Crop Growth Prediction System and Method Based on Agricultural Unmanned Aerial Vehicle Remote Sensing Technology

Abstract

This invention discloses a crop growth prediction system and method based on agricultural drone remote sensing technology, belonging to the field of agricultural information and intelligent monitoring technology. This system utilizes a drone remote sensing module and a ground-based IoT sensor module to collect multi-source data from banana cultivation, fishpond aquaculture, and pig farming as data recording carriers. An edge computing node module preprocesses the data, while a cloud-based intelligent platform module incorporates multimodal data fusion models and growth prediction models as processing and recording carriers, enabling in-depth data analysis and prediction. Through blockchain notarization and smart contracts, automated control of the production process is achieved based on the data. A disease control model is constructed using drone hyperspectral imagery and related data. This invention enhances the ability to collect, process, and apply agricultural production data, achieving accurate prediction and intelligent regulation, improving resource utilization efficiency, reducing disease losses, and promoting the intelligent development of agriculture.

Description

Technical Field

[0001] This invention relates to the field of agricultural information and intelligent monitoring technology, specifically to a crop growth prediction system and method based on agricultural drone remote sensing technology. Background Technology

[0002] In agricultural production, accurately grasping crop growth trends and adjusting production strategies in a timely manner are crucial for improving yield and quality. However, traditional agricultural production methods rely heavily on manual experience, lacking precise and real-time monitoring of the crop's growth environment and its own condition. Taking integrated banana-fishpond-pig farm farming as an example, on the one hand, it is difficult to comprehensively obtain multi-source data on the banana planting, fishpond farming, and pig farming environments, such as banana canopy information, fishpond water quality parameters, and pigsty air quality. This limitation in data acquisition prevents a deep understanding of the agricultural production system's operation. On the other hand, existing data recording and processing methods are outdated, resulting in scattered data that lacks effective integration, failing to form comprehensive and systematic agricultural production information and hindering precise decision-making. This leads to problems such as low resource utilization efficiency, untimely pest and disease control, and difficulty in improving economic benefits in agricultural production, urgently requiring innovative data representation, recording, and processing technologies to improve the current situation.

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

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

[0005] To solve the above-mentioned technical problems, the present invention provides a crop growth prediction system based on agricultural drone remote sensing technology, comprising:

[0006] UAV remote sensing module: equipped with a multispectral camera, thermal infrared sensor and lidar, used to acquire multispectral images of banana canopy, thermal infrared temperature distribution of fish ponds and three-dimensional point cloud data of pigsty area;

[0007] Ground-based IoT sensor modules include ammonia / methane sensors deployed in pigsties, pH / EC probes and algae density sensors in fishponds, and soil moisture sensors in banana fields, used to collect real-time environmental parameters for aquaculture and planting.

[0008] Edge computing node module: It is communicatively connected to the UAV remote sensing module and the ground IoT sensor module, and is used to preprocess real-time data, including image noise reduction, outlier detection and feature extraction;

[0009] Cloud-based intelligent platform module: Communicates with the edge computing node module, and has built-in multimodal data fusion model, growth prediction model and cyclic system optimization algorithm for in-depth analysis and prediction of preprocessed data;

[0010] User terminal module: Communicatively connected to the cloud-based intelligent platform module, used to visualize prediction results and receive control commands;

[0011] This system architecture enables efficient collection and integration of multi-source data, providing a comprehensive data foundation for accurate forecasting. The UAV remote sensing module can quickly acquire information on large areas of crops and livestock, while ground-based IoT sensors accurately collect local environmental parameters. Edge computing node modules alleviate data transmission pressure and perform initial data processing, the cloud-based intelligent platform module conducts in-depth analysis, and the user terminal module allows users to intuitively understand and intervene in production. This system significantly improves the efficiency of agricultural production data acquisition and processing, providing strong support for optimizing agricultural production decisions.

[0012] Furthermore, the multimodal data fusion model includes: a spatiotemporal data cube construction unit, used to fuse UAV remote sensing data, ground sensor data, and meteorological data according to a three-dimensional structure of "field-time-indicator" to generate a spatiotemporal dataset containing banana leaf area index, fishpond algae density, and pigsty environmental parameters; and a Bayesian network association unit, used to establish a spatiotemporal association model between pig manure discharge, nutrient content, banana growth indicators, and fishpond water quality parameters, optimizing pig manure input strategies. The multimodal data fusion model can uncover potential connections between different data. The spatiotemporal data cube construction unit integrates multi-source data, enabling the data to reflect correlations in time and space dimensions, facilitating a comprehensive understanding of the dynamic changes in agricultural production. The Bayesian network association unit, by establishing correlations between pig manure and crop and aquaculture parameters, can precisely adjust pig manure input according to actual conditions, improving resource utilization efficiency, reducing resource waste, and ensuring the stability of crop growth and the aquaculture environment.

