Intelligent control method of field water and fertilizer based on Internet of Things

Abstract

The present invention relates to an intelligent field water and fertilizer control method based on the Internet of Things. The method collects crop images, soil water and fertilizer data, and meteorological forecast data to identify crop types and growth stages, generates an initial water and fertilizer control plan for the current time node, creates a soil water and fertilizer prediction model to predict soil water and fertilizer change trends within a specified time period, adjusts the initial water and fertilizer control plan for the current time node according to the predicted soil water and fertilizer change trends, obtains growth cycle characteristics corresponding to crop types based on big data, formulates a dynamic water and fertilizer control plan, executes the adjusted water and fertilizer control plan at the current time node, and executes the dynamic water and fertilizer control plan at the specified time node. The intelligent field water and fertilizer control method integrates Internet of Things perception, intelligent identification, trend prediction, dynamic programming, and closed-loop feedback, realizes full-process intelligence from data collection to plan execution, improves water and fertilizer utilization efficiency, and reduces agricultural production costs.

Description

Technical Field

[0001] The present invention belongs to the technical field of agricultural water and fertilizer control, and in particular relates to an intelligent field water and fertilizer control method based on the Internet of Things. Background Art

[0002] In agricultural production, field water and fertilizer management is a key link influencing crop growth quality and yield. Traditional field water and fertilizer management relies primarily on manual experience, with growers irrigating and fertilizing based on subjective judgment or fixed cycles. This leads to problems such as low precision in water and fertilizer supply and severe resource waste. For example, manual irrigation often fails to accurately grasp the actual soil moisture content, resulting in over-irrigation causing soil compaction or waterlogging, or insufficient irrigation affecting crop growth. Manual fertilization is prone to imbalanced fertilizer ratios, which not only reduces fertilizer utilization but can also cause soil pollution and water eutrophication due to excess nutrients.

[0003] With the advancement of agricultural modernization, some sensor-based water and fertilizer control technologies have emerged. These use soil sensors to collect moisture and nutrient data, enabling preliminary automated irrigation and fertilization. However, existing technologies still have significant limitations: First, they lack targeted adaptation to crop types and growth stages, making it difficult for the same control logic to meet the differentiated needs of different crops or the different growth stages of the same crop. Second, they lack predictive capabilities, relying solely on passive adjustments based on real-time data. They are unable to plan water and fertilizer plans in advance based on meteorological changes and natural trends in soil water and fertilizer, making adjustments prone to lags in response to rainfall, high temperatures, and other weather conditions. Third, closed-loop feedback mechanisms are imperfect, lacking a dynamic correlation between crop growth status and water and fertilizer plans. When crop growth deviates from expectations, it is difficult to quickly identify the cause and optimize the control strategy.

[0004] In addition, existing crop identification technologies mostly rely on manual input and lack the ability to automatically identify species and growth stages based on images, resulting in limited accuracy in the initial settings of water and fertilizer plans; soil water and fertilizer prediction models mostly use simple linear fitting, which does not fully integrate the influence of historical data and meteorological factors, resulting in low prediction accuracy; at the same time, plan modification and model optimization lack standardized processes, making it difficult to adapt to complex and changeable field environments.

[0005] To solve the above problems, the present invention intends to provide a field water and fertilizer intelligent control method that integrates Internet of Things perception, intelligent identification, trend prediction, dynamic planning and closed-loop feedback, to realize the intelligence of the entire process from data collection to solution execution, improve the efficiency of water and fertilizer utilization, and reduce agricultural production costs. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent field water and fertilizer control method based on the Internet of Things, so as to realize the intelligence of the entire process from data collection to program execution, improve the efficiency of water and fertilizer utilization, and reduce agricultural production costs.

[0007] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0008] The field water and fertilizer intelligent control method based on the Internet of Things includes the following steps:

[0009] S1: Capture images of crops in the controlled field, collect specified water and fertilizer data of the current field soil, obtain meteorological forecast data based on big data, and transmit all data to the cloud server;

[0010] S2: The cloud server identifies crop species based on the collected crop images and identifies the growth stage based on the crop species;

[0011] S3: Based on the crop type, growth stage and current field soil water and fertilizer data, the basic water and fertilizer parameter library of the corresponding crop is called to generate the initial water and fertilizer control plan for the current time node;

[0012] S4: Create a soil water and fertilizer prediction model and predict the soil water and fertilizer change trend within a specified time period using historical soil water and fertilizer data and historical meteorological data;

[0013] S5: adjusting the initial water and fertilizer control plan at the current time node according to the predicted soil water and fertilizer change trend to obtain an adjusted water and fertilizer control plan at the current time node;

[0014] S6: Obtain the growth cycle characteristics corresponding to the crop types based on big data, and formulate a dynamic water and fertilizer control plan for a specified time node based on the growth cycle characteristics and soil water and fertilizer change trends, and execute the adjustment water and fertilizer control plan at the current time node, and execute the dynamic water and fertilizer control plan at the specified time node.

[0015] Preferably, it also includes a real-time crop growth monitoring and feedback process, as follows:

[0016] S7: Collecting actual crop growth data at a specified time point, comparing the actual crop growth data with the growth cycle characteristics of the crop type obtained based on big data, and determining whether the specified indicators in the actual crop growth data meet the preset growth requirements. If so, executing the dynamic water and fertilizer control plan at the specified time point according to step S6; if not, executing step S8;

[0017] S8: Collecting actual soil water and fertilizer data at a specified time point, comparing the actual soil water and fertilizer data with the soil water and fertilizer data at the corresponding time point in the predicted soil water and fertilizer change trend, and determining whether the actual soil water and fertilizer data and the predicted soil water and fertilizer data at the corresponding time point are both within a preset error range. If so, revising the initial water and fertilizer control plan; if not, executing step S9;

[0018] S9: Compare the actual soil water and fertilizer data with the various indicators in the predicted soil water and fertilizer data at the corresponding time node one by one, obtain the indicators whose differences exceed the corresponding preset error range, and adjust the parameters of the soil water and fertilizer prediction model based on the differences of the indicators that exceed the corresponding preset error range.

[0019] Preferably, the detailed process of revising the initial water and fertilizer control scheme in step S8 is as follows:

[0020] S81: extracting specified parameters from the initial water and fertilizer control plan, including irrigation water volume, irrigation time node, fertilizer type, fertilizer volume, and fertilizer time node;

[0021] S82: setting correction coefficients for various designated parameters in the initial water and fertilizer control plan based on actual soil water and fertilizer data and the needs of the crop at the corresponding growth stage;

[0022] S83: Modify the initial water and fertilizer control plan based on the correction coefficient, re-execute steps S4-S6, and synchronously update to the cloud server.

[0023] Preferably, the specific process of step S2 is as follows:

[0024] S21: perform uniform resizing, grayscale processing and denoising on the crop images;

[0025] S22: Identify the crop body and background from the crop image using a preset semantic segmentation model, and retain the initial valid areas of the crop leaves, stems, and fruits in the image;

[0026] S23: Binarizing the initial effective area, selecting a structural element that is suitable for crop characteristics, traversing each pixel in the crop image, marking the pixel as foreground if the pixel and all pixels in its 8-neighborhood are foreground, otherwise marking it as background, and eliminating isolated noise points whose area is smaller than a specified number of pixels;

[0027] S24: traverse each pixel in the crop image from which isolated noise has been eliminated, mark the pixel and any pixel in its 8-neighborhood that has at least one foreground pixel as foreground, repair minor breaks in the crop area, and restore the complete shape of the crop body;

[0028] S25: extracting multidimensional features from the crop image that has restored the complete morphology of the crop body using the pre-trained neural network model, calling a preset crop species feature library to match the extracted multidimensional features with the crop species using a preset feature vector comparison algorithm, and selecting the crop species with the highest matching degree as the species recognition result;

[0029] S26: Loading the growth stage feature library of the corresponding crop according to the species recognition result, obtaining the corresponding growth state features in the crop image from the multidimensional features, and performing feature matching based on the growth state features and the growth stage feature library of the corresponding crop.

