Crop growth prediction system and method based on agricultural unmanned aerial vehicle remote sensing technology

Abstract

The invention discloses a crop growth prediction system and method based on an agricultural unmanned aerial vehicle remote sensing technology, and relates to the technical field of agricultural information and intelligent monitoring, the system uses an unmanned aerial vehicle remote sensing module and a ground Internet of Things sensor module to collect multi-source data of banana planting, fish pond breeding and hog house breeding, and the multi-source data is used as a data expression recording carrier. The edge computing node module preprocesses data, and the cloud intelligent platform module is internally provided with a multi-modal data fusion model, a growth prediction model and the like as processing record carriers to realize data depth analysis and prediction. And realizing automatic control of the production process according to the data through block chain evidence storage and the intelligent contract. And constructing a disease prevention and control model by using the hyperspectral image of the unmanned aerial vehicle and related data. According to the invention, the collection, processing and application capabilities of agricultural production data are improved, accurate prediction and intelligent regulation are realized, the resource utilization efficiency is improved, the disease loss is reduced, and the agricultural intelligent development is promoted.

Description

Technical Field

[0001] The present invention relates to the field of agricultural information and intelligent monitoring technology, and specifically to a crop growth prediction system and method based on agricultural unmanned aerial vehicle remote sensing technology. Background Art

[0002] In agricultural production, accurately understanding crop growth trends and adjusting production strategies in a timely manner are crucial for improving yield and quality. However, traditional agricultural production methods rely primarily on manual experience and lack accurate, real-time monitoring of the crop's growth environment and its own status. Taking the integrated banana-fish pond-pigsty farming system as an example, on the one hand, it is difficult to comprehensively obtain multi-source data on the banana planting, fish pond farming, and pigsty farming environments, such as banana canopy information, fish pond water quality parameters, and pigsty air quality. These limitations in data acquisition make it difficult to gain a deep understanding of the operating status of the agricultural production system. On the other hand, existing data recording and processing methods are outdated, with data being scattered and lacking effective integration. This makes it impossible to form comprehensive and systematic agricultural production information, making it difficult to support accurate decision-making. This leads to agricultural production facing problems such as low resource utilization efficiency, untimely pest and disease control, and difficulty in improving economic benefits. Innovative data expression, recording, and processing technologies are urgently needed to improve the current situation.

[0003] In view of this, this application is hereby filed. Summary of the Invention

[0004] The purpose of the present invention is to provide a crop growth prediction system and method based on agricultural drone remote sensing technology to solve the problems raised in the above background technology.

[0005] To solve the above technical problems, the present invention provides a crop growth prediction system based on agricultural drone remote sensing technology, comprising: UAV remote sensing module: equipped with a multispectral camera, thermal infrared sensor, and lidar, used to obtain multispectral images of banana canopies, thermal infrared temperature distribution of fish ponds, and 3D point cloud data of piggery areas; Ground IoT sensor modules: These include ammonia / methane sensors deployed in pig houses, pH / EC probes and algae density sensors in fish ponds, and soil moisture sensors in banana fields, used to collect real-time parameters of aquaculture and planting environments. Edge computing node module: communicates with the UAV remote sensing module and the ground IoT sensor module to pre-process real-time data, including image noise reduction, outlier detection, and feature extraction; Cloud intelligent platform module: communicates with the edge computing node module, and has a built-in multimodal data fusion model, growth prediction model and circulatory system optimization algorithm for in-depth analysis and prediction of pre-processed data; User terminal module: communicates with the cloud intelligent platform module and is used to visualize the prediction results and receive control instructions; The above system architecture enables efficient collection and integration of multi-source data, providing a comprehensive data foundation for accurate forecasting. The drone remote sensing module rapidly acquires information on large-scale crops and breeding areas, while ground-based IoT sensors accurately collect local environmental parameters. The edge computing node module reduces data transmission pressure and performs preliminary data processing. The cloud-based intelligent platform module conducts in-depth analysis, and the user terminal module facilitates intuitive understanding and intervention in production. This system significantly improves the efficiency of agricultural production data acquisition and processing, providing strong support for optimizing agricultural production decisions.

[0006] Furthermore, the multimodal data fusion model includes: a spatiotemporal data cube construction unit, which is used to fuse drone remote sensing data, ground sensor data, and meteorological data according to a three-dimensional "field-time-indicator" structure to generate a spatiotemporal dataset containing banana leaf area index, fish pond algae density, and pig house environmental parameters; a Bayesian network association unit, which is used to establish a spatiotemporal correlation model between pig manure discharge, nutrient content, banana growth indicators, and fish pond water quality parameters to optimize pig manure input strategies; the multimodal data fusion model can explore potential connections between different data. The spatiotemporal data cube construction unit integrates multi-source data, allowing the correlation of data in time and space dimensions to be reflected, facilitating a comprehensive understanding of the dynamic changes in agricultural production. By establishing correlations between pig manure and crop and breeding parameters, the Bayesian network association unit can accurately adjust pig manure input according to actual conditions, improving resource utilization efficiency and reducing resource waste, while ensuring the stability of crop growth and the breeding environment.

[0007] Furthermore, the cloud-based intelligent platform module also includes a blockchain evidence storage unit, which uses consortium blockchain technology to encrypt and store records of pig manure treatment, fish pond fertilization schedules, and banana fertilization data, ensuring that the data cannot be tampered with. A smart contract execution unit, based on the data in the blockchain evidence storage unit, automatically triggers instructions such as pig manure input, pond mud return to the field, and aerator start and stop. The blockchain evidence storage unit provides a reliable storage method for agricultural production data. Its tamper-proof nature ensures the authenticity and credibility of the data, facilitates agricultural product traceability, and enhances consumer trust in agricultural product quality. The smart contract execution unit automatically executes relevant instructions based on the stored evidence data, automating the agricultural production process, reducing manual intervention, improving production efficiency, and ensuring the accuracy and timeliness of operations.

[0008] Furthermore, the growth prediction model includes an LSTM neural network, which analyzes spatiotemporal series data of algae density and combines it with water temperature and light intensity to predict the risk of cyanobacteria outbreaks. A Transformer hybrid model, which integrates multispectral drone imagery and ground sensor data to predict banana yield per plant and fish growth rate, can accurately predict key indicators in agricultural production. By analyzing algae density and environmental factors, the LSTM neural network provides early warning of cyanobacteria outbreak risks, enabling farmers to take timely measures to reduce the risk of fish mortality due to cyanobacteria outbreaks and safeguard fishery production. The Transformer hybrid model integrates multiple data sets to predict crop yields and fish growth rates, providing a scientific basis for farmers to rationally plan production and estimate profits, thereby helping to improve the economic benefits of agricultural production.

[0009] Furthermore, the edge computing node module includes a YOLOv8 target detection unit for processing drone visible light imagery to identify abnormal targets such as banana leaf spot and floating objects in fish ponds; and a Kalman filter unit for denoising real-time sensor data and calibrating remote sensing inversion parameters based on ground-based data. The built-in units in the edge computing node module enhance the real-time and accuracy of data processing. The YOLOv8 target detection unit can quickly identify abnormalities such as banana leaf spot, enabling farmers to prevent and control diseases in a timely manner and reduce crop losses. The Kalman filter unit denoises sensor data and calibrates remote sensing inversion parameters, improving data quality and, in turn, the accuracy of subsequent data analysis and prediction, providing more reliable data support for agricultural production decisions.