[0013] Furthermore, the cloud-based intelligent platform module also includes: a blockchain evidence storage unit: using consortium blockchain technology to encrypt and store records of pig manure treatment, fishpond fertilization times, and banana fertilization data, ensuring the data is tamper-proof; and a smart contract execution unit: based on the data from the blockchain evidence storage unit, automatically triggering commands such as pig manure input, pond mud return to the field, and aerator start / stop. The blockchain evidence storage unit provides a reliable storage method for agricultural production data, and its tamper-proof nature ensures the authenticity and credibility of the data, facilitating agricultural product traceability and enhancing consumer trust in agricultural product quality. The smart contract execution unit automatically executes relevant commands based on the evidence storage data, automating the agricultural production process, reducing manual intervention, improving production efficiency, and ensuring the accuracy and timeliness of operations.

[0014] Furthermore, the growth prediction model includes: an LSTM neural network for analyzing spatiotemporal data of algal density and combining it with water temperature and light intensity to predict the risk of cyanobacterial blooms; and a Transformer hybrid model that integrates UAV multispectral image features and ground sensor data to predict banana yield per plant and fish growth rate. This growth prediction model can accurately predict key indicators in agricultural production. The LSTM neural network, by analyzing algal density and environmental factors, provides early warnings of cyanobacterial bloom risks, enabling fish farmers to take timely measures to reduce the risk of fish mortality due to cyanobacterial blooms and ensure fishery production. The Transformer hybrid model integrates multiple data sources to predict crop yield and fish growth rate, providing a scientific basis for farmers to rationally plan production and estimate profits, thus helping to improve the economic efficiency of agricultural production.

[0015] Furthermore, the edge computing node module includes: a YOLOv8 target detection unit for processing UAV visible light images and identifying abnormal targets such as banana leaf spot disease and floating debris in fishponds; and a Kalman filter unit for denoising real-time sensor data and calibrating remote sensing inversion parameters based on ground-measured data. The built-in units of the edge computing node module enhance the real-time performance and accuracy of data processing. The YOLOv8 target detection unit can quickly identify abnormalities such as banana leaf spot disease, enabling farmers to prevent and control diseases in a timely manner and reduce crop losses. The Kalman filter unit denoises the sensor data and calibrates remote sensing inversion parameters, improving data quality and thus enhancing the accuracy of subsequent data analysis and prediction, providing more reliable data support for agricultural production decisions.

[0016] The prediction method of a crop growth prediction system based on agricultural drone remote sensing technology includes the following steps:

[0017] Data acquisition steps: Use the drone remote sensing module to acquire multispectral images of banana canopy, thermal infrared data of fishpond, and 3D point cloud of pigsty; use ground sensors to collect pigsty environment, fishpond water quality and soil parameters.

[0018] Preprocessing steps: The edge computing node module is used to perform noise reduction, feature extraction and outlier filtering on the collected data to generate a standardized dataset;

[0019] Fusion processing steps: Input the preprocessed multi-source data into the spatiotemporal data cube of the cloud intelligent platform module, and establish the correlation between pig manure treatment data and crop growth and aquaculture parameters through Bayesian network;

[0020] Model prediction steps: Use LSTM neural network to predict the risk of algal blooms in fish ponds, and use Transformer hybrid model to predict banana yield and fish growth rate;

[0021] Command output steps: Generate control commands based on the prediction results and send them to the execution equipment (such as fertilizer applicator and aerator) through the user terminal module.

[0022] This forecasting method forms a complete closed loop of data processing and decision-making. The data acquisition step comprehensively collects relevant agricultural production data; the preprocessing step ensures data quality; the fusion processing step uncovers data correlations; the model prediction step yields accurate forecast results; and the instruction output step transforms the forecast results into actual production control instructions. This series of interconnected steps effectively optimizes the agricultural production process, improves resource utilization efficiency, reduces production costs, and enhances the overall benefits of agricultural production.

[0023] Furthermore, in the data acquisition step, the UAV remote sensing module includes multispectral image processing: identifying algae-covered areas in fishponds based on NDVI and NDWI indices; the multispectral image processing method in the UAV remote sensing module can accurately acquire algae information in fishponds. Based on NDVI and NDWI indices, algae-covered areas and boundaries can be clearly identified, providing intuitive data for algae monitoring, helping to more accurately understand algae growth status, and providing a scientific basis for fishpond ecological environment assessment and fishery production management.