[0030] Preferably, the specific process of performing multidimensional feature extraction in step S25 is as follows:

[0031] S251: extracting different features respectively through the multi-layer feature extraction layer of the neural network model, and extracting low-order features including color features and edge features through the shallow convolution layer;

[0032] S252: extracting morphological and texture features through the middle convolutional layer;

[0033] S253: Extract growth status features and semantic association features through deep convolutional layers;

[0034] S254: low-order, mid-order, and high-order feature vectors are merged into a fused feature vector of the crop image by using feature splicing.

[0035] Preferably, the specific process of step S3 is as follows:

[0036] S31: Performing a primary index based on the crop type and locating the water and fertilizer parameter unit corresponding to the crop type;

[0037] S32: Perform secondary indexing based on the growth stage, locate the corresponding growth stage sub-library under the water and fertilizer parameter unit corresponding to the crop type, and extract basic water and fertilizer parameters;

[0038] S33: Setting an initial water and fertilizer control plan based on the current field soil water and fertilizer data and the basic water and fertilizer parameters:

[0039] The current field soil water and fertilizer data are compared with the basic water and fertilizer parameters one by one, the difference between each water and fertilizer data is calculated, the actual supply is calculated based on the difference between each water and fertilizer data, and the execution parameters of the initial water and fertilizer control plan are generated based on the actual supply, including irrigation parameters and fertilization parameters.

[0040] Preferably, the specific process of step S4 is as follows:

[0041] S41: Preprocess historical soil water and fertilizer data and historical meteorological data, identify and replace abnormal data, and fill in missing values;

[0042] S42: constructing a training set and a test set based on the pre-processed historical soil water and fertilizer data and historical meteorological data in a specified ratio and inputting them into the soil water and fertilizer prediction model in sequence;

[0043] S43: The LSTM in the soil water and fertilizer prediction model is used to extract the time series features of soil water and fertilizer data. The random forest algorithm is used as a feature importance assessment tool to screen meteorological factors that significantly affect water and fertilizer changes.

[0044] S44: Optimize the hyperparameters of the soil water and fertilizer prediction model by determining the optimal number of LSTM neurons and time window length through grid search;

[0045] S45: Acquire soil water and fertilizer data at the prediction start time node and weather forecast data within the specified time period, and predict the soil water and fertilizer change trend within the specified time period based on the soil water and fertilizer data at the prediction start time node and weather forecast data within the specified time period through a soil water and fertilizer prediction model.

[0046] Preferably, the specific process of step S5 is as follows:

[0047] S51: Performing hysteresis data modeling based on the predicted soil water and fertilizer change trends, and establishing a hysteresis coefficient matrix based on the hysteresis data. The hysteresis coefficient matrix is ​​divided into dimensions according to "soil type × crop type × water and fertilizer type", and hysteresis coefficients of various water and fertilizer indicators corresponding to each dimension are set after each dimension;

[0048] S52: based on the predicted soil water and fertilizer change trends, drawing change curves of various water and fertilizer indicators, and performing time offset and amplitude correction on the change curves of various water and fertilizer indicators based on the hysteresis coefficients of the various water and fertilizer indicators;

[0049] S53: extracting the forecast data after lag correction, and marking the lag impact interval based on the forecast curve of the original specified time period;

[0050] S54: Retrieve the hysteresis-related parameters of the initial plan: clarify the theoretical time difference between "execution action-expected effect" in the initial water and fertilizer control plan;

[0051] S55: Adjust impact weights and identify lagging risk points;

[0052] S56: Adjust the initial water and fertilizer control plan at the current time node, dynamically correct the irrigation amount, and adjust the fertilizer amount.

[0053] Preferably, the specific process of step S6 is as follows:

[0054] S61: Based on the agricultural database and historical planting records, filter data by crop type and extract the characteristics of the specified growth cycle:

[0055] Stage division: germination stage, seedling stage), flowering stage, fruiting stage, and maturity stage;

[0056] Key indicators at this stage: peak fertilizer demand during the fruiting period and water-sensitive period;

[0057] Specify growth cycle characteristics for structuring: convert the specified growth cycle characteristics into computable parameters and generate a crop type-growth cycle characteristic table;

[0058] S62: performing regional adaptation and modification of growth characteristic parameters and variety adaptation for the designated growth cycle characteristics, and establishing a local exclusive growth cycle characteristic database;

[0059] S63: Plan designated time nodes, including key growth nodes and regular nodes, and set corresponding dynamic water and fertilizer control plans for each designated time node. Based on the growth cycle characteristics as the demand benchmark and the soil water and fertilizer change trends as the supply basis, calculate the supply and demand gap and establish the basic framework of the dynamic water and fertilizer control plan;

[0060] S64: Calculate dynamic parameters of the basic framework of the dynamic water and fertilizer control scheme, and establish the dynamic water and fertilizer control scheme at the specified time node based on the dynamic parameters.

[0061] The beneficial effects of the present invention include:

[0062] The field water and fertilizer intelligent control method provided by the present invention based on the Internet of Things collects crop images, soil water and fertilizer data, and meteorological forecast data to identify crop types and growth stages, generate an initial water and fertilizer control plan for the current time node, create a soil water and fertilizer prediction model to predict the soil water and fertilizer change trend within a specified time period, adjust the initial water and fertilizer control plan for the current time node according to the predicted soil water and fertilizer change trend, obtain the growth cycle characteristics corresponding to the crop type based on big data, formulate a dynamic water and fertilizer control plan, execute the adjusted water and fertilizer control plan at the current time node, and execute the dynamic water and fertilizer control plan at the specified time node. The field water and fertilizer intelligent control method integrates Internet of Things perception, intelligent identification, trend prediction, dynamic programming and closed-loop feedback, realizes the intelligence of the entire process from data collection to plan execution, improves water and fertilizer utilization efficiency, and reduces agricultural production costs.

[0063] First, by using image acquisition to identify crop types and growth stages, and combining soil water and fertilizer data to generate customized water and fertilizer plans, this solves the issues of empirical fertilization and blind irrigation in traditional manual management. Based on the differentiated needs of crops at different growth stages, the system automatically adjusts potassium fertilizer dosage and irrigation frequency, improving the match between water and fertilizer supply and crop needs, reducing fertilizer waste and water resource consumption, and lowering the risk of soil contamination caused by excess nutrients.

[0064] Secondly, with the help of the soil water and fertilizer prediction model, historical data and weather forecasts are combined to predict the trend of future soil conditions, and the prediction results are corrected through the lag coefficient matrix, which solves the limitation of existing technology that only passively responds to real-time data.

[0065] Furthermore, by comparing crop growth monitoring with soil data, a closed loop of plan execution, performance evaluation, and model optimization is formed. When crop growth deviates from expectations, the system can quickly identify whether the problem lies with the plan or the prediction model, and make targeted corrections. If actual soil water and fertilizer levels deviate significantly from the predictions, the model parameters are automatically adjusted. If crop growth remains poor after the plan is implemented, irrigation and fertilization parameters are corrected. This mechanism shortens the iterative optimization cycle of water and fertilizer plans and, with long-term use, improves crop yield stability.