[0010] The prediction method of the crop growth prediction system based on agricultural UAV remote sensing technology includes the following steps: Data collection steps: The UAV remote sensing module acquires multispectral images of banana canopies, thermal infrared data of fish ponds, and 3D point clouds of pig houses. Ground sensors are used to collect pig house environment, fish pond water quality, and soil parameters. Preprocessing step: Use edge computing node modules to reduce noise, extract features, and filter outliers on the collected data to generate a standardized data set; Fusion processing step: The pre-processed multi-source data is input into the spatiotemporal data cube of the cloud-based intelligent platform module, and the association between pig manure treatment data and crop growth and aquaculture parameters is established through the Bayesian network; Model prediction steps: Use LSTM neural networks to predict the risk of algae outbreaks in fish ponds, and use Transformer hybrid models to predict banana yields and fish growth rates; Instruction output step: Generate control instructions based on the prediction results and send them to the execution equipment (such as fertilizer spreader, aerator) through the user terminal module; This prediction method forms a complete closed-loop process of data processing and decision-making. The data collection step comprehensively gathers relevant agricultural production data; the preprocessing step ensures data quality; the fusion processing step discovers data associations; the model prediction step produces accurate predictions; and the instruction output step converts the predictions into actual production control instructions. This interconnected series of steps can effectively optimize the agricultural production process, improve resource utilization efficiency, reduce production costs, and enhance the overall benefits of agricultural production.

[0011] Furthermore, during the data collection step, the drone remote sensing module includes multispectral image processing to identify algae-covered areas in fish ponds based on the NDVI and NDWI indices. This multispectral image processing method accurately captures algae information in fish ponds. Based on the NDVI and NDWI indices, algae-covered areas and boundaries can be clearly identified, providing intuitive data for algae monitoring, enabling a more accurate understanding of algae growth conditions and providing a scientific basis for fish pond ecological environment assessment and fishery production management.

[0012] Furthermore, the fusion processing step also includes: using a blockchain evidence storage unit to store records of pig manure input, fertilizer application time, and pond mud return to the field, creating an unalterable production log; and utilizing a reinforcement learning algorithm (PPO) to dynamically adjust the pig manure fermentation cycle and the frequency of fish pond fertilization and watering to optimize resource recycling efficiency. The application of blockchain evidence storage and reinforcement learning algorithms in this fusion processing step has brought significant benefits. The unalterable production log created by blockchain evidence storage facilitates the traceability and management of agricultural production processes, enhancing the credibility of agricultural product quality. The reinforcement learning algorithm dynamically adjusts the pig manure fermentation cycle and the frequency of fish pond fertilization and watering, optimizing resource recycling based on actual production conditions, improving the utilization efficiency of resources such as pig manure, reducing production costs, and minimizing negative environmental impacts.

[0013] Furthermore, the model prediction step includes banana disease identification: spectral features of banana leaves are extracted using drone hyperspectral imagery and input into a YOLOv8 model to identify early symptoms of leaf spot disease with an accuracy rate of at least 95%. A disease transmission probability model is developed, combining ground spore capture data with meteorological parameters, providing a 72-hour advance warning of disease risk. The banana disease identification method used in this model prediction step is effective in preventing and controlling banana diseases. Using drone hyperspectral imagery and the YOLOv8 model to accurately identify early symptoms of leaf spot disease allows farmers to take timely preventive measures in the early stages of the disease and reduce its spread. The disease transmission probability model, developed by combining ground spore capture data with meteorological parameters, provides early warning of disease risk, giving farmers more time to prevent and control the disease and reducing banana yield losses due to disease.

[0014] Furthermore, in the instruction output step, when the LSTM neural network predicts that the probability of a cyanobacteria outbreak exceeds a threshold, it automatically generates aerator start / stop instructions and probiotic dosage recommendations. When the Transformer model predicts that bananas are nitrogen deficient, it automatically adjusts the pig manure input to an optimal value (with an error of ±5%) based on the Bayesian network calculation results. This instruction output step can promptly translate the prediction results into effective production control measures. When the predicted probability of a cyanobacteria outbreak exceeds a threshold, the automatically generated aerator start / stop instructions and probiotic dosage recommendations can effectively prevent the damage caused by cyanobacteria outbreaks to the fish pond ecosystem. When the bananas are predicted to be nitrogen deficient, the pig manure input is automatically adjusted to the optimal value. This not only meets the nutrient needs of banana growth but also avoids resource waste and environmental pollution caused by excessive pig manure application, achieving precision fertilization and improving the scientific and sustainable nature of agricultural production.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. Efficient Multi-Source Data Collection and Integration: By leveraging drone remote sensing modules equipped with multispectral cameras, thermal infrared sensors, and lidar, along with various sensors from ground-based IoT sensor modules, a comprehensive data representation and recording platform was constructed, enabling efficient collection of multi-source data from banana canopies, fish ponds, and piggeries. Using spatiotemporal data cubes as building blocks, this data was integrated into a three-dimensional "field-time-indicator" structure, providing a complete and systematic data foundation for subsequent analysis and improving data usability.

[0016] 2. Accurate Data Processing and Prediction: Edge computing node modules are used to pre-process collected data. Combined with multimodal data fusion models and growth prediction models within the cloud-based intelligent platform module, this data is then deeply analyzed and predicted. For example, LSTM neural networks analyze spatiotemporal series data on algae density to predict the risk of cyanobacteria outbreaks, while Transformer hybrid models fuse multi-source data to predict banana yields and fish growth rates, providing precise decision-making for agricultural production.

[0017] 3. Data-driven intelligent control: Leveraging blockchain evidence storage and smart contract execution, data-driven automation of agricultural production processes is achieved. Based on predictions, smart contracts automatically trigger instructions such as pig manure application, pond mud return to the fields, and aerator startup and shutdown, improving production efficiency. Furthermore, reinforcement learning algorithms dynamically adjust the pig manure fermentation cycle and pond fertilization frequency based on data, optimizing resource recycling and reducing production costs.

[0018] 4. Data-based disease control: Utilizing drone hyperspectral imagery and the YOLOv8 model, combined with ground spore capture data and meteorological parameters, we construct a disease transmission probability model. Through early identification and risk warning of banana leaf spot, we develop targeted prevention and control measures based on this data, reducing disease losses and ensuring stable agricultural production. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the principle block diagram of the crop growth prediction system based on agricultural UAV remote sensing technology; Figure 2 Flowchart of the prediction method of the crop growth prediction system based on agricultural UAV remote sensing technology. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] See also Figure 1-Figure 2 The present invention provides a technical solution: a crop growth prediction system and method based on agricultural UAV remote sensing technology, comprising: 1. System Construction and Equipment Deployment (1) Hardware equipment installation 1. Drone Remote Sensing Module: A certain brand of Mavic 3M drone equipped with a multispectral camera and thermal infrared sensor was selected, while a certain brand of P100 drone equipped with a lidar was selected. Based on the distribution of banana plantations, fish ponds, and piggeries, drone route planning software (such as DJI Terra) was used to set circular or grid flight paths. The flight altitude was controlled at 50-80 meters to ensure full coverage of the monitoring area. For example, for a 100-mu banana plantation, a route with 10-meter intervals was planned. Flight missions were carried out every Tuesday and Friday morning between 9:00 and 11:00 AM (during stable sunlight hours) to collect imagery and data.