[0024] Furthermore, the fusion processing step also includes: storing the amount of pig manure input, fertilization time, and pond mud return records on the blockchain through a blockchain notarization unit to form an immutable production log; and dynamically adjusting the pig manure fermentation cycle and fishpond fertilization frequency using a reinforcement learning algorithm (PPO) to optimize resource recycling efficiency. The application of blockchain notarization and reinforcement learning algorithms in the fusion processing step brings significant benefits. The immutable production log formed by blockchain notarization facilitates the traceability and management of the agricultural production process, improving the credibility of agricultural product quality. The reinforcement learning algorithm dynamically adjusts the pig manure fermentation cycle and fishpond fertilization frequency, 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.

[0025] Furthermore, the model prediction step includes banana disease identification: extracting spectral features of banana leaves using UAV hyperspectral imagery, inputting this data into the YOLOv8 model to identify early symptoms of leaf spot disease with an accuracy rate of no less than 95%; and establishing a disease transmission probability model by combining ground spore capture data and meteorological parameters, providing an early warning of disease risk 72 hours in advance. The banana disease identification method in the model prediction step can effectively control banana diseases. The high accuracy of identifying early symptoms of leaf spot disease using UAV hyperspectral imagery and the YOLOv8 model allows farmers to take timely control measures in the early stages of the disease, reducing its spread. The disease transmission probability model established by combining ground spore capture data and meteorological parameters can provide early warning of disease risk, giving farmers more time for prevention and control, and reducing yield losses caused by banana diseases.

[0026] Furthermore, in the instruction output step, when the LSTM neural network predicts that the probability of a cyanobacterial bloom exceeds a threshold, it automatically generates instructions to start / stop the aerator and suggestions for the dosage of probiotics. When the Transformer model predicts nitrogen deficiency in bananas, it automatically adjusts the amount of pig manure input to the optimal value (error ±5%) based on the Bayesian network calculation results. The instruction output step can promptly transform the prediction results into effective production control measures. When the predicted probability of a cyanobacterial bloom exceeds the threshold, the automatically generated instructions to start / stop the aerator and suggestions for the dosage of probiotics can effectively prevent the damage of cyanobacterial blooms to the fishpond ecosystem. When a nitrogen deficiency in bananas is predicted, automatically adjusting the amount of pig manure input to the optimal value can both meet the nutrient requirements for banana growth and avoid excessive input of pig manure, which would lead to resource waste and environmental pollution, achieving precision fertilization and improving the scientific and sustainable nature of agricultural production.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. Efficient Acquisition and Integration of Multi-Source Data: By utilizing a UAV remote sensing module equipped with a multispectral camera, thermal infrared sensor, and lidar, along with various sensors from a ground-based IoT sensor module, a comprehensive data representation and recording carrier was constructed, enabling efficient acquisition of multi-source data from banana canopy, fishpond, and pigsty areas. Spatiotemporal data cube construction units were used to fuse these data according to a three-dimensional structure of "field-time-indicator," providing a complete and systematic data foundation for subsequent analysis and improving data usability.

[0029] 2. Precise Data Processing and Prediction: Edge computing node modules are used to preprocess the collected data. Combined with multimodal data fusion models and growth prediction models in the cloud-based intelligent platform module, the data is subjected to in-depth analysis and prediction. For example, LSTM neural networks analyze spatiotemporal sequence data of algal density to predict the risk of cyanobacterial blooms, and Transformer hybrid models fuse multi-source data to predict banana yield and fish growth rates, providing precise decision-making support for agricultural production.

[0030] 3. Data-Driven Intelligent Control: Utilizing blockchain-based notarization units and smart contract execution units, data-driven automation of agricultural production processes is achieved. Based on prediction results, smart contracts automatically trigger commands such as pig manure input, pond mud return to the fields, and aerator start / stop, improving production efficiency. Simultaneously, reinforcement learning algorithms dynamically adjust the pig manure fermentation cycle and fishpond fertilization frequency based on data, optimizing resource recycling and reducing production costs.

[0031] 4. Data-driven disease control: Utilizing UAV hyperspectral imagery and the YOLOv8 model, combined with ground-based spore capture data and meteorological parameters, a disease transmission probability model is constructed. Through early identification and risk warning of banana leaf spot disease, targeted control measures are developed based on data to reduce disease losses and ensure stable agricultural production. Attached Figure Description

[0032] Figure 1 This is a block diagram illustrating the principle of a crop growth prediction system based on agricultural drone remote sensing technology.