[0066] Thirdly, the entire process from crop identification and plan generation to execution and regulation is automated, reducing more than 90% of manual intervention: crop features are extracted through semantic segmentation and neural networks, replacing manual judgment of crop growth stages; fertilization and irrigation time is automatically planned through a dynamic plan generation module, replacing manual recording and decision-making, while avoiding the risk of misjudgment in manual operations.

[0067] Finally, by modifying growth cycle characteristics based on regional and crop adaptation, combined with distributed sensors and multi-protocol communication, the system can adapt to different soil types, crop varieties, and growing regions. This allows for automatic adjustment of phosphate fertilizer absorption efficiency parameters in alkaline soils in the north, and for enhanced drainage-linked irrigation strategies in rainy southern regions, achieving precise water and fertilizer control for each region and crop. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The figure is a flow chart of the field water and fertilizer intelligent control method based on the Internet of Things of the present invention.

[0069] Figure 2 Schematic diagram of the soil water and fertilizer prediction model of the present invention.

[0070] Figure 3 Schematic diagram of the architecture of the neural network model of the present invention. DETAILED DESCRIPTION

[0071] The following is combined with Figures 1 to 3 , the present invention is further described in detail using tomatoes from a certain region as crops:

[0072] Example 1

[0073] See attached Figure 1 As shown, the field water and fertilizer intelligent control method based on the Internet of Things includes the following steps:

[0074] S1: Capture images of crops in the controlled field, collect specific water and fertilizer data for the current field soil, obtain meteorological forecast data based on big data, and transmit all data to a cloud server. Using high-definition industrial cameras and multispectral cameras, the multispectral cameras cover the 450nm, 550nm, and 650nm bands and are distributedly installed on field poles at a density of one group per 500 square meters, at a height of 2-3 meters, to ensure coverage of the crop canopy. Clock synchronization signals are sent via the IoT gateway to ensure that all cameras start collecting data at the same time, avoiding image deviations caused by differences in lighting. Specific water and fertilizer data, including soil moisture and soil nutrients, are collected using soil moisture sensors and soil nutrient sensors, respectively. Access to the agricultural meteorological service network through an API interface provides access to meteorological data for the field area.

[0075] S2: The cloud server identifies crop species based on the captured crop images and their growth stages. The crop images are resized, grayscaled, and denoised. The crop body and background are identified to obtain valid areas. Binarization is then performed to select structural elements that match the crop's characteristics. Multidimensional features are extracted and a pre-defined crop feature library is used to match the extracted multidimensional features to the crop species. The corresponding crop growth stage feature library is loaded to obtain the corresponding growth state features from the crop images.

[0076] S3: Based on the crop type, growth stage and current field soil water and fertilizer data, the basic water and fertilizer parameter library of the corresponding crop is called to generate the initial water and fertilizer control plan for the current time node.

[0077] S4: Create a soil water and fertilizer prediction model and predict the soil water and fertilizer change trend within a specified time period using historical soil water and fertilizer data and historical meteorological data;

[0078] S5: adjusting the initial water and fertilizer control plan at the current time node according to the predicted soil water and fertilizer change trend to obtain an adjusted water and fertilizer control plan at the current time node;

[0079] S6: Obtain the growth cycle characteristics corresponding to the crop types based on big data, and formulate a dynamic water and fertilizer control plan for a specified time node based on the growth cycle characteristics and soil water and fertilizer change trends, and execute the adjustment water and fertilizer control plan at the current time node, and execute the dynamic water and fertilizer control plan at the specified time node.

[0080] Another implementation of this embodiment also includes a real-time crop growth monitoring and feedback process, as follows:

[0081] S7: Collecting actual crop growth data at a specified time point, comparing the actual crop growth data with the growth cycle characteristics of the crop type obtained based on big data, and determining whether the specified indicators in the actual crop growth data meet the preset growth requirements. If so, executing the dynamic water and fertilizer control plan at the specified time point according to step S6; if not, executing step S8;

[0082] S8: Collect actual soil water and fertilizer data at a specified time point, compare the actual soil water and fertilizer data with the soil water and fertilizer data at the corresponding time point in the predicted soil water and fertilizer change trend, and determine whether the actual soil water and fertilizer data and the predicted soil water and fertilizer data at the corresponding time point are both within a preset error range. If so, it indicates that the predicted soil water and fertilizer data at the corresponding time point are normal, and it is necessary to re-execute steps S3-S6, and revise the initial water and fertilizer control plan. If not, it indicates that the prediction performance of the soil water and fertilizer prediction model is insufficient, and it is necessary to optimize the soil water and fertilizer prediction model, and execute step S9;

[0083] S9: Compare the actual soil water and fertilizer data with the various indicators in the predicted soil water and fertilizer data at the corresponding time node one by one, obtain the indicators whose differences exceed the corresponding preset error range, and adjust the parameters of the soil water and fertilizer prediction model based on the differences of the indicators that exceed the corresponding preset error range.

[0084] The detailed process of revising the initial water and fertilizer control plan in step S8 is as follows:

[0085] S81: extracting specified parameters from the initial water and fertilizer control plan, including irrigation water volume, irrigation time node, fertilizer type, fertilizer volume, and fertilizer time node;

[0086] S82: setting correction coefficients for various designated parameters in the initial water and fertilizer control plan based on actual soil water and fertilizer data and the needs of the crop at the corresponding growth stage;

[0087] S83: Modify the initial water and fertilizer control plan based on the correction coefficient, re-execute steps S4-S6, and synchronously update to the cloud server.

[0088] Since the implementation of the water and fertilizer control plan will cause the soil pH value to change to a certain extent, it is necessary to further adjust the initial water and fertilizer control plan based on the real-time monitoring of the soil pH value. The specific adjustment process is as follows:

[0089] Data is collected synchronously at designated time points. For example, if one of the designated time points is the 50th day of tomato fruiting, the actual growth data for tomatoes is: plant height measured at 110 cm, number of leaves measured at 8, and fruit diameter measured at 4.5 cm. Soil data includes actual water and fertilizer data (moisture content 26%, nitrogen 42 mg / kg, phosphorus 15 mg / kg, potassium 58 mg / kg), plus a newly measured soil pH value of 6.2.

[0090] Comparing the measured plant height of 110 cm with the standard plant height of 120 ± 5 cm for the mid-fruiting period in the growth cycle characteristic database, the deviation was -8.3%, not meeting the preset requirements, triggering soil data verification. Soil data comparison: The predicted phosphorus content was 18 mg / kg, the measured value was 15 mg / kg, a difference of -3 mg / kg, exceeding the ±2 mg / kg error range. The predicted potassium content was 55 mg / kg, the measured value was 58 mg / kg, a difference of +3 mg / kg, within the error range.

[0091] The above data indicates an abnormal phosphorus content forecast, requiring adjustment in step S9. Meanwhile, the other indicators require correction of the initial plan in step S8. The irrigation parameters extracted from the initial plan are: irrigation volume 12 L / m², irrigation time 2:00 PM, and a 3-day irrigation cycle. Based on the difference between the actual soil moisture (26%) and the crop requirement (30-35%), a basic correction factor K1 is set as (30-26) / (30-initial moisture 28) = 2.0. Based on the weather forecast (5 mm of rainfall the next day), an environmental correction factor K2 is set as 0.8 to compensate for rainfall. The overall correction factor K is calculated as K1 × K2 = 1.6.

[0092] Corrected irrigation water volume = 12 L / m² × 1.6 = 19.2 L / m² (rounded up to 20 L / m²); advance irrigation time to 10:00 AM to avoid rainy periods; shorten irrigation cycle to 2 days to accelerate humidity recovery. Re-execute S4-S6: Input the corrected parameters into the forecast model to generate a new 3-day humidity trend (targeting 32% on the third day).

