency is determined by the rice nutrient deficiency type identification and prediction model, and then the type and amount of nutrient deficiency are predicted and output. S65: Output a grayscale image of the degree of nutrient deficiency in rice, using white as the feature of no nutrient deficiency and black as the feature of severe nutrient deficiency, and use the grayscale image to represent the degree of nutrient deficiency in rice to generate a fertilizer prescription map. S7: Based on the application rates of various fertilizers generated from the fertilizer prescription diagram, control the variable fertilizer application device to discharge fertilizer; S8: After fertilization is completed, continue with the determination of fertilizer type and fertilization operation in the next area.
2. The method for determining rice nutrient deficiency types and applying variable fertilizers based on deep learning according to claim 1, characterized in that, In step S1, the steps for establishing a rice nutrient deficiency experimental field are as follows: S11: Select three square paddy fields of equal area with the same soil moisture content and soil fertility; S12: Perform land preparation treatment on the three selected square paddy fields, including but not limited to weeding, tilling the land, and leveling the soil surface; S13: Apply compound fertilizers that are "nitrogen-deficient", "phosphorus-deficient", and "potassium-deficient" to three square paddy fields as base fertilizers for the early stage of rice planting; S14: In subsequent planting, topdressing should be carried out according to the principle of "no nitrogen application to nitrogen-deficient plots", "no phosphorus application to phosphorus-deficient plots", and "no potassium application to potassium-deficient plots".
3. The method for determining rice nutrient deficiency types and applying variable fertilizers based on deep learning according to claim 2, characterized in that, In step S3, the feature engineering process includes image preprocessing and numerical preprocessing, wherein, The image preprocessing includes the following steps: S31: By contrast stretching and adaptive histogram equalization, the contrast and clarity of rice nutrient deficiency sample images in the nutrient deficiency sample dataset are improved, making the leaf color in the rice nutrient deficiency sample images in the nutrient deficiency sample dataset more prominent, and eliminating the influence of light and shadow in rice growth images. S32: All collected images of rice samples showing normal growth and those showing nutrient deficiency are uniformly cropped and resized to 224×224 pixels, and then converted to ToTensor format; The numerical preprocessing specifically involves: channel normalization. For each pixel in the rice nutrient deficiency sample image after image preprocessing, the ratio of its R, G, and B channel values to the sum of these three channels is calculated. The calculation expression is as follows: 。 4. The method for determining rice nutrient deficiency types and applying variable fertilizers based on deep learning according to claim 1, characterized in that, In step S4, the steps for constructing the rice nutrient deficiency type identification and prediction model are as follows: S41: The rice nutrient deficiency sample images after feature engineering are divided into training set and validation set in a 7:3 ratio; S42: Create separate folders for rice images of nitrogen-deficient, phosphorus-deficient, and potassium-deficient fertilizers, as subfolders for the dataset and test set, and place them under the main folder; S43: Construct a convolutional neural network image classification model and iteratively train it on rice images in the training set that show normal rice growth, rice deficient in nitrogen, phosphorus, and potassium fertilizer. S44: Evaluate the convolutional neural network image classification model. By calculating three metrics—accuracy, recall, and F1 score—optimize the structure, algorithm, and learning rate parameters of the convolutional neural network image classification model to improve its performance. S45: Use the convolutional neural network image classification model with the required accuracy as the model for determining and predicting the types of rice lacking fertilizer.
5. The method for determining rice nutrient deficiency types and applying variable fertilizers based on deep learning according to claim 4, characterized in that, In step S43, the iterative training steps for rice images in the training set showing normal growth, nitrogen deficiency, phosphorus deficiency, and potassium deficiency are as follows: S431: The preprocessed images of normally growing rice and rice samples lacking various fertilizers are fed into the input layer of the convolutional neural network image classification model. The output results are obtained by passing the data through convolutional layers, pooling layers, and fully connected layers. S432: Calculate the loss function to evaluate the accuracy of the prediction results of the convolutional neural network image classification model; S433: Calculate the gradient of the loss function with respect to each parameter, update the model parameters based on the gradient, select the stochastic gradient descent (SGD) optimization algorithm to minimize the loss function, and make the convolutional neural network image classification model more accurate in the next iteration. S434: Perform iterative training by classifying rice sample images of rice growing normally and rice lacking various fertilizers multiple times, so that the classification accuracy of the convolutional neural network image classification model meets the requirements.
6. The method for determining rice nutrient deficiency types and applying variable fertilizers based on deep learning according to claim 1, characterized in that, In step S5, a normal rice planting experimental field is established, the field is divided into regions, each region is marked, and its geographic coordinates are transformed; this includes the following steps: S51: The rice planting experimental field is prepared in the same way as in steps S11-S12, and a compound fertilizer containing nitrogen, phosphorus and potassium in a predetermined ratio is applied as the base fertilizer for rice planting. S52: Based on the boundary and area information of the rice planting experimental field and the effective sowing width of the fertilizer application equipment, the rice planting experimental field is divided into several square areas of equal area; wherein, the effective sowing width of the fertilizer application equipment is the side length 'a' of each square area, and the area of each square area is... ; S53: Use RTK technology to label and locate the square area.
7. The method for determining rice nutrient deficiency types and applying variable fertilizers based on deep learning according to claim 1, characterized in that, In step S7, the application rates of various fertilizers are generated according to the fertilizer prescription diagram, and the variable fertilizer application device is controlled to apply fertilizer separately. This specifically includes the following steps: S71: Fertilization operation is carried out using a fertilization drone equipped with a variable fertilizer application device; wherein, the three fertilizer tanks in the variable fertilizer application device are respectively filled with nitrogen, phosphorus and potassium fertilizers, and a screw conveyor is arranged at the bottom of the fertilizer tank to apply fertilizers by using a screw conveyor forced fertilizer application method. S72: The drone performs fixed-point and quantitative operations based on the fertilizer prescription map. It uses RTK technology to locate the position information of the fertilizer drone in real time and transmits it to the control console to match the fertilizer prescription map of the corresponding area. S73: Controls the rotation speed of the auger fertilizer dispensers under the three fertilizer tanks according to the electrical signals transmitted from the control console; S74: The control console monitors and controls the rotation time of each auger fertilizer applicator, thereby controlling the application rate of various fertilizers in the corresponding area.
8. The method for determining rice nutrient deficiency types and applying variable fertilizers based on deep learning according to claim 1, characterized in that, The mathematical expression for the fertilizer discharge rate of the variable fertilization device is as follows: ； Where q represents the amount of fertilizer discharged per forced layer; d represents the diameter of the fertilizer discharge wheel; f represents the cross-sectional area of the groove; and r represents the fertilizer capacity. represents the fertilizer filling rate; x represents the tooth spacing.