[0033] Figure 2 This is a flowchart of the prediction method for a crop growth prediction system based on agricultural drone remote sensing technology. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Please see Figures 1-2 This invention provides a technical solution: a crop growth prediction system and method based on agricultural unmanned aerial vehicle (UAV) remote sensing technology, comprising:

[0036] I. System Setup and Equipment Deployment

[0037] (a) Hardware equipment installation

[0038] 1. UAV Remote Sensing Module: A Mavic 3M UAV of a certain brand was selected, equipped with a multispectral camera and a thermal infrared sensor; a P100 UAV of another brand was equipped with a lidar. Based on the distribution of banana plantations, fishponds, and pigsties, UAV flight path planning software (such as DJI Terra) was used to set circular or grid-like flight paths, with flight altitude controlled between 50-80 meters to ensure coverage of the entire monitoring area. For example, for a 100-acre banana plantation, flight paths were planned at 10-meter intervals, and flight missions were carried out every Tuesday and Friday morning from 9-11 am (during periods of stable sunlight) to collect images and data.

[0039] 2 Ground IoT sensor modules

[0040] Pigsty: Alphasense NH3-B4 ammonia and methane sensors are installed every 10 meters in the pigsty, 1.2 meters above the ground (the height at which pigs breathe), and the data is transmitted to the edge computing node in real time via a 4G DTU module.

[0041] Fishpond: YSI ProDSS pH / EC probes (0.5 meters underwater) and Turner Designs Cyclops-7 algae density sensors (0.3 meters underwater) are installed at 50-square-meter intervals. A buoy-type data acquisition terminal is installed at the edge of the pond, and the data is aggregated to the edge computing node via LoRa wireless communication.

[0042] Banana field: Decagon EM5 soil moisture sensors are buried every 20 meters at a depth of 15 centimeters. They are connected to form a sensor array via a ZigBee network to transmit data to edge computing nodes.

[0043] 3 Edge Computing Node Module: Deploy NVIDIA Jetson AGX Orin edge servers in the farm's central control room, connect to the drone data receiving base station and ground sensor gateway via gigabit Ethernet, configure data receiving and preliminary processing programs to ensure real-time data transmission and preprocessing.

[0044] 4. Cloud-based intelligent platform module: Leasing computing resources on a certain platform to build a cloud server cluster based on the Linux system, and deploying programs such as multimodal data fusion model, growth prediction model, cyclic system optimization algorithm, blockchain evidence storage unit and smart contract execution unit.

[0045] 5. User Terminal Module: Develop a dedicated mobile app and web management platform, compatible with Android, iOS systems, and mainstream browsers. Users log in with an account and password to view forecast results visually, send control commands, and monitor device status.

[0046] (II) Software System Configuration

[0047] 1. Edge computing node: Install Ubuntu 20.04 operating system, deploy Python 3.9 environment, and install libraries such as OpenCV, NumPy, and PyTorch.

[0048] Edge computing node data preprocessing process

[0049] 1.1 YOLOv8 Object Detection Program Flow

[0050] 1.1.1 Data Input

[0051] Visible light images captured 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 then transmitted to edge computing nodes and stored in a designated folder.

[0052] 1.1.2 Data Preprocessing

[0053] The input image is standardized by normalizing the pixel values ​​to the [0,1] interval, using the following formula:

[0054] ;

[0055] in, These are the original pixel values. and These are the maximum and minimum pixel values ​​in the image, respectively. For example, if a pixel has RGB values ​​of (128, 64, 32), after normalization it becomes (0.5, 0.25, 0.125).

[0056] 1.1.3 Model Loading and Inference

[0057] Load the pre-trained YOLOv8 weight file (e.g., yolov8n.pt, model size approximately 28MB). This model is trained on the COCO dataset and can identify 80 target classes. In this system, the model is retrained to identify two target classes: banana leaf spot disease lesions and floating debris in fishponds. The learning rate is set to 0.001 and the number of iterations is 500.

[0058] During inference, the preprocessed image is input into the model, and the model outputs the prediction results, including the target category (banana leaf spot disease / fishpond floating debris) and bounding box coordinates. And a confidence score (ranging from 0 to 1). For example, a test result showed that banana leaf spot disease was detected, with the bounding box coordinates as follows: The confidence level is 0.85.

[0059] 1.1.4 Result Filtering and Labeling

[0060] Set a confidence threshold of 0.5 to filter out predictions with a confidence level below this threshold. Label the retained results on the original image, draw bounding boxes, add category labels, and store the labeled image in a folder.