[0093] Based on the measured soil pH value, the calcium, phosphorus and potassium in the initial water and fertilizer control plan are modified:

[0094] First, analyze the impact of pH on nutrient availability:

[0095] Effects of soil pH 6.2 (weakly acidic) on nutrients:

[0096] Phosphorus: It is most effective at pH 6.5-7.0. The current pH value reduces the phosphorus absorption rate by about 15%.

[0097] Calcium: It is easily fixed in weakly acidic soil, and its effectiveness is reduced by about 10%;

[0098] Potassium: pH value has little impact and its effectiveness is basically stable.

[0099] Secondly, extract calcium, phosphorus and potassium parameters and set correction coefficients:

[0100] Assume that the initial fertilization parameters in the initial water and fertilizer control plan are: 10g / m² of phosphorus fertilizer, 8g / m² of potassium fertilizer, and 5g / m² of calcium fertilizer; set the nutrient correction coefficient based on the pH value:

[0101] Phosphate fertilizer correction factor K 磷 = 1 / (1-15%)=1.18;

[0102] Calcium fertilizer correction coefficient K 钙 =1 / (1-10%)=1.11;

[0103] Potassium fertilizer correction factor K 钾 = 1.0 (no correction required);

[0104] Combined with the actual soil value and the demand difference (phosphorus demand 20mg / kg, actual value 15mg / kg, difference +5mg / kg), set the supplementary correction factor K 补 = 5 / (20-initial value 18)=2.5;

[0105] Final phosphate coefficient = K 磷 ×K 补 = 1.18×2.5=2.95;

[0106] Final calcium fertilizer coefficient = K 钙 ×1.2 (to prevent fruit cracking due to calcium deficiency) = 1.33.

[0107] Again, make the calcium, phosphorus and potassium program correction:

[0108] Corrected fertilizer amount:

[0109] Phosphorus fertilizer = 10g / m²×2.95=29.5g / m² (applied twice, with an interval of 1 day);

[0110] Calcium fertilizer = 5g / m²×1.33=6.65g / m² (apply simultaneously with phosphorus fertilizer to promote absorption);

[0111] Potassium fertilizer is maintained at 8g / m² (within the error range);

[0112] Adjustment of fertilization time: Avoid the period when soil moisture is greater than 30% (to prevent phosphorus leaching), and choose to fertilize at 9:00 am;

[0113] Simultaneous optimization: Add 0.1% citric acid to irrigation water (adjust the root zone pH to 6.5) to improve the effectiveness of phosphorus and calcium.

[0114] Example 2

[0115] The specific process of step S2 is as follows:

[0116] S21: perform uniform resizing, grayscale processing and denoising on the crop images;

[0117] S22: Identify the crop body and background from the crop image using a preset semantic segmentation model, and retain the initial valid areas of the crop leaves, stems, and fruits in the image;

[0118] S23: Binarize the initial valid area and select a structuring element that is suitable for the crop characteristics. For foliage crops, a 3×3 pixel square structuring element is set; for leaf crops, a 5×5 pixel circular structuring element is set. Each pixel in the crop image is traversed, and pixels where the pixel and all pixels in its 8-neighborhood are foreground (value 1) are marked as foreground. Otherwise, the pixel is marked as background (value 0). Isolated noise points with an area smaller than the specified number of pixels are eliminated, including false targets caused by reflections from soil particles and image sensor noise.

[0119] S24: Traverse each pixel in the crop image after eliminating isolated noise, mark the pixel and its eight-neighborhood pixels with at least one foreground pixel as foreground, and repair minor fractures in the crop area, including notches on leaf edges and fractures on thin stems caused by corrosion, to restore the complete shape of the crop body;

[0120] S25: extracting multidimensional features from the crop image that has restored the complete morphology of the crop body using the pre-trained neural network model, calling a preset crop species feature library to match the extracted multidimensional features with the crop species using a preset feature vector comparison algorithm, and selecting the crop species with the highest matching degree as the species recognition result;

[0121] S26: Automatically load the growth stage feature library of the corresponding crop based on the species recognition result, obtain the corresponding growth state features in the crop image from the multidimensional features, perform feature matching based on the growth state features and the growth stage feature library of the corresponding crop, calculate the matching score using a weighted scoring method, and use the growth stage with the highest score as the growth stage recognition result.

[0122] The specific process of multi-dimensional feature extraction in step S25 is as follows:

[0123] S251: Different features are extracted respectively through the multi-layer feature extraction layers of the neural network model, low-order features including color features and edge features are extracted through the shallow convolution layer, pixel distribution features of the RGB channels are extracted through the 1st-2nd convolution layer, a color histogram vector is generated, and the color saturation and hue of the crop leaves and fruits are quantified. The Sobel operator is combined with the feature map output by the convolution layer to extract the edge contour features of the crop body, including the serrated edges of the leaves and the circular contours of the fruits, and an edge direction histogram is generated.

[0124] S252: The middle convolutional layer extracts morphological and texture features. The third and fourth convolutional layers and pooling layers extract the geometric morphological features of the crop, including: Leaf features: number of leaves, aspect ratio, and tip shape. A 128-dimensional feature vector is generated using a contour fitting algorithm. Fruit features: number of fruits, diameter, and roundness. A 64-dimensional feature vector is generated using a minimum circumscribed circle calculation. Plant type features: number of stems and branch angles. A 64-dimensional feature vector is generated using a skeleton extraction algorithm. Plant type features: number of stems and branch angles. A 64-dimensional feature vector is generated using a skeleton extraction algorithm. The middle-layer feature map is processed using a gray-level co-occurrence matrix (GLCM) to extract texture features of the leaf surface, including smoothness / wrinkles and vein density. A 16-dimensional feature vector is generated, including contrast and entropy.

[0125] S253: Through deep convolutional layers, features including growth status and semantic association features are extracted. Through deep convolutional layers, the overall growth status characteristics of crops can be extracted for the bottleneck structure of the ResNet network model, including leaf stretch (calculated based on the contour convex hull ratio) and fruit maturity (combined with a comprehensive evaluation of color and texture), generating a 256-dimensional feature vector. Using the self-attention mechanism, the Transformer module can capture the association between various parts of the crop, including the spatial position of leaves and fruits, and the supporting structure of stems and branches, generating a 128-dimensional association feature vector.

[0126] S254: The feature fusion layer uses feature splicing to merge low-order, mid-order, and high-order feature vectors into a fused feature vector of the crop image.

[0127] See also Figure 3 The neural network model consists of an input layer, intermediate convolutional layers, and deep convolutional layers. The input layer feeds a preprocessed tomato crop image (2048×1536 pixels, RGB channels). The preprocessed tomato crop image, which has been resized, grayscaled, and denoised, serves as the raw input for feature extraction.

[0128] Low-level features are first extracted through shallow convolutional layers. The network structure consists of two convolutional layers (Conv1 and Conv2) and one pooling layer (Pool1). Conv1 has 32 3×3 convolution kernels with a stride of 1 and a ReLU activation function to extract basic color features. Conv2 has 64 3×3 convolution kernels with a stride of 1 and a ReLU activation function to enhance edge features. Pool1 has a size of 2×2 and performs max pooling to preserve features while reducing dimensionality.

[0129] Feature extraction logic: Color feature extraction (based on Conv1 output) generates RGB three-channel pixel distribution features, which are quantized using color histograms to form a 384-dimensional vector (3 channels x 128 dimensions). Output: Quantized features of the dark green tomato leaf (saturation 0.7-0.8) and the light green tomato fruit (hue 100-120°).