[0022] 2Ground IoT sensor module Pig houses: Alphasense NH3-B4 ammonia sensors and methane sensors are installed every 10 meters in the pig house, 1.2 meters above the ground (pig breathing height), and data is transmitted in real time to the edge computing node via a 4G DTU module.

[0023] Fish ponds: YSI ProDSS pH / EC probes (0.5 meters underwater) and Turner Designs Cyclops-7 algae density sensors (0.3 meters underwater) are arranged at intervals of 50 square meters. Buoy-type data collection terminals are installed at the edge of the ponds, and data is aggregated to edge computing nodes via LoRa wireless communication.

[0024] Banana fields: Decagon EM5 soil moisture sensors are buried every 20 meters at a depth of 15 cm. A sensor array is formed via a ZigBee network, transmitting data to edge computing nodes.

[0025] 3. Edge computing node module: Deploy NVIDIA Jetson AGX Orin edge servers in the farm control room, connect to the drone data receiving base station and ground sensor gateway via Gigabit Ethernet, configure data reception and preliminary processing programs, and ensure real-time data transmission and preprocessing.

[0026] 4. Cloud intelligent platform module: rent computing resources on a certain platform, build a cloud server cluster based on the Linux system, and deploy programs such as multimodal data fusion models, growth prediction models, circulatory system optimization algorithms, blockchain evidence storage units, and smart contract execution units.

[0027] 5. User Terminal Module: Develop a dedicated mobile app and web management platform compatible with Android, iOS, and mainstream browsers. Users log in with their username and password to visualize forecast results, send control commands, and monitor device status.

[0028] (2) Software system configuration 1. Edge computing node: Install the Ubuntu 20.04 operating system, deploy the Python 3.9 environment, and install libraries such as OpenCV, NumPy, and PyTorch.

[0029] Edge computing node data preprocessing process 1.1YOLOv8 target detection program flow 1.1.1 Data Input The visible light images collected by the drone are stored in RGB format with a resolution of 4000 × 3000 pixels and a data size of approximately 12MB per image. The image data is transmitted to the edge computing node and stored in a designated folder.

[0030] 1.1.2 Data Preprocessing Normalize the input image and normalize the pixel values to the [0,1] interval. The formula is: ; in, is the original pixel value, and The maximum and minimum pixel values in the image are respectively. For example, the RGB value of a pixel is (128, 64, 32), which becomes (0.5, 0.25, 0.125) after normalization.

[0031] 1.1.3 Model Loading and Inference Load the pretrained YOLOv8 weight file (e.g., yolov8n.pt, model size approximately 28MB). This model is trained on the COCO dataset and can recognize 80 object categories. In this system, retrain the model to recognize two object categories: banana leaf spot lesions and floating objects in fish ponds. During training, set the learning rate to 0.001 and the number of iterations to 500.

[0032] During the inference process, the preprocessed image is input into the model, and the model outputs the prediction results, including the target category (banana leaf spot / fish pond floating objects), bounding box coordinates And the confidence score (value range 0-1). For example, a test result shows that banana leaf spot is identified, and the bounding box coordinates are , with a confidence level of 0.85.

[0033] 1.1.4 Result screening and annotation Set the confidence threshold to 0.5 and filter out predictions with confidence levels below this threshold. Annotate the retained results on the original image, draw bounding boxes, and add category labels. The generated annotated images are stored in a folder.

[0034] 1.2 Kalman filter algorithm program flow 1.2.1 State Space Definition For sensor data (taking soil moisture sensor as an example), define the state vector: ; in is the soil moisture value, is the humidity change rate. State transfer matrix: , in is the sampling interval, assuming that the sampling interval of the soil moisture sensor is 30 minutes (i.e. Second).

[0035] 1.2.2 Initialization Parameters Initialize state estimate and the estimated error covariance matrix Assuming that the soil moisture value collected for the first time is 50%, then: ; ; represents the uncertainty of the initial estimate.

[0036] 1.2.3 Prediction Steps Make predictions based on the state transition equation: ; ; in, is the process noise covariance matrix, which is set according to the sensor characteristics: , Indicates the uncertainty in the soil moisture change process. For example, at the second sampling moment, ; .

[0037] 1.2.4 Update steps When a new measurement value is obtained (For example, the soil moisture measured at the second sampling moment is 52%), calculate the Kalman gain: ; Among them, the measurement matrix (because only soil moisture values are measured), the measurement noise covariance matrix (Set according to the sensor's measurement accuracy, indicating the uncertainty of the measurement value).

[0038] Update the state estimate and the estimation error covariance matrix: ; ; After calculation, the updated state estimate and error covariance matrix are obtained as the initial value for the next prediction.

[0039] 1.3 Data integration and upload The image data annotated by YOLOv8 object detection and sensor data processed by Kalman filtering are integrated and packaged into a specific format (such as JSON), containing information such as the image path, object detection results, sensor ID, and processed data values. This integrated data is uploaded to the cloud-based intelligent platform via the network communication module. The transmission protocol uses TCP / IP, and the data transmission rate is adjusted according to network conditions to ensure stable data transmission.

[0040] 2Cloud Intelligent Platform 2.1 Construction of multimodal data fusion model 2.1.1 Spatiotemporal Data Cube Construction Unit Data reception and cleaning: Receive UAV remote sensing data (such as multispectral image reflectance, thermal infrared temperature value, lidar point cloud coordinates), ground sensor data (ammonia concentration in pig houses, pH value in fish ponds, soil moisture in banana fields), and meteorological data (temperature, humidity, light intensity) from edge computing nodes. Fill missing values (such as using mean imputation) and remove outliers (using For example, if a pH sensor in a fish pond shows an abnormal value of 15 (the normal range is 6-9), the data will be discarded.

[0041] Three-dimensional structure construction: The three-dimensional dimensions are the field's geographic coordinates (e.g., the longitude and latitude of banana field area A is 113.2°E, 23.1°N), timestamps (accurate to the minute, such as 20XX-XX-XX 10:30), and indicator types (NDVI value, dissolved oxygen content, etc.). Cleaned data is stored according to this structure to form a spatiotemporal dataset. For example, data such as an NDVI value of 0.6, soil moisture of 55%, and temperature of 28°C in banana field area A at 20XX-XX-XX 10:30 are integrated into the corresponding spatiotemporal location.

[0042] 2.1.2 Bayesian Network Association Unit Variable definition: Determine the pig manure discharge (unit: tons / day), pig manure nutrient content (nitrogen content N%, phosphorus content P%, potassium content K%), banana growth indicators (plant height cm, leaf area index LAI), fish pond water quality parameters (dissolved oxygen DOmg / L, pH value, ammonia nitrogen content NH4 + mg / L) and other variables.

[0043] Conditional probability table construction: Statistical analysis is performed on historical data (e.g., daily data from the past year) to calculate the conditional probabilities between variables. For example, if the daily pig manure discharge is 2 tons, the probability of a banana plant growing 1-2 cm is 0.7; if the dissolved oxygen in a fish pond is below 5 mg / L and the nitrogen content of pig manure is above 3%, the probability of a blue-green algae outbreak is 0.8.

[0044] Model reasoning: Based on Bayesian formula: ; Based on the current known variable values (e.g., daily pig manure output of 3 tons and dissolved oxygen in the fish pond of 4 mg / L), the posterior probabilities of other variables are calculated to optimize the pig manure input strategy. For example, the calculation shows that under the current conditions, adding 0.5 tons of pig manure has the highest probability of increasing the banana leaf area index by 10%.