[0061] 1.2 Kalman Filter Algorithm Program Flow

[0062] 1.2.1 State Space Definition

[0063] For sensor data (taking a soil moisture sensor as an example), define a state vector:

[0064] ;

[0065] in This represents the soil moisture value. Let be the rate of change of humidity. State transition matrix:

[0066] ,

[0067] in The sampling time interval is assumed to be 30 minutes (i.e., ...). Second).

[0068] 1.2.2 Initialization Parameters

[0069] Initial state estimate and the estimated error covariance matrix Assuming the initial soil moisture content is 50%, then:

[0070] ;

[0071] ;

[0072] This indicates the uncertainty of the initial estimate.

[0073] 1.2.3 Prediction Steps

[0074] Prediction is made based on the state transition equation:

[0075] ;

[0076] ;

[0077] in, The process noise covariance matrix is ​​set according to the sensor characteristics:

[0078] ,

[0079] This indicates the uncertainty in the process of soil moisture change. For example, at the second sampling time,

[0080] ;

[0081]

[0082]

[0083] .

[0084] 1.2.4 Update Steps

[0085] When new measurement values ​​are obtained (For example, if the soil moisture measured at the second sampling time is 52%), calculate the Kalman gain:

[0086] ;

[0087] Among them, the measurement matrix (Because only soil moisture values ​​are measured), measure the noise covariance matrix. (Based on the sensor's measurement accuracy setting, this indicates the uncertainty of the measured value.)

[0088] Update the state estimate and the covariance matrix of the estimation error:

[0089] ;

[0090] ;

[0091] After calculation, the updated state estimate and error covariance matrix are obtained, which are used as the initial values ​​for the next prediction.

[0092] 1.3 Data Integration and Upload

[0093] The image data annotated with YOLOv8 target detection and the sensor data processed by Kalman filtering are integrated and packaged into a specific format (such as JSON), including information such as image path, target detection results, sensor ID, and processed data values. The integrated data is then uploaded to a cloud-based intelligent platform via a network communication module. The transmission protocol is TCP / IP, and the data transmission rate is adjusted according to network conditions to ensure stable data transmission.

[0094] 2. Cloud-based intelligent platform

[0095] 2.1 Construction of Multimodal Data Fusion Model

[0096] 2.1.1 Spatiotemporal Data Cube Building Unit

[0097] Data reception and cleaning: This involves receiving UAV remote sensing data (such as multispectral image reflectance, thermal infrared temperature values, and lidar point cloud coordinates), ground sensor data (ammonia concentration in pigsties, pH values ​​in fishponds, and soil moisture in banana fields) and meteorological data (temperature, humidity, and light intensity) from edge computing nodes. Missing value imputation (e.g., using mean imputation) and outlier removal (through...) are then performed. (Principles), for example, if the pH sensor of a fishpond shows an abnormal value of 15 (the normal range is 6-9), then that data is removed.

[0098] Three-dimensional structure construction: The three dimensions are: field geographic coordinates (e.g., banana field A area has latitude and longitude of 113.2°E, 23.1°N), timestamp (accurate to the minute, e.g., 20XX-XX-XX 10:30), and indicator types (NDVI value, dissolved oxygen content, etc.). The cleaned data is stored according to this structure to form a spatiotemporal dataset. For example, data such as the NDVI value of 0.6, soil moisture of 55%, and air temperature of 28℃ for banana field A area at 20XX-XX-XX 10:30 are integrated into the corresponding spatiotemporal location.

[0099] 2.1.2 Bayesian Network Association Unit

[0100] Variable definitions: Determine the following parameters: pig manure discharge (unit: tons / day), pig manure nutrient content (nitrogen N%, phosphorus P%, potassium K%), banana growth indicators (plant height cm, leaf area index LAI), and fishpond water quality parameters (dissolved oxygen DO mg / L, pH value, ammonia nitrogen NH4). + Variables such as mg / L.

[0101] Conditional probability table construction: Statistical analysis is performed using historical data (such as daily data from the past year) to calculate the conditional probabilities between variables. For example, when the daily discharge of pig manure is 2 tons, the probability of banana plant height increasing by 1-2 cm is 0.7; when the dissolved oxygen in the fishpond is below 5 mg / L and the nitrogen content of pig manure is above 3%, the probability of cyanobacterial bloom is 0.8.

[0102] Model inference: Based on Bayes' theorem:

[0103] ;

[0104] Based on the known values ​​of variables (e.g., daily pig manure discharge of 3 tons, dissolved oxygen in fishpond of 4 mg / L), the posterior probabilities of other variables are calculated to optimize the pig manure input strategy. For example, calculations show that under the current conditions, increasing the input of 0.5 tons of pig manure has the highest probability of increasing the banana leaf area index by 10%.