[0130] Edge feature extraction (based on Conv2 output): Combines the Sobel operator to extract edge contours and generates an 8-directional edge direction histogram. Output: Leaf serrated edge features (direction angles 30° and 150°) and fruit rounded outline features (uniformly distributed in all directions).

[0131] Mid-level features are then extracted through the mid-level convolutional layers. The network structure consists of two convolutional layers (Conv3 and Conv4) and one pooling layer (Pool2). Conv3 includes 128 3×3 convolution kernels with a stride of 1 and a ReLU activation function (to extract basic morphological features). Conv4 includes 256 3×3 convolution kernels with a stride of 1 and a ReLU activation function (to enhance texture features). Pool2 uses a 2×2 max pooling to retain key features.

[0132] Feature extraction logic: Morphological feature extraction (based on Conv3+Pool2 output). Leaf features: Generate a 128-dimensional vector using a contour fitting algorithm (number of leaves: 3-5, aspect ratio: 1.5-2.0). Fruit features: Generate a 64-dimensional vector using a minimum circumscribed circle calculation (number of fruits: 2-4, diameter: 2-3 cm). Plant type features: Generate a 64-dimensional vector using a skeleton extraction algorithm (number of stems: 1-2, branching angle: 45°-60°). Texture feature extraction (based on Conv4 output): Use a gray-level co-occurrence matrix to process the feature map, generating a 16-dimensional vector (contrast: 0.3-0.5, entropy: 0.6-0.8). Output: Leaf surface smoothness and vein density characteristics.

[0133] High-level features are extracted through deep convolutional layers. The network architecture consists of a ResNet bottleneck structure (three residual blocks) followed by a Transformer attention module. Residual blocks 1-3 each contain a 1×1 convolution for dimensionality reduction, a 3×3 convolution for feature enhancement, and a 1×1 convolution for dimensionality increase. The Transformer module employs a four-head self-attention mechanism to capture feature correlations.

[0134] Feature extraction logic: Growth state features are extracted based on residual blocks. Leaf spread: Calculated using the convex hull ratio of the outline, convex hull area / actual area = 0.8-0.9. Fruit maturity: A 256-dimensional vector is generated by combining color and texture. Semantic association feature extraction is implemented using a Transformer-based module. Leaf and fruit spatial position: Leaf coverage of less than 30% of the fruit area. Stem and branch support structure: The angle between branches and the main stem is 45°-60°. Output: A 128-dimensional association feature vector.

[0135] The feature fusion layer uses feature concatenation to integrate the outputs of each layer. The low-order feature vector is: 384 (color) + 256 (edge) = 640 dimensions. The mid-order feature vector is: 128 (leaf) + 64 (fruit) + 64 (plant type) + 16 (texture) = 272 dimensions. The high-order feature vector is: 256 (growth state) + 128 (semantic association) = 384 dimensions. The final output is: a 640+272+384=1296-dimensional fused feature vector of the crop image, which is used to identify tomato species and growth stages.

[0136] In step S3, based on the crop type, growth stage, and current field soil water and fertilizer data, the specific process of calling the basic water and fertilizer parameter library of the corresponding crop to generate the initial water and fertilizer control plan for the current time node is as follows:

[0137] S31: Performing a primary index based on the crop type and locating the water and fertilizer parameter unit corresponding to the crop type;

[0138] S32: Perform secondary indexing based on the growth stage, locate the corresponding growth stage sub-library under the water and fertilizer parameter unit corresponding to the crop type, and extract basic water and fertilizer parameters;

[0139] S33: Setting an initial water and fertilizer control plan based on the current field soil water and fertilizer data and the basic water and fertilizer parameters:

[0140] Comparing the current field soil water and fertilizer data with the basic water and fertilizer parameters one by one, calculating the difference between each water and fertilizer data, and calculating the actual supply based on the difference between each water and fertilizer data;

[0141] The execution parameters of the initial water and fertilizer control plan are generated based on the actual supply, including irrigation parameters and fertilization parameters. The irrigation parameters include single irrigation amount, single irrigation duration, and irrigation time nodes. The fertilization parameters include single fertilizer type, single fertilizer type ratio, and fertilization time nodes.

[0142] Specifically, the system first receives three core data types: the crop type is tomato, the growth stage is the early fruiting stage, the current field soil water and fertilizer data is soil moisture of 25%, nitrogen content of 45mg / kg, phosphorus content of 20mg / kg, and potassium content of 60mg / kg, and the basic crop growth parameters are the early fruiting growth characteristics of tomatoes obtained from big data. A data mapping relationship is established: the growth stage is associated with the timeline, and the early fruiting stage of tomatoes corresponds to the 30th-40th day after planting, providing a benchmark for the time parameters of the plan.

[0143] Perform primary indexing based on crop types and locate the "Tomato Exclusive Parameter Library" by crop type, including cherry tomatoes and large-fruit tomatoes.

[0144] Perform secondary indexing based on growth stage, locate the early fruiting sub-database by growth stage, and extract basic water and fertilizer parameters: target soil conditions: 30%-35% moisture, 40-50mg / kg nitrogen, 25-30mg / kg phosphorus, and 70-80mg / kg potassium. Basic supply parameters: single irrigation volume 8-10L / m², NPK ratio 3:2:5, and irrigation cycle 2-3 days.

[0145] Dynamic parameter adaptation is performed, with fine-tuning based on the characteristics of the variety. For cherry tomatoes, the target potassium value is increased by 10% to enhance fruit sweetness. A regional correction factor is introduced: Based on the growing region, the phosphorus absorption efficiency parameter is adjusted for alkaline soils in northern China, with a correction factor of 0.85.

[0146] Calculation of initial water and fertilizer control plan:

[0147] Calculate the supply and demand difference and compare the current soil water and fertilizer status with the target value:

[0148] Humidity difference: 30%-25%=5% (need to add water);

[0149] Nitrogen content: 45mg / kg is within the target range (no additional nitrogen supplementation is required);

[0150] Phosphorus content difference: 25mg / kg-20mg / kg=5mg / kg (phosphorus supplementation is required);

[0151] Potassium content difference: 70mg / kg-60mg / kg=10mg / kg (potassium supplementation is required).

[0152] Converted to actual supply:

[0153] Irrigation amount: Based on the soil bulk density (1.3g / cm³), the required watering amount is 5% × the thickness of the tillage layer (20cm) area = 1.3L / m²;

[0154] Fertilizer dosage: Phosphorus 5mg / kg × soil volume × absorption rate (60%) = 3.6g / m², potassium 10mg / kg × corresponding coefficient = 8.2g / m².

[0155] Execution parameter generation:

[0156] Irrigation parameters: Single irrigation volume 1.3L / m² (added with basic evaporation compensation of 0.5L / m², actual volume 1.8L / m²), irrigation duration 8 minutes (based on a drip irrigation system flow rate of 2L / h / m²);

[0157] Fertilization parameters: phosphorus and potassium fertilizer ratio 2:5, concentration controlled by fertilization device, phosphorus concentration 0.3%, potassium concentration 0.75%;

[0158] Time planning: The first execution time is the current node, and the next execution interval is 2 days.

[0159] Example 3

[0160] Based on Example 1 or Example 2, the specific process of step S4 is as follows:

[0161] S41: Preprocess historical soil water, fertilizer, and meteorological data to identify and replace abnormal data. The 3σ principle is used to identify abnormal data. For example, a sudden 100% increase in soil moisture is considered a sensor failure. Missing values ​​are filled using the mean of three consecutive days. Missing soil nutrient data, such as missing samples for three days, are filled based on meteorological correlation. For example, when rainfall exceeds 50 mm, nitrogen content is estimated based on a 15% decrease over the same period in history.