[0045] 2.2 Growth Prediction Model Deployment 2.2.1 LSTM Neural Network 1. Data Preparation: Collect historical data on algae density in fish ponds (at one-hour intervals for one week), along with water temperature (in °C) and light intensity (in lux) at the corresponding time. Split the data into a training set (80%) and a test set (20%) based on the time series.

[0046] 2. Model architecture settings: Set the number of LSTM layer neurons to 64, the number of hidden layers to 2, and the number of output layer neurons to 1 (predicting the probability of cyanobacteria outbreak, with a value range of 0-1).

[0047] 3. Training and Optimization: We used the mean squared error (MSE) as the loss function and the Adam optimizer for training. The learning rate was set to 0.001 and the number of iterations was 1000. For example, after training, the model achieved an MSE of 0.05 on the test set, and the average error in predicting the probability of a cyanobacteria bloom was 8%.

[0048] 2.2.2 Transformer Hybrid Model Feature extraction and fusion: Extract vegetation indices (NDVI, EVI) and texture features from drone multispectral imagery; obtain soil nutrient content and meteorological data from ground sensors; and combine these with historical yield data (unit: kg / mu). These features are normalized and then fused through splicing.

[0049] Model Architecture: Build a Transformer model that includes a multi-head attention mechanism (8 heads) and a multi-layer feedforward neural network (128 hidden layer dimensions). The input sequence length is determined by the number of data features. For example, if there are 10 features, the sequence length is 10.

[0050] Training and Prediction: We used the mean squared error (MSE) as the loss function, the stochastic gradient descent (SGD) optimizer, a learning rate of 0.01, and trained for 50 epochs. For example, after training, the model predicted banana yield per plant with a mean absolute error of 0.5 kg and predicted weekly growth rate of fish with an error of 5 g.

[0051] 2.3 Blockchain Evidence Storage Unit Configuration 2.3.1 Environment Construction Install Hyper Ledger Fabric 2.5, configure network nodes (including sorting nodes and peer nodes), set node IP addresses (for example, sorting node IP is 192.168.1.10, peer node IP is 192.168.1.11-192.168.1.15) and port numbers (default 7050-7054).

[0052] 2.3.2 Data encryption rule setting The SHA-256 hash algorithm is used to encrypt pig manure processing records (including input time and input amount), fish pond fertilization time, banana fertilization data, etc. For example, for the pig manure input time "20XX-XX-XX14:00" and input amount "2 tons", the SHA-256 algorithm generates a fixed-length 256-bit hash value, such as "5e884898da28047151d0e56f8dc6292773603d0d6aabbdd62a11ef721d1542d8".

[0053] 2.3.3 Smart Contract Writing and Trigger Condition Setting Smart contract writing: Use Go language to write smart contracts and define operation functions such as data storage, query, and update.

[0054] Trigger condition setting: Set the conditions for triggering the smart contract. For example, when the Bayesian network association unit calculates that the amount of pig manure input needs to be adjusted, it is triggered; when the LSTM neural network predicts that the probability of cyanobacteria outbreak exceeds the threshold of 20%, the fish pond aerator control-related smart contract is triggered.

[0055] 2.4 Reinforcement Learning Algorithm (PPO) Deployment 2.4.1 Environment and State Definition Environmental definition: The pig manure fermentation cycle (unit: day, range 5-10 days) and the frequency of fish pond fertilization and watering (unit: times / week, range 1-3 times) are used as controllable environmental variables; banana yield (unit: kg / mu), fish growth (unit: g / week), and resource utilization rate (pig manure utilization rate, pond mud utilization rate) are used as environmental feedback indicators.

[0056] State definition: The state vector contains information such as the current pig manure fermentation cycle, fish pond fertilization frequency, banana growth indicators, and fish pond water quality parameters. For example, [7, 2, 1200, 6.5, 8] respectively indicates a pig manure fermentation cycle of 7 days, a fish pond fertilization frequency of 2 times per week, a banana yield of 1200 kg / mu, a fish pond dissolved oxygen of 6.5 mg / L, and a pig manure utilization rate of 80%.

[0057] 2.4.2 Action and Reward Function Settings 1 Action definition: The actions that the agent can perform include adjusting the pig manure fermentation cycle (±1 day each time) and changing the frequency of fertilizing and watering the fish pond (±1 time each time).

[0058] 2 Reward function design: Reward function: ; in is the banana yield change rate, is the fish growth rate change, is the rate of change in resource utilization. For example, after adjusting the pig manure fermentation cycle and the frequency of fish pond fertilization, banana production increases by 10%, fish growth increases by 8%, and resource utilization increases by 5%. The reward value is: .

[0059] 2.4.3 Model Training and Strategy Optimization The PPO algorithm was used for training, with a learning rate of 0.0003, a batch size of 64, and 1000 training rounds. During training, the agent's strategy was continuously optimized based on the reward function. For example, after training, the optimal strategy was a six-day pig manure fermentation cycle and a fish pond fertilization frequency of twice a week, which achieved the best overall system performance.

[0060] User terminal: Develop APP and web interface based on Vue.js and React Native, call cloud API interface to obtain prediction results and device status data, and realize visual display and command interaction functions.

[0061] 2. Data Collection 1. UAV remote sensing data collection 1. The multispectral camera collects band images with a resolution of 0.1-0.5 meters and records banana canopy data.

[0062] 2 Thermal infrared sensors obtain the surface temperature distribution of the fish pond with an accuracy of ±0.1°C and identify areas with abnormal water temperature.

[0063] 3. LiDAR scans the pig house area and generates three-dimensional point cloud data for monitoring changes in the pig house structure and surrounding environment.

[0064] (2) Data collection from ground IoT sensors 1. The ammonia / methane sensor in the pig house monitors the gas concentration in real time. When the ammonia concentration is ≥20ppm and the methane concentration is ≥5%LEL, the alarm data will be uploaded immediately.

[0065] 2. The pH / EC probe in the fish pond collects data every 15 minutes (pH range: 0-14, accuracy: ±0.01; conductivity range: 0-20mS / cm, accuracy: ±0.01mS / cm); the algae density sensor monitors in real time and triggers an alarm when the data reaches 5000 / mL.

[0066] 3. The banana field soil moisture sensor collects data every 30 minutes (measurement range 0-100%, accuracy ±2%) and simultaneously records soil temperature data.

[0067] (3) Acquisition of meteorological data Obtain real-time meteorological data through the meteorological API interface, including temperature, humidity, light intensity, wind speed, precipitation, etc., to supplement environmental parameter information.

[0068] 3. Data Preprocessing 1Edge computing node processing 1.1 Data Input 1.1.1 Image acquisition and storage Visible light images collected by drones are stored in standard RGB format, with a typical resolution of 4000 × 3000 pixels. A single image is approximately 12MB in size. Image data is transmitted to edge computing nodes in real time via wireless transmission (e.g., Wi-Fi 5GHz, with a transmission rate of up to 867Mbps). The image data is stored in a designated folder, such as / data / uav_images / , with file names containing the acquisition timestamp (e.g., 20XXXXXX_103000.jpg), facilitating subsequent data management and traceability.

[0069] 1.1.2 Basic image information recording The image also records metadata such as its geographic location (latitude and longitude, with an accuracy of ±0.001°, such as 113.250°E, 23.120°N), shooting altitude (unit: meters, such as 50 meters), and flight attitude (pitch, roll, and yaw angles, with an accuracy of ±1°). These metadata are stored in a JSON file with the same name (such as 20XXXXXX_103000.json), providing auxiliary information for subsequent target positioning and analysis.