[0105] 2.2 Deployment of Growth Prediction Model

[0106] 2.2.1 LSTM Neural Network

[0107] 1. Data Preparation: Collect historical data on algae density in fishponds (with an hourly time interval, collected continuously for one week), water temperature (unit: °C), and light intensity (unit: lux) at the corresponding times. Divide the data into a training set (80%) and a test set (20%) according to the time series.

[0108] 2. Model architecture settings: The number of neurons in the LSTM layer is set to 64, the number of hidden layers is 2, and the number of neurons in the output layer is 1 (to predict the probability of cyanobacterial blooms, with a value range of 0-1).

[0109] 3. Training and Optimization: The mean squared error (MSE) was used as the loss function, and the Adam optimizer was used for training. The learning rate was set to 0.001, and the number of iterations was 1000. For example, after training, the model had an MSE of 0.05 on the test set, and the average error in predicting the probability of cyanobacterial blooms was 8%.

[0110] 2.2.2 Transformer Hybrid Model

[0111] Feature extraction and fusion: Vegetation indices (NDVI, EVI) and texture features were extracted from UAV multispectral imagery; soil nutrient content and meteorological data were obtained from ground sensors; and historical yield data (unit: kg / mu) were combined. These features were normalized and then fused by stitching.

[0112] Model Structure: Construct a Transformer model that includes a multi-head attention mechanism (8 heads) and a multi-layer feedforward neural network (128 hidden layers). The length of the input sequence is determined by the number of data features; for example, if there are 10 features, the sequence length is 10.

[0113] Training and Prediction: The model was trained for 50 epochs using a stochastic gradient descent (SGD) optimizer with a learning rate of 0.01 and a loss function of mean squared error (MSE). For example, after training, the model's average absolute error in predicting the yield of a single banana plant was 0.5 kg, and its error in predicting the weekly growth rate of fish was 5 g.

[0114] 2.3 Blockchain Evidence Storage Unit Configuration

[0115] 2.3.1 Environment Setup

[0116] Install Hyper Ledger Fabric 2.5, configure network nodes (including sorting nodes and peer nodes), set node IP addresses (e.g., 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).

[0117] 2.3.2 Setting Data Encryption Rules

[0118] The SHA-256 hash algorithm is used to encrypt records of pig manure treatment (including input time and amount), fishpond fertilization time, and banana fertilization data. For example, for the pig manure input time "20XX-XX-XX 14:00" and input amount "2 tons", a fixed-length 256-bit hash value is generated using the SHA-256 algorithm, such as "5e884898da28047151d0e56f8dc6292773603d0d6aabbdd62a11ef721d1542d8".

[0119] 2.3.3 Smart Contract Writing and Trigger Condition Setting

[0120] Smart contract writing: Use the Go language to write smart contracts and define functions for data storage, querying, updating, and other operations.

[0121] Trigger Condition Setting: Set the conditions for triggering the smart contract, such as triggering when the Bayesian network association unit calculates that the amount of pig manure input needs to be adjusted; triggering the smart contract related to the fishpond aerator control when the LSTM neural network predicts that the probability of cyanobacteria bloom exceeds the threshold of 20%.

[0122] 2.4 Deployment of Reinforcement Learning Algorithm (PPO)

[0123] 2.4.1 Definition of Environment and State

[0124] Environmental definition: The fermentation cycle of pig manure (unit: days, range 5-10 days) and the frequency of fertilization in fish ponds (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.

[0125] State definition: The state vector contains information such as the current pig manure fermentation cycle, fishpond fertilization frequency, banana growth indicators, and fishpond water quality parameters, such as [7,2,1200,6.5,8]; representing a pig manure fermentation cycle of 7 days, a fishpond fertilization frequency of 2 times / week, a banana yield of 1200 kg / mu, a fishpond dissolved oxygen of 6.5 mg / L, and a pig manure utilization rate of 80%, respectively.

[0126] 2.4.2 Setting the Action and Reward Functions

[0127] 1. Action definition: Actions that the agent can perform include adjusting the fermentation cycle of pig manure (±1 day each time) and changing the frequency of fertilization in fish ponds (±1 time each time).

[0128] 2. Reward Function Design: Reward Function:

[0129] ;

[0130] in The rate of change in banana yield. The rate of change in fish growth. This represents the rate of change in resource utilization. For example, if adjusting the fermentation cycle of pig manure and the frequency of fertilization in fishponds results in a 10% increase in banana yield, an 8% increase in fish growth, and a 5% increase in resource utilization, then the reward value is:

[0131] .