[0162] S42: Construct training and test sets based on the pre-processed historical soil water and fertilizer data and historical meteorological data in a specified ratio and input them into the soil water and fertilizer prediction model in sequence. Figure 2 The soil water and fertilizer prediction model consists of an input layer (128-dimensional features) → a hidden layer (two LSTM layers, each containing 64 and 32 neurons, respectively) → an output layer (four-dimensional, corresponding to moisture and nitrogen, phosphorus, and potassium). A temporal attention module is added to the hidden layer to automatically increase the weight of recent data (for example, data from the past seven days is twice as weighted as data from one month ago). Model training parameters include a batch size of 32, 50 iterations, the Adam optimizer, and an initial learning rate of 0.001. An early stopping strategy halts training when the validation set loss fails to decrease for five consecutive epochs to prevent overfitting. Hyperparameter tuning involves using a grid search to determine the optimal number of LSTM neurons (64 → 32) and time window length (30 days).

[0163] S43: The LSTM in the soil water and fertilizer prediction model is used to extract the time series features of soil water and fertilizer data. The random forest algorithm is used as a feature importance assessment tool to screen meteorological factors that have a significant impact on water and fertilizer changes. For example, rainfall weights greater than 30% are prioritized for inclusion in the model.

[0164] S44: Optimize the hyperparameters of the soil water and fertilizer prediction model by determining the optimal number of LSTM neurons and time window length through grid search;

[0165] S45: Acquire soil water and fertilizer data at the prediction start time node and weather forecast data within the specified time period, and predict the soil water and fertilizer change trend within the specified time period based on the soil water and fertilizer data at the prediction start time node and weather forecast data within the specified time period through a soil water and fertilizer prediction model.

[0166] Executed via rolling forecast:

[0167] Step 1: Predict the water and fertilizer values ​​for Day 1 based on the data for Day 0;

[0168] Step 2: Take the Day 1 forecast as input, superimpose the Day 1 weather forecast, and predict Day 2;

[0169] Repeat until the 7-day forecast is completed and a continuous change curve is generated;

[0170] Confidence interval calculation: A confidence interval of ±5% is added to each predicted value based on the error distribution of the test set. For example, if the humidity on Day 3 is predicted to be 30%, the interval is labeled 30% ± 1.5%.

[0171] Generate multi-dimensional forecast report:

[0172] Trend chart: 7-day variation curve of soil moisture and nitrogen, phosphorus and potassium content, with key nodes marked. For example, due to no rainfall on Day 5, the potassium content is predicted to decrease.

[0173] Influencing factor analysis: Attached is a table of meteorological factor contributions, such as rainfall's influence on humidity accounts for 42%;

[0174] Data interface: Convert the prediction results into JSON format, synchronize them to the cloud server, and call the water and fertilizer plan adjustment module.

[0175] Model dynamic update:

[0176] Regular iteration: Incorporate new historical data every 30 days, retrain the model, and update weight parameters;

[0177] Error feedback: When the deviation between the actual value and the predicted value is greater than 10%, such as the predicted humidity is 30% and the actual humidity is 25%, an emergency update is triggered to increase the weight of the training samples corresponding to the meteorological-soil combination scenario.

[0178] The specific process of step S5 is as follows:

[0179] S51: Perform hysteresis data modeling based on the predicted soil water and fertilizer change trend, and establish a hysteresis coefficient matrix based on the hysteresis data. The hysteresis coefficient matrix is ​​divided into dimensions according to "soil type × crop type × water and fertilizer type", and the hysteresis coefficients of the corresponding water and fertilizer indicators are set after each dimension.

[0180] S52: Based on the predicted trend of soil water and fertilizer changes, the change curves of various water and fertilizer indicators are drawn, and the change curves of various water and fertilizer indicators are time-shifted and amplitude-corrected based on the hysteresis coefficients of each water and fertilizer indicator. For example, after clay irrigation, the Day1 humidity forecast value is linked to the actual Day1+6 hours to achieve the time shift of the curve; the theoretical increase of potassium fertilizer is 10mg / kg, and the actual correction is 8mg / kg, and the remaining 2mg / kg is delayed until the next day to achieve the amplitude correction.

[0181] S53: Extract the lag-corrected forecast data and, based on the forecast curve of the original specified time period, mark the lag impact interval: the humidity change lag period is within 6 hours after irrigation, and the nutrient absorption lag period is 3 days after potassium fertilizer application.

[0182] S54: Retrieve the hysteresis-related parameters of the initial plan: This determines the theoretical time difference between "executed action - expected effect" in the initial plan. For example, if the initial irrigation plan is executed at 3:00 PM and the humidity is expected to reach 30% by 5:00 PM, the hysteresis-corrected predicted value for 5:00 PM must be matched.

[0183] S55: Adjust the impact weight: Increasing the trend weight within the lag period by 20% can increase the weight of the moisture prediction value 6 hours after irrigation from 60% to 72%.

[0184] It is also necessary to identify lag risk points: if it is predicted that the soil water and fertilizer will exceed the safe range during the lag period, for example, due to clay lag after irrigation, the humidity will drop to the appropriate value after 24 hours, and there may be short-term waterlogging during this period, and buffer space needs to be reserved during adjustment.

[0185] S56: Adjust the initial water and fertilizer control plan at the current time node.

[0186] Dynamic correction of irrigation amount:

[0187] For example, assuming an initial irrigation rate of 1.8 L / m², clay soil (6-hour lag), a predicted rainfall of 8 mm on Day 1 (with a 0.4 reduction factor), and a predicted humidity of 32% on Day 2 (above the target of 30%);

[0188] Adjusted irrigation amount = 1.8 × (1-0.4) × (1-(32%-30%) / 30%) × (1-0.25)

[0189] =1.8×0.6×0.93×0.75=0.753L / m², rounded to one decimal place is 0.8L / m².

[0190] Adjustment of fertilizer dosage:

[0191] For example, the initial potassium fertilizer dosage is 8.2 g / m², for tomatoes (potassium fertilizer lags behind by 3 days), and the predicted potassium loss rate on Day 3 is 12% (compensation coefficient 0.144). The adjusted potassium fertilizer dosage = 8.2×(1+0.144)×(1+0.43)=8.2×1.144×1.43≈13.4 g / m²).

[0192] The specific process of step S6 is as follows:

[0193] S61: Based on agricultural databases (such as national crop germplasm banks, regional planting archives) and historical planting records, filter data by crop type (such as tomato) and extract the characteristics of a specified growth cycle:

[0194] Stage division: germination period (1-7 days), seedling period (8-21 days), flowering period (22-35 days), fruiting period (36-60 days), maturity period (61-75 days);

[0195] Key indicators during this stage: peak fertilizer demand during the fruiting period (potassium demand increases by 80% compared to the seedling stage), water-sensitive period (soil moisture needs to be stable at 28%-32% during the flowering period);

[0196] Environmental response rules: If the average daily temperature during the fruiting period is greater than 28°C, transpiration will increase by 30%, and the irrigation frequency needs to be increased;

[0197] Specify growth cycle characteristics for structuring: Convert the specified growth cycle characteristics into calculable parameters (e.g. "10th day of fruiting period" corresponds to the 45th day after planting, "potassium requirement threshold" is set to 80mg / kg), and generate a "crop type-growth cycle characteristic table"

[0198] S62: Specify growth cycle characteristics for regional adaptation (modify characteristic parameters based on the planting area (e.g., the northern temperate zone) (e.g., extend the tomato ripening period from 75 days to 85 days due to lower accumulated temperature)), variety adaptation (for early-maturing tomato varieties, compress the duration of each stage by 20% (e.g., shorten the fruiting period from 25 days to 20 days)), and establish a localized growth cycle characteristic library;

[0199] S63: Plan designated time nodes, including key growth nodes and regular nodes. Set corresponding dynamic water and fertilizer control plans for each designated time node, using growth cycle characteristics as the demand benchmark: for example, 80mg / kg of potassium is required 40 days after planting (mid-fruiting period), and set the target value. Use soil water and fertilizer change trends as the basis for supply: Based on the forecast curve for the next 10 days, if the current potassium is 65mg / kg and natural loss is 2mg / kg per day, calculate the supply and demand gap and establish the basic framework of the dynamic water and fertilizer control plan.