[0070] 1.11 Data Preprocessing 1.11.1 Pixel Value Normalization Normalize the pixel values of the input image and map them to the [0,1] interval. The formula is: in, is the original pixel value, and The maximum and minimum pixel values in the image are respectively. For example, the RGB value of a pixel is (128, 64, 32), which becomes (0.5, 0.25, 0.125) after calculation and normalization to meet the model input requirements and improve detection accuracy.

[0071] 1.11.2 Size Adjustment Resize the normalized image to the model input size, assuming the YOLOv8 model input size is 640×640 pixels. Use bilinear interpolation for scaling to maintain image clarity and integrity, and avoid loss of target features due to resizing.

[0072] 1.12 Model Detection 1.12.1 Model loading Load the pretrained YOLOv8 weight file (e.g., yolov8n.pt, model size approximately 28MB). This model is trained on the COCO dataset and then retrained for two target categories: banana leaf spot lesions and floating objects in fish ponds. During training, set the learning rate to 0.001, the batch size to 16, the number of iterations to 500, and the Adam optimizer to improve the model's ability to detect specific objects.

[0073] 1.12.2 Forward Reasoning The preprocessed image is input into the YOLOv8 model, which extracts image features through a convolutional neural network (CNN), fuses feature maps of different scales through a multi-layer feature pyramid network (FPN), and finally predicts the bounding box coordinates of the target through the anchor box mechanism. Category (Banana Leaf Spot / Floating Objects in Fish Pond) and confidence score (range 0-1). For example, a detection result shows that floating objects in a fish pond were identified, with bounding box coordinates of (50, 100, 150, 200) and a confidence score of 0.88.

[0074] 1.2 Results screening 1.2.1 Confidence Filtering Set the confidence threshold to 0.5 and filter the model output. Prediction boxes with a confidence level below 0.5 are discarded, while high-confidence object detection results are retained to reduce false positives. For example, if a prediction box has a confidence level of 0.45, it is discarded.

[0075] 1.2.2 Non-Maximum Suppression (NMS) For multiple overlapping prediction boxes for the same target, a non-maximum suppression algorithm is used to process them. The intersection over union (IoU) between the prediction boxes is calculated, and the prediction box with the lowest IoU value (for example, an IoU threshold of 0.3) and the highest confidence level is retained to eliminate redundant detections and ensure that only one optimal detection result is retained for each target.

[0076] 1.3 Result Annotation 1.3.1 Coordinate transformation and annotation drawing Convert the filtered bounding box coordinates from the normalized coordinates output by the model to the original image coordinates and scale them according to the image size. Use annotation tools (such as OpenCV's drawing functions) to draw a rectangular box on the original image to mark the target location and add a category label (such as "Banana Leaf Spot Disease" or "Floating Objects in Fish Pond") and a confidence score (such as "0.88").

[0077] 1.3.2 Annotation result storage Store the annotated images in a new folder, such as / data / annotated_images / , and generate a corresponding annotation information file (such as JSON format) to record information such as target category, bounding box coordinates, confidence level, and image geographic location to facilitate subsequent data analysis and system calls.

[0078] 2.1 Data Input Visible light images collected by drones are stored in RGB format with a resolution of 4000 × 3000 pixels and a data size of approximately 12MB per image. Images are accompanied by metadata such as geolocation information (latitude and longitude, with an accuracy of ±0.001°) and shooting altitude (typically 50-80 meters). These images are wirelessly transmitted (at a rate of approximately 867 Mbps) to edge computing nodes and stored in the designated directory / uav_images / . File names include the acquisition timestamp (e.g., 20XXXXXX_103000.jpg).

[0079] 2.11 Data Preprocessing 2.11.1 Pixel Value Normalization Using the formula: Normalize the image pixel values and map them to the [0, 1] range. For example, the RGB value of a pixel is (150, 75, 30), which becomes (0.6, 0.3, 0.12) after calculation and normalization to meet the model input requirements.

[0080] 2.11.2 Size Adjustment The normalized image is adjusted to the YOLOv8 model input size of 640×640 pixels, and the bilinear interpolation algorithm is used to maintain image clarity and avoid distortion of target features.

[0081] 2.12 Model Detection 2.12.1 Model loading Load the YOLOv8 weight file (e.g., yolov8n_custom.pt, approximately 28MB) pre-trained on the COCO dataset and then retrained for banana leaf spot and fish pond floating debris. Training parameters: learning rate 0.001, batch size 16, iterations 500, and Adam optimizer.

[0082] 2.12.2 Forward Reasoning The model uses a convolutional neural network (CNN) to extract features, a feature pyramid network (FPN) to fuse multi-scale information, and an anchor box mechanism to predict the target. The output contains the bounding box coordinates. (normalized value, range 0-1), category (Banana Leaf Spot / Fish Pond Floating Objects), confidence score (range 0-1). For example, a detection result: bounding box coordinates (0.1, 0.2, 0.3, 0.4), category "Fish Pond Floating Objects", confidence score 0.85.

[0083] 2.2 Results screening 2.2.1 Confidence Filtering Set the confidence threshold to 0.5 and remove prediction boxes with confidence lower than this value. For example, if the confidence of a prediction box is 0.4, it will be discarded directly.

[0084] 2.2.2 Non-maximum suppression (NMS) Calculate the intersection over union (IoU) between prediction boxes, set the IoU threshold to 0.3, retain prediction boxes with high confidence and low overlap, and eliminate redundant detections.

[0085] 2.3 Result Annotation 2.3.1 Coordinate Conversion and Drawing Convert the normalized bounding box coordinates to the original image coordinates and scale them according to the image size (4000 × 3000 pixels). Use the annotation tool to draw a rectangular box on the original image to mark the target location and add a category label (such as "Banana Leaf Spot") and a confidence score (such as "0.85").

[0086] 2.3.2 Result Storage The annotated images are stored in the / annotated_images / directory and corresponding JSON files are generated to record information such as target category, coordinates, confidence level, image metadata, etc. for subsequent analysis and call.

[0087] 3. Data Denoising: A Kalman filter algorithm is used to denoise sensor data. Based on historical sensor data and measurement error characteristics, filter parameters are dynamically adjusted to remove outliers. For example, the updated state estimate (such as soil moisture value and rate of change) after the Kalman filter is used as the denoised data output for subsequent data analysis and processing. For example, the denoised soil moisture value of 61.6875% and the rate of change of 0.625% / hour are output, providing more accurate data support for agricultural production decision-making.

[0088] 4. Feature extraction: 4.1 Multispectral Image Feature Extraction 4.1.1 Calculation of vegetation index NDVI (Normalized Difference Vegetation Index) calculation: Use the formula: in is the reflectivity in the near-infrared band, is the red light band reflectance. Assume that the near infrared band reflectance of a pixel point in the multispectral image is , red light band reflectivity ,but: NDVI values range from -1 to 1, with higher values indicating better vegetation cover.

[0089] EVI (Enhanced Vegetation Index) calculation: Use the formula in is the reflectivity of the blue light band. , , ,but: Compared with NDVI, EVI can better reduce the influence of atmospheric and soil background.