[0132] 2.4.3 Model Training and Policy Optimization

[0133] 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. After training, the optimal strategy was determined to be a 6-day fermentation cycle for pig manure and a 2-times-per-week fertilization frequency for fishponds, resulting in the best overall system efficiency.

[0134] 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 visualization display and command interaction functions.

[0135] II. Data Collection

[0136] (I) Unmanned Aerial Vehicle (UAV) Remote Sensing Data Acquisition

[0137] 1. A multispectral camera was used to acquire band images with a resolution of 0.1-0.5 meters to record banana canopy data.

[0138] 2. Thermal infrared sensors acquire the surface temperature distribution of fishponds with an accuracy of ±0.1℃, and identify areas with abnormal water temperature.

[0139] 3. The LiDAR scanner scans the pigsty area to generate three-dimensional point cloud data, which is used to monitor changes in the pigsty's building structure and surrounding environment.

[0140] (ii) Ground-based Internet of Things (IoT) sensor data acquisition

[0141] 1. The ammonia / methane sensor in the pigsty monitors the gas concentration in real time. When the ammonia concentration is ≥20ppm or the methane concentration is ≥5%LEL, alarm data is immediately uploaded.

[0142] 2. The pH / EC probe in the fishpond 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, triggering an alarm when the data reaches 5000 algae / mL.

[0143] 3. The soil moisture sensor in the banana field collects data every 30 minutes (measurement range 0-100%, accuracy ±2%), and records soil temperature data simultaneously.

[0144] (III) Meteorological Data Acquisition

[0145] Real-time meteorological data, including temperature, humidity, light intensity, wind speed, and precipitation, can be obtained through the meteorological API interface to supplement environmental parameter information.

[0146] III. Data Preprocessing

[0147] 1 Edge computing node processing

[0148] 1.1 Data Input

[0149] 1.1.1 Image Acquisition and Storage

[0150] Visible light images captured by drones are stored in standard RGB format, with a common resolution of 4000×3000 pixels and a single image data size of approximately 12MB. The image data is transmitted in real time to edge computing nodes via wireless transmission (such as Wi-Fi 5GHz band, with a transmission rate of up to 867Mbps), and stored in a designated folder, such as / data / uav_images / . The filename includes the acquisition timestamp (such as 20XXXXXX_103000.jpg), which facilitates subsequent data management and traceability.

[0151] 1.1.2 Recording of Basic Image Information

[0152] Simultaneously, the image's geographic location information (latitude and longitude, accuracy ±0.001°, such as 113.250°E, 23.120°N), shooting altitude (unit: meters, such as 50 meters), flight attitude (pitch angle, roll angle, yaw angle, accuracy ±1°) and other metadata are recorded and stored in a JSON file with the same name (such as 20XXXXXX_103000.json) to provide auxiliary information for subsequent target positioning and analysis.

[0153] 1.11 Data Preprocessing

[0154] 1.11.1 Pixel Value Normalization

[0155] The pixel values ​​of the input image are normalized and mapped to the [0,1] interval, using the following formula:

[0156]

[0157] in, These are the original pixel values. and These represent the maximum and minimum pixel values ​​in the image. For example, if a pixel has RGB values ​​of (128, 64, 32), after normalization, it becomes (0.5, 0.25, 0.125) to adapt to the model input requirements and improve detection accuracy.

[0158] 1.11.2 Size Adjustment

[0159] The normalized image is adjusted to the model input size, assuming the YOLOv8 model input size is 640×640 pixels. Bilinear interpolation is used for scaling to maintain the sharpness and integrity of the image content and avoid loss of target features due to size transformation.

[0160] 1.12 Model Detection

[0161] 1.12.1 Model Loading

[0162] Load the pre-trained YOLOv8 weight file (e.g., yolov8n.pt, model size approximately 28MB). This model is initially trained on the COCO dataset, and then undergoes secondary training targeting two categories of objects: banana leaf spot disease lesions and floating debris in fishponds. During training, the learning rate is set to 0.001, the batch size to 16, and the number of iterations to 500. The Adam optimizer is used to improve the model's ability to detect specific objects.

[0163] 1.12.2 Forward Reasoning

[0164] 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 an anchor box mechanism. The category (banana leaf spot disease / fishpond floating debris) and the confidence score (range 0-1). For example, a detection result showed that fishpond floating debris was identified, with bounding box coordinates of (50, 100, 150, 200) and a confidence score of 0.88.