[0200] S64: Calculate dynamic parameters of the basic framework of the dynamic water and fertilizer control scheme, and establish the dynamic water and fertilizer control scheme at the specified time node based on the dynamic parameters.

[0201] For example, if three days of rain are predicted during the middle of fruiting, potassium loss will accelerate to 3 mg / kg per day, and the target value will increase by 10% to 88 mg / kg. Water and fertilizer control plan: Single potassium fertilizer application rate = (88 - 65 + 3 × 10) × 1.2 (hysteresis compensation factor) = (23 + 30) × 1.2 = 63.6 g / m² (applied in two applications, three days apart).

[0202] Example 4

[0203] Based on Example 1 or Example 3, in fact, in the actual process of intelligent field water and fertilizer control, meteorological factors will affect the soil moisture in water and fertilizer control. The impact of meteorology on soil moisture is mainly reflected in evaporation consumption and rainfall replenishment. The formula needs to strip out the nutrient factors and calculate the irrigation amount only based on meteorological data and soil moisture status. Therefore, the decoupling control formula of meteorology on soil moisture is set as follows:

[0204] I=[( S T - S) ×Kᵥ]×K b -E0×K e +R0×K r ×K i .

[0205] I is the irrigation volume, unit is L / m², S T is the target humidity, unit is %, S is the current humidity, unit is %, Kᵥ is the soil volume moisture conversion coefficient, unit is L / m²・%, K b is the soil water retention coefficient, E0 is the average daily evaporation in the past 30 days, in mm / d, K e is the evaporation correction coefficient for the day, R0 is the average daily rainfall in the past 30 days, in mm / d, K r is the rainfall correction coefficient for the day, K i is the rainfall infiltration rate.

[0206] The above formula does not contain any nutrient-related parameters. It is calculated only through specified meteorological parameters and soil moisture parameters, completely eliminating the interference of nutrients on irrigation. For example, even if the soil phosphorus content is insufficient, it will not affect the calculation of irrigation amount.

[0207] Since the impact of meteorology on calcium, phosphorus, and potassium is mainly reflected in rainfall leaching loss (calcium / easy to leach) and temperature-affected absorption (phosphorus absorption is sensitive to temperature), the calculation formula for calcium, phosphorus, and potassium needs to remove the influence of water irrigation and calculate the fertilizer amount based only on meteorological data and nutrient status. This example uses phosphate fertilizer as an example to illustrate the calculation of phosphate fertilizer application rate, and calcium and potassium can be applied accordingly. The calculation formula for phosphate fertilizer application rate is as follows:

[0208] F p =[(P T -P)×K s ]×K w + R0×K p ×K c ×K pH ;

[0209] F p is the amount of phosphate fertilizer applied, in g / m², P T is the target phosphorus content, K s is the nutrient-fertilization conversion coefficient, K w is the temperature absorption coefficient, R0 is the average daily rainfall in the past 30 days, in mm / d, K p is the phosphorus leaching coefficient K p , K c is the soil clay content correction coefficient, K pH is the soil P correction factor.

[0210] There are no parameters related to irrigation amount in the formula. Even if the irrigation amount is adjusted, such as increasing irrigation, the amount of fertilizer applied will remain unchanged as long as the phosphorus content, climate, and pH remain unchanged. At the same time, the meteorological impact is quantified separately through the temperature absorption coefficient and leaching coefficient, achieving complete decoupling from water control.

[0211] In summary, the IoT-based intelligent field water and fertilizer control method provided by this invention uses image acquisition to identify crop types and growth stages, and combines soil water and fertilizer data to generate an adaptive water and fertilizer plan. It automatically adjusts potash fertilizer dosage and irrigation frequency based on the differentiated needs of crops at different growth stages, improving the matching of water and fertilizer supply with crop needs. Using a soil water and fertilizer prediction model, it combines historical data and weather forecasts to predict future soil conditions. The prediction results are then corrected using a hysteresis coefficient matrix, overcoming the limitations of existing technologies that only passively respond to real-time data.

[0212] By comparing crop growth monitoring with soil data, a closed loop of plan execution, performance evaluation, and model optimization is formed. When crop growth deviates from expectations, the system can quickly identify whether the problem lies with the plan or the prediction model, and targeted corrections can be made. If actual soil water and fertilizer deviate significantly from the prediction, model parameters are automatically adjusted. If crop growth remains poor after the plan is implemented, irrigation and fertilization parameters are corrected, shortening the iterative optimization cycle for water and fertilizer plans and improving crop yield stability. The entire process, from crop identification and plan generation to execution and control, is fully automated. Crop features are extracted through semantic segmentation and neural networks, replacing manual judgment of crop growth stages. A dynamic plan generation module automatically schedules fertilization and irrigation schedules, replacing manual record-keeping and decision-making while minimizing the risk of manual misjudgment. Growth cycle characteristics are corrected through regional and variety adaptation. Combined with distributed sensors and multi-protocol communication, the system can adapt to different soil types, crop varieties, and growing regions, achieving precise water and fertilizer control for each region and crop.

Claims

1. The intelligent control method of field water and fertilizer based on the Internet of Things is characterized by: The following steps are involved: S1: Capture images of crops in the controlled field, collect specified water and fertilizer data of the current field soil, obtain meteorological forecast data based on big data, and transmit all data to the cloud server; S2: The cloud server identifies crop species based on the collected crop images and identifies the growth stage based on the crop species; S3: Based on the crop type, growth stage and current field soil water and fertilizer data, the basic water and fertilizer parameter library of the corresponding crop is called to generate the initial water and fertilizer control plan for the current time node; S4: Create a soil water and fertilizer prediction model and predict the soil water and fertilizer change trend within a specified time period using historical soil water and fertilizer data and historical meteorological data; S5: adjusting the initial water and fertilizer control plan at the current time node according to the predicted soil water and fertilizer change trend to obtain an adjusted water and fertilizer control plan at the current time node; S6: Based on big data, the growth cycle characteristics of the crop species are obtained, and a dynamic water and fertilizer control plan is formulated at a specified time node based on the growth cycle characteristics and soil water and fertilizer change trends. The water and fertilizer control plan is then adjusted at the current time node, and the dynamic water and fertilizer control plan is executed at the specified time node. It also includes real-time crop growth monitoring and feedback processes, as follows: S7: Collecting actual crop growth data at a specified time point, comparing the actual crop growth data with the growth cycle characteristics of the crop type obtained based on big data, and determining whether the specified indicators in the actual crop growth data meet the preset growth requirements. If so, executing the dynamic water and fertilizer control plan at the specified time point according to step S6; if not, executing step S8; S8: Collecting actual soil water and fertilizer data at a specified time point, comparing the actual soil water and fertilizer data with the soil water and fertilizer data at the corresponding time point in the predicted soil water and fertilizer change trend, and determining whether the actual soil water and fertilizer data and the predicted soil water and fertilizer data at the corresponding time point are both within a preset error range. If so, revising the initial water and fertilizer control plan; if not, executing step S9; S9: Compare the actual soil water and fertilizer data with the various indicators in the predicted soil water and fertilizer data at the corresponding time node one by one, obtain the indicators whose differences exceed the corresponding preset error range, and adjust the parameters of the soil water and fertilizer prediction model based on the differences of the indicators that exceed the corresponding preset error range.