[0090] 4.1.2 Leaf Area Index (LAI) Calculation Calculate LAI using empirical models or lookup tables. For example, based on the NDVI-LAI empirical relationship model ,in and are the parameters obtained by fitting the experimental data, assuming , , when the average hour, , the unit is , which represents the total area of vegetation leaves per unit land area.

[0091] 4.11 Thermal Infrared Image Feature Extraction 4.11.1 Temperature Inversion The atmospheric correction method is used to invert the surface temperature. The formula is: Calculate the surface temperature, where is Planck's constant, is the speed of light, is the wavelength of thermal infrared band, is the Boltzmann constant, atmospheric transmittance (Assuming it is 0.8 measured by atmospheric parameters), surface emissivity (For farmland surface, it is assumed to be 0.95), is the radiation brightness received by the sensor, then the surface temperature of the area (Converted to approximately 27° Celsius).

[0092] 4.12 LiDAR Point Cloud Data Feature Extraction 4.12.1 Pig House Structural Feature Extraction Point cloud filtering and segmentation: Statistical filtering is used to remove outliers. The distance between each point and its neighboring points is calculated. If the distance exceeds the mean, the times the standard deviation (assuming ) are considered as outliers and removed. Then, the pig house point cloud is segmented from the background using a region growing algorithm or an edge detection-based method.

[0093] Structural parameter calculation: Extract the pig house's geometric structural features, such as calculating the minimum bounding rectangle of the point cloud to obtain the length and width of the pig house. The average and maximum heights of the pig house are calculated based on the point cloud's height distribution. For example, if the segmented pig house point cloud calculates the minimum bounding rectangle to be 50 meters long, 20 meters wide, with an average height of 4 meters and a maximum height of 5 meters, features such as point cloud density (the number of points per unit area, assuming 10 points per square meter) can also be calculated.

[0094] 4.2 Data Integration Integrate the extracted multispectral vegetation indices (NDVI, EVI), leaf area index (LAI), thermal infrared image temperature characteristics, and piggery structural features extracted from LiDAR point cloud data. For each region (e.g., each field or each piggery area), combine the relevant feature data into a single data record, for example: [region number, NDVI, EVI, LAI, surface temperature, piggery length, piggery width, piggery average height, piggery maximum height, point cloud density].

[0095] 4.3 Data Upload The integrated data is packaged and compressed using a compression algorithm (such as ZIP) to reduce its size. The data is uploaded to the cloud-based intelligent platform via a network transmission protocol (such as TCP / IP). During transmission, a data verification mechanism (such as MD5 checksum) is implemented to ensure data integrity. Assuming the packaged data size is 5MB, with a network bandwidth of 100Mbps, the theoretical transmission time is approximately 0.4 seconds. The processed data is packaged and uploaded to the cloud-based intelligent platform.

[0096] 4. Data Fusion and Analysis (1) Construction of spatiotemporal data cube The cloud-based intelligent platform's spatiotemporal data cube construction unit integrates drone remote sensing data, ground sensor data, and meteorological data into a three-dimensional "field-time-indicator" structure. For example, the NDVI value (0.65), soil moisture (60%), and temperature (25°C) for banana field area A at 10:00 AM on XX / XX / XXXX are integrated to form a multidimensional data record for that time point.

[0097] (2) Bayesian network association analysis 1. Data Preparation Agricultural production data from the past three years was collected, covering pig manure discharge (in tons / day, ranging from 0-5 tons, recorded daily), pig manure nutrient content (nitrogen content ranges from 1%-5%, phosphorus content 0.5%-3%, potassium content 0.3%-2%, tested weekly), banana growth indicators (plant height in cm, ranging from 50-300 cm, measured weekly; leaf area index (LAI) ranges from 0-5, measured weekly; fruit weight in kg / plant, recorded daily during the harvest period), and fish pond water quality parameters (dissolved oxygen ranges from 3-12 mg / L, pH 6-9, ammonia nitrogen content 0-2 mg / L, tested every two hours). This data was collected through sensor collection and laboratory testing to ensure data accuracy.

[0098] 2. Variable definition and network structure construction 2.1 Variable Definition Identify variable nodes in the Bayesian network, including: Pig manure emissions (A) are divided into three states: low (<1 ton / day), medium (1-3 tons / day), and high (>3 tons / day); The nitrogen content of pig manure (B) is divided into three states: low (<2%), medium (2%-3%), and high (>3%); Pig manure phosphorus content (C) is divided into three states: low (<1.5%), medium (1.5%-2%), and high (>2%); The potassium content (D) of pig manure is divided into three states: low (<1%), medium (1%-1.5%), and high (>1.5%); Banana plant height (E) is divided into three states: short (<100 cm), medium (100-200 cm), and tall (>200 cm); Banana leaf area index (F) is divided into three states: small (<2), medium (2-3.5), and large (>3.5); Banana fruit weight (G) is divided into three states: light (<1 kg / plant), medium (1-2 kg / plant), and heavy (>2 kg / plant); Dissolved oxygen (H) in fish ponds is divided into three states: low (<5mg / L), medium (5-8mg / L), and high (>8mg / L); The pH value (I) of fish ponds is divided into three states: low (<7), medium (7-8), and high (>8); The ammonia nitrogen content (J) in fish ponds is divided into three states: low (<0.5mg / L), medium (0.5-1mg / L), and high (>1mg / L).

[0099] 2.2 Network structure construction Based on agricultural production knowledge and experience, we identify causal relationships between variable nodes and construct a Bayesian network structure. For example, pig manure discharge affects its nutrient content, which in turn affects banana growth indicators and fish pond water quality parameters. Banana growth indicators and fish pond water quality parameters also have a certain correlation, thus forming a directed acyclic graph.

[0100] 3.1 Calculating Conditional Probability By using historical data statistics, we calculate the conditional probabilities of each variable node under different states. For example, if we calculate the probability of pig manure nitrogen content being "high" when pig manure emissions are "high," we can calculate the probability of pig manure nitrogen content being "high" by comparing the probability of 200 records of "high" pig manure emissions and 120 records of "high" pig manure nitrogen content over the past three years. This gives a probability of 0.6. Similarly, we calculate the conditional probabilities between other variable nodes.

[0101] 4. Model training and optimization 4.1 Training Method The Bayesian network is trained using the maximum likelihood estimation method, which adjusts the parameters in the conditional probability table to maximize the likelihood function of the model on the training data. In other words, a set of conditional probability values is found that maximizes the probability of observing the training data given these probabilities.

[0102] 4.2 Optimization process The historical data was divided into a training set (80%) and a test set (20%). The model was trained on the training set, and then its predictive accuracy was evaluated on the test set. The optimized Bayesian network model was obtained by adjusting the conditional probability table parameters multiple times, for example, by 0.05 each time, until the model's predictive accuracy on the test set no longer improved significantly (for example, the improvement was less than 1%).

[0103] 5. Model Decision Recommendations The posterior probabilities of different pig manure emission states are compared, and the pig manure input strategy corresponding to the state with the highest posterior probability is selected. If the calculated posterior probability of "high" pig manure emission is the highest, and the current pig manure input is at a medium level, a 10% increase in pig manure input is recommended (assuming that, based on experience and model calculations, the change from "medium" to "high" is 10%). This decision recommendation is also fed back to the user terminal for farmers' reference and implementation.

[0104] (3) Blockchain Evidence Storage The blockchain evidence storage unit uses the SHA-256 encryption algorithm to generate hash values for pig manure processing records (input time and amount), fish pond fertilization schedules, and banana fertilization data, and then stores them on the consortium chain. Each data update automatically records a timestamp and operation log to ensure that the data cannot be tampered with. For example, if 2 tons of pig manure were added to the banana field at 2:00 PM on XX / XX / 20XX, this record would be immediately stored on the chain.

[0105] (IV) Reinforcement Learning to Optimize Resource Cycles Using a reinforcement learning algorithm (PPO), the system dynamically adjusts the pig manure fermentation cycle and the frequency of fish pond fertilization using banana yield, fish growth, and resource utilization as reward functions. For example, if the initial pig manure fermentation cycle is 7 days, the PPO algorithm may discover that adjusting it to 6 days increases banana yield by 5%. The algorithm then automatically updates the fermentation cycle parameters and issues instructions through the smart contract execution unit.

[0106] 5. Model Prediction 1. LSTM neural network prediction of algae bloom risk The system uses historical data on algae density in fish ponds (one record per hour for the past seven days), water temperature, light intensity, and other time series data to feed into an LSTM neural network. After model training (learning rate 0.001, 1000 iterations), it predicts the probability of a cyanobacteria outbreak within the next 72 hours. If the predicted probability is ≥ 20% (threshold), an early warning mechanism is triggered.

[0107] 2. Transformer Hybrid Model Predicting Yield and Growth Rate 1. Data Preparation 1.1 UAV multispectral image data Multispectral drone imagery data was collected over the past two years for banana plantations and fish ponds. The images include red, green, blue, and near-infrared bands, with a resolution of 0.3 meters. Vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), were extracted from the images using the following formula: in is the reflectivity in the near-infrared band, is the red light band reflectance. Assume that in a certain area of multispectral image, a certain pixel point , , then the pixel point: At the same time, other spectral features are extracted, such as the ratio vegetation index (RVI), to reflect the vegetation growth status, where: .

[0108] 1.2 Ground sensor data Banana growing area: Soil moisture sensors collect data every 30 minutes, with a measurement range of 0-100% and an accuracy of ±2%. For example, the soil moisture value collected on one occasion was 65%. Soil nutrient sensors (detecting nitrogen, phosphorus, and potassium content) are tested once a week, with nitrogen content showing 1.2%, phosphorus content showing 0.8%, and potassium content showing 1.0%.

[0109] Fish pond area: The pH / EC probe collects data every 15 minutes, with a pH measurement range of 0-14 and an accuracy of ±0.01. The pH value of one measurement was 7.5. The dissolved oxygen sensor monitors in real time, with a measurement range of 0-20 mg / L and an accuracy of ±0.1 mg / L. For example, the current dissolved oxygen content is 8.2 mg / L. The algae density sensor collects data in real time, with the unit being pieces / mL.

[0110] 2. Model Input The input sequence contains 7 days of data. Each time step of the data includes drone multispectral imagery features (such as NDVI and RVI), ground sensor data (soil moisture, pH, etc.), and features related to historical yield data, forming a multidimensional tensor. For example, the shape of the input tensor is [7,10], where 7 represents the time step and 10 represents the feature dimension (including 5 multispectral imagery features, 3 soil sensor features, and 2 historical yield-related features).

[0111] 3. Result output After the model calculates, it obtains predicted values for banana yield per plant and weekly fish growth rate. For example, the predicted banana yield per plant is 16.4 kg, and the predicted weekly fish growth rate is 56 g. The predicted results are output for use.

[0112] (3) Identification of banana diseases Banana leaf spot disease identification and early warning process 1. Data Collection 1.1 UAV hyperspectral image acquisition A drone equipped with a hyperspectral imager (e.g., a hyperspectral camera from Headwall Photonics) was used to collect images of the banana plantation area. The acquisition time was selected between 9:00 AM and 11:00 AM (when lighting conditions were stable) at an altitude of 60 meters, covering all banana fields. The hyperspectral imagery covers the 400-1000 nm band, with a spectral resolution of 5 nm and a spatial resolution of 0.2 meters. The collected image data was stored as radiometric brightness values in units of .

[0113] 1.2 Ground spore capture data collection Five spore traps were evenly spaced 50 meters apart in the banana planting area. Spore samples were collected from the traps at 8:00 AM each day. The number of spores per cubic meter of air was counted using a microscope in the laboratory and the data was recorded. For example, one test revealed a spore count of 80 per cubic meter in a particular area.

[0114] 1.3 Meteorological parameter collection Weather stations installed in the banana growing areas collect real-time meteorological data, including humidity (measurement range 0-100%, accuracy ±1%, e.g., current humidity is 75%) and wind speed (measurement range 0-30 m / s, accuracy ±0.1 m / s, e.g., current wind speed is 2.5 m / s). Data collection intervals are 15 minutes.

[0115] 2. Identification of banana leaf spot 2.1 Spectral feature extraction Spectral reflectance data for banana leaf pixels was extracted from the preprocessed hyperspectral image, with reflectance values within the 400-1000nm band selected as features. For example, the reflectance values for a leaf pixel at 450nm, 550nm, 650nm, 750nm, 850nm, and 950nm were 0.2, 0.3, 0.25, 0.4, 0.5, and 0.45, respectively.

[0116] 2.2YOLOv8 Model Recognition Model Training: The YOLOv8 model was trained using a dataset of annotated banana leaf spot hyperspectral images (2,000 images, 1,600 for training and 400 for validation). The training parameters were: learning rate 0.001, batch size 16, iterations 300, and the Adam optimizer.

[0117] Model inference: The extracted spectral feature data of banana leaves are input into the trained YOLOv8 model, and the model outputs the judgment result of whether the leaf has leaf spot disease and the location of the diseased spot (expressed in bounding box coordinates, such as ) and confidence score (range 0-1). When the confidence score is ≥0.95, it is judged as an early symptom of leaf spot disease.

[0118] 3. Establishment of disease transmission probability model 3.1 Data Integration The leaf spot disease areas identified by the YOLOv8 model, ground spore capture data (number of spores per cubic meter), and meteorological parameters (humidity and wind speed) are integrated to form a dataset containing the diseased area identifier, spore count, humidity value, wind speed value, and disease status (1 indicates diseased, 0 indicates diseased). For example, a record might be [Area A, 80, 75%, 2.5 m / s, 1].

[0119] 3.2 Logistic regression model construction Variable definition: Let the dependent variable Indicates the incidence of banana leaf spot disease (1 means diseased, 0 means no disease), the independent variable is the number of spores per cubic meter, is the air humidity, is the wind speed.

[0120] Model formula: The logistic regression model formula is: .

[0121] in For a given variable The probability of disease occurring at 、 、 、 are model parameters.

[0122] Parameter estimation: Use maximum likelihood estimation to estimate the model parameters, calculated using the training data set (assuming it contains 1000 records) , , , ,.

[0123] 3.3 Prediction of disease spread probability The logistic regression model is used to calculate the probability of disease spread in each region by inputting the weather forecast data (humidity, wind speed) for the next 72 hours and the predicted spore count (based on historical data trends and current data estimates). For example, if the predicted spore count for a region in the next 72 hours is 100, the humidity is 80%, and the wind speed is 3 m / s, the model calculates: .

[0124] 4. Warning Release Based on the calculated probability of disease spread, the risk level is divided into: Low risk: probability < 0.3; Medium risk: 0.3≤probability<0.6; High risk: probability ≥ 0.6.

[0125] Early warning information is issued to farmers through text messages, APP push, etc., including the diseased area, risk level and prevention and control recommendations (such as the recommendation to spray fungicides immediately in high-risk areas).