2. The field water and fertilizer intelligent control method based on the Internet of Things according to claim 1 is characterized in that: The detailed process of revising the initial water and fertilizer control plan in step S8 is as follows: S81: extracting specified parameters from the initial water and fertilizer control plan, including irrigation water volume, irrigation time node, fertilizer type, fertilizer volume, and fertilizer time node; S82: setting correction coefficients for various designated parameters in the initial water and fertilizer control plan based on actual soil water and fertilizer data and the needs of the crop at the corresponding growth stage; S83: Modify the initial water and fertilizer control plan based on the correction coefficient, re-execute steps S4-S6, and synchronously update to the cloud server.

3. The field water and fertilizer intelligent control method based on the Internet of Things according to claim 1 is characterized in that: The specific process of step S2 is as follows: S21: perform uniform resizing, grayscale processing and denoising on the crop images; S22: Identify the crop body and background from the crop image using a preset semantic segmentation model, and retain the initial valid areas of the crop leaves, stems, and fruits in the image; S23: Binarizing the initial effective area, selecting a structural element that is suitable for crop characteristics, traversing each pixel in the crop image, marking the pixel as foreground if the pixel and all pixels in its 8-neighborhood are foreground, otherwise marking it as background, and eliminating isolated noise points whose area is smaller than a specified number of pixels; S24: traverse each pixel in the crop image from which isolated noise has been eliminated, mark the pixel and any pixel in its 8-neighborhood that has at least one foreground pixel as foreground, repair minor breaks in the crop area, and restore the complete shape of the crop body; S25: extracting multidimensional features from the crop image that has restored the complete morphology of the crop body using the pre-trained neural network model, calling a preset crop species feature library to match the extracted multidimensional features with the crop species using a preset feature vector comparison algorithm, and selecting the crop species with the highest matching degree as the species recognition result; S26: Loading the growth stage feature library of the corresponding crop according to the species recognition result, obtaining the corresponding growth state features in the crop image from the multidimensional features, and performing feature matching based on the growth state features and the growth stage feature library of the corresponding crop.

4. The field water and fertilizer intelligent control method based on the Internet of Things according to claim 3 is characterized in that: The specific process of multi-dimensional feature extraction in step S25 is as follows: S251: extracting different features respectively through the multi-layer feature extraction layer of the neural network model, and extracting low-order features including color features and edge features through the shallow convolution layer; S252: extracting morphological and texture features through the middle convolutional layer; S253: Extract growth status features and semantic association features through deep convolutional layers; S254: low-order, mid-order, and high-order feature vectors are merged into a fused feature vector of the crop image by using feature splicing.

5. The field water and fertilizer intelligent control method based on the Internet of Things according to claim 1 is characterized in that: The specific process of step S3 is as follows: S31: Performing a primary index based on the crop type and locating the water and fertilizer parameter unit corresponding to the crop type; S32: Perform secondary indexing based on the growth stage, locate the corresponding growth stage sub-library under the water and fertilizer parameter unit corresponding to the crop type, and extract basic water and fertilizer parameters; S33: Setting an initial water and fertilizer control plan based on the current field soil water and fertilizer data and the basic water and fertilizer parameters: The current field soil water and fertilizer data are compared with the basic water and fertilizer parameters one by one, the difference between each water and fertilizer data is calculated, the actual supply is calculated based on the difference between each water and fertilizer data, and the execution parameters of the initial water and fertilizer control plan are generated based on the actual supply, including irrigation parameters and fertilization parameters.

6. The field water and fertilizer intelligent control method based on the Internet of Things according to claim 1 is characterized in that: The specific process of step S4 is as follows: S41: Preprocess historical soil water and fertilizer data and historical meteorological data, identify and replace abnormal data, and fill in missing values; S42: constructing a training set and a test set based on the pre-processed historical soil water and fertilizer data and historical meteorological data in a specified ratio and inputting them into the soil water and fertilizer prediction model in sequence; S43: The LSTM in the soil water and fertilizer prediction model is used to extract the time series features of soil water and fertilizer data. The random forest algorithm is used as a feature importance assessment tool to screen meteorological factors that significantly affect water and fertilizer changes. S44: Optimize the hyperparameters of the soil water and fertilizer prediction model by determining the optimal number of LSTM neurons and time window length through grid search; S45: Acquire soil water and fertilizer data at the prediction start time node and weather forecast data within the specified time period, and predict the soil water and fertilizer change trend within the specified time period based on the soil water and fertilizer data at the prediction start time node and weather forecast data within the specified time period through a soil water and fertilizer prediction model.

7. The field water and fertilizer intelligent control method based on the Internet of Things according to claim 1 is characterized in that: The specific process of step S5 is as follows: S51: Based on the predicted soil water and fertilizer change trends, hysteresis data modeling is performed. A hysteresis coefficient matrix is ​​established based on the hysteresis data. The hysteresis coefficient matrix is ​​divided into dimensions according to "soil type × crop type × water and fertilizer type". The hysteresis coefficients of the corresponding water and fertilizer indicators are set after each dimension. S52: based on the predicted soil water and fertilizer change trends, drawing change curves of various water and fertilizer indicators, and performing time offset and amplitude correction on the change curves of various water and fertilizer indicators based on the hysteresis coefficients of the various water and fertilizer indicators; S53: extracting the forecast data after lag correction, and marking the lag impact interval based on the forecast curve of the original specified time period; S54: Retrieve the hysteresis-related parameters of the initial plan: clarify the theoretical time difference between "execution action - expected effect" in the initial water and fertilizer control plan; S55: Adjust impact weights and identify lagging risk points; S56: Adjust the initial water and fertilizer control plan at the current time node, dynamically correct the irrigation amount, and adjust the fertilizer amount.

8. The field water and fertilizer intelligent control method based on the Internet of Things according to claim 1 is characterized in that: The specific process of step S6 is as follows: S61: Based on the agricultural database and historical planting records, filter data by crop type and extract the characteristics of the specified growth cycle: Stage division: germination stage, seedling stage, flowering stage, fruiting stage, and maturity stage; Key indicators at this stage: peak fertilizer demand during the fruiting period and water-sensitive period; Specify growth cycle characteristics for structuring: convert the specified growth cycle characteristics into computable parameters and generate a crop type-growth cycle characteristic table; S62: performing regional adaptation and modification of growth characteristic parameters and variety adaptation for the designated growth cycle characteristics, and establishing a local exclusive growth cycle characteristic database; S63: Plan designated time nodes, including key growth nodes and regular nodes, and set corresponding dynamic water and fertilizer control plans for each designated time node. Based on the growth cycle characteristics as the demand benchmark and the soil water and fertilizer change trends as the supply basis, calculate the supply and demand gap and establish the basic framework of the dynamic water and fertilizer control plan; S64: Calculate dynamic parameters of the basic framework of the dynamic water and fertilizer control scheme, and establish the dynamic water and fertilizer control scheme at the specified time node based on the dynamic parameters.

Images