[0126] 6. Command Output and Execution Cyanobacteria outbreak prevention instructions: When the LSTM neural network predicts that the probability of a cyanobacteria outbreak exceeds a threshold, the smart contract execution unit automatically generates an instruction: start the fish pond aerator at 4:00 PM (for 14 hours) and simultaneously push a recommendation to the farmer via the app to administer 1 kg of probiotics per acre. After the farmer confirms the recommendation, the system remotely activates the aerator and records the operation log.

[0127] Banana fertilization instructions: If the Transformer model predicts that bananas are nitrogen deficient, the Bayesian network calculates that the amount of pig manure input needs to be increased to 2 tons (with an error of ±5%). The smart contract execution unit sends instructions to the fertilizer spreader, which completes the spreading of pig manure according to the set amount and feeds back the execution results to the cloud and user terminal.

[0128] User interaction and monitoring: Users can view prediction results and command execution status through the APP or web platform, and can manually intervene in commands (such as shutting down the aerator in advance). The system records user operations in real time and updates the production log.

[0129] In summary: In the present invention, the UAV remote sensing module and the ground Internet of Things sensor module are used as data expression and recording carriers to collect multi-source data of banana canopy, fish pond and pig house areas, and the spatiotemporal data cube construction unit is used for integration to achieve comprehensive and efficient collection and organization of agricultural production data. The edge computing node module is used to pre-process the data, and the multimodal data fusion model, growth prediction model and other processing and recording carriers in the cloud intelligent platform module are combined to conduct in-depth analysis and prediction of the data, providing an accurate basis for agricultural production decision-making. With the help of the blockchain evidence storage unit and the smart contract execution unit, the automated production process is driven by data, and the reinforcement learning algorithm is used to optimize the resource cycle based on the data to achieve intelligent regulation of agricultural production. A disease prevention and control model is constructed based on UAV hyperspectral images, YOLOv8 model and multi-source data to achieve early identification and early warning of banana leaf spot disease, reduce disease losses, and ensure the stable development of agricultural production.

Claims

1. A crop growth prediction system based on agricultural drone remote sensing technology, characterized by: include: UAV remote sensing module: equipped with a multispectral camera, thermal infrared sensor, and lidar, used to obtain multispectral images of banana canopies, thermal infrared temperature distribution of fish ponds, and 3D point cloud data of piggery areas; Ground IoT sensor modules: These include ammonia / methane sensors deployed in pig houses, pH / EC probes and algae density sensors in fish ponds, and soil moisture sensors in banana fields, used to collect real-time parameters of aquaculture and planting environments. Edge computing node module: communicates with the UAV remote sensing module and the ground IoT sensor module to pre-process real-time data, including image noise reduction, outlier detection, and feature extraction; Cloud intelligent platform module: communicates with the edge computing node module, has a built-in multimodal data fusion model, a growth prediction model and a circulatory system optimization algorithm, and is used to analyze and predict the pre-processed data; User terminal module: communicates with the cloud-based intelligent platform module and is used to visualize the prediction results and receive control instructions.

2. The crop growth prediction system based on agricultural UAV remote sensing technology according to claim 1, characterized in that: The multimodal data fusion model includes a spatiotemporal data cube construction unit, which is used to fuse drone remote sensing data, ground sensor data, and meteorological data according to the three-dimensional "field-time-indicator" structure to generate a spatiotemporal dataset containing banana leaf area index, fish pond algae density, and pig house environmental parameters; and a Bayesian network association unit, which is used to establish a spatiotemporal correlation model between pig manure discharge and nutrient content, banana growth indicators, and fish pond water quality parameters to optimize pig manure input strategies.

3. The crop growth prediction system based on agricultural UAV remote sensing technology according to claim 1, characterized in that: The cloud-based intelligent platform module also includes: a blockchain evidence storage unit: which uses alliance chain technology to encrypt and store pig manure treatment records, fish pond fertilization time, and banana fertilization data to ensure that the data cannot be tampered with; a smart contract execution unit: based on the data of the blockchain evidence storage unit, triggers pig manure input, pond mud return to the field, and aerator start and stop instructions.

4. The crop growth prediction system based on agricultural UAV remote sensing technology according to claim 1, characterized in that: The growth prediction model includes: an LSTM neural network, which is used to analyze the spatiotemporal sequence data of algae density and predict the risk of blue-green algae outbreaks by combining water temperature and light intensity; and a Transformer hybrid model, which integrates drone multispectral image features and ground sensor data to predict the yield of individual banana plants and the growth rate of fish.

5. The crop growth prediction system based on agricultural UAV remote sensing technology according to claim 1, characterized in that: The edge computing node module includes: a YOLOv8 target detection unit, which is used to process the visible light images of the drone and identify abnormal targets such as banana leaf spot and floating objects in fish ponds; a Kalman filter unit, which is used to denoise the real-time sensor data and calibrate the remote sensing inversion parameters based on ground-based measured data.

6. A prediction method for a crop growth prediction system based on agricultural UAV remote sensing technology, characterized by: The following steps are involved: Data collection steps: The UAV remote sensing module acquires multispectral images of banana canopies, thermal infrared data of fish ponds, and 3D point clouds of pig houses. Ground sensors are used to collect pig house environment, fish pond water quality, and soil parameters. Preprocessing step: Use edge computing node modules to reduce noise, extract features, and filter outliers on the collected data to generate a standardized data set; Fusion processing step: The pre-processed multi-source data is input into the spatiotemporal data cube of the cloud-based intelligent platform module, and the association between pig manure treatment data and crop growth and aquaculture parameters is established through the Bayesian network; Model prediction steps: Use LSTM neural networks to predict the risk of algae outbreaks in fish ponds, and use Transformer hybrid models to predict banana yields and fish growth rates; Instruction output step: Generate control instructions based on the prediction results and send them to the execution device through the user terminal module.

7. The prediction method of the crop growth prediction system based on agricultural UAV remote sensing technology according to claim 6, characterized in that: In the data collection step, the UAV remote sensing module includes multispectral image processing: identifying the algae coverage area of the fish pond based on the NDVI and NDWI indices.

8. The prediction method of the crop growth prediction system based on agricultural UAV remote sensing technology according to claim 6, characterized in that: The fusion processing steps also include: using the blockchain evidence unit to store the amount of pig manure input, fertilization time, and pond mud return to the field on the chain to form an unalterable production log; using the reinforcement learning algorithm to dynamically adjust the pig manure fermentation cycle and the frequency of fish pond fertilization and watering.

9. The prediction method of the crop growth prediction system based on agricultural UAV remote sensing technology according to claim 6, characterized in that: The model prediction step includes banana disease identification: using drone hyperspectral images to extract the spectral characteristics of banana leaves, inputting them into the YOLOv8 model to identify early symptoms of leaf spot disease; combining ground spore capture data and meteorological parameters to establish a disease transmission probability model to provide early warning of disease risks.

10. The prediction method of the crop growth prediction system based on agricultural UAV remote sensing technology according to claim 6, characterized in that: In the instruction output step, when the LSTM neural network predicts that the probability of cyanobacteria outbreak exceeds a threshold, an aerator start-stop instruction and a probiotic dosage recommendation are generated; when the Transformer model predicts that bananas are nitrogen deficient, the pig manure input amount is adjusted to the optimal value based on the Bayesian network calculation results.

Images