[0165] 1.2 Result Screening

[0166] 1.2.1 Confidence Filtering

[0167] Set a confidence threshold of 0.5 to filter the model output. Discard predicted boxes with a confidence level below 0.5, and retain high-confidence target detection results to reduce false positives. For example, if a predicted box has a confidence level of 0.45, it will be removed.

[0168] 1.2.2 Non-maximum suppression (NMS)

[0169] For multiple overlapping predicted bounding boxes that may exist for the same target, a non-maximum suppression algorithm is used. The intersection-over-union (IoU) ratio between each predicted box is calculated, and the predicted box with the lowest IoU value (e.g., an IoU threshold of 0.3) and the highest confidence is retained to eliminate redundant detections and ensure that only one optimal detection result is retained for each target.

[0170] 1.3 Result Labeling

[0171] 1.3.1 Coordinate Transformation and Annotation Drawing

[0172] The filtered bounding box coordinates are converted from the normalized coordinates output by the model to the original image coordinates, and scaled according to the image size. Using annotation tools (such as OpenCV's drawing functions), rectangular boxes are drawn on the original image to mark the target locations, and category labels (such as "banana leaf spot disease" and "fishpond floating debris") and confidence scores (such as "0.88") are added.

[0173] 1.3.2 Storage of Annotation Results

[0174] The labeled images are stored in a new folder, such as / data / annotated_images / , and a corresponding annotation information file (such as JSON format) is generated to record information such as target category, bounding box coordinates, confidence level, and image geographic location, which facilitates subsequent data analysis and system calls.

[0175] 2.1 Data Input

[0176] Visible light images captured 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 images are accompanied by metadata such as geolocation information (latitude and longitude, accuracy ±0.001°) and shooting altitude (usually 50-80 meters). They are transmitted wirelessly (transmission rate approximately 867Mbps) to edge computing nodes and stored in a designated directory / uav_images / . The filename includes the acquisition timestamp (e.g., 20XXXXXX_103000.jpg).

[0177] 2.11 Data Preprocessing

[0178] 2.11.1 Pixel Value Normalization

[0179] Formula used:

[0180]

[0181] The image pixel values ​​are normalized and mapped to the [0,1] interval. For example, if the RGB value of a pixel is (150,75,30), it becomes (0.6,0.3,0.12) after normalization to adapt to the model input requirements.

[0182] 2.11.2 Size Adjustment

[0183] The normalized image was adjusted to the YOLOv8 model input size of 640×640 pixels, and bilinear interpolation was used to maintain image clarity and avoid distortion of target features.

[0184] 2.12 Model Detection

[0185] 2.12.1 Model Loading

[0186] Load the YOLOv8 weight file (e.g., yolov8n_custom.pt, model size approximately 28MB) that was pre-trained on the COCO dataset and then retrained for banana leaf spot disease and fishpond floating debris. Training parameters: learning rate 0.001, batch size 16, number of iterations 500, and Adam optimizer.

[0187] 2.12.2 Forward Reasoning

[0188] The input is a preprocessed image. The model extracts features using a convolutional neural network (CNN), fuses multi-scale information via a feature pyramid network (FPN), and predicts targets using an anchor box mechanism. The output includes bounding box coordinates. (Normalized value, range 0-1), category (banana leaf spot / fishpond floating debris), confidence score (range 0-1). For example, in a certain test result: bounding box coordinates (0.1, 0.2, 0.3, 0.4), category "fishpond floating debris", confidence score 0.85.

[0189] 2.2 Result Screening

[0190] 2.2.1 Confidence Filtering

[0191] Set a confidence threshold of 0.5 and discard predicted boxes with a confidence level lower than this value. For example, if a predicted box has a confidence level of 0.4, it is discarded directly.

[0192] 2.2.2 Non-maximum suppression (NMS)

[0193] Calculate the Intersection over Union (IoU) ratio between predicted boxes, set the IoU threshold to 0.3, retain predicted boxes with high confidence and low overlap, and eliminate redundant detections.

[0194] 2.3 Result Labeling

[0195] 2.3.1 Coordinate Transformation and Plotting

[0196] 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 rectangles on the original image to mark the target locations, and add category labels (such as "banana leaf spot disease") and confidence scores (such as "0.85").

[0197] 2.3.2 Result Storage

[0198] The labeled images are stored in the / annotated_images / directory, and a corresponding JSON file is generated to record information such as target category, coordinates, confidence level, and image metadata for subsequent analysis.

[0199] 3. Data Denoising: A Kalman filter algorithm is used to denoise the sensor data. Based on historical sensor data and measurement error characteristics, the filter parameters are dynamically adjusted to remove outliers. For example, the updated state estimate (such as soil moisture value and rate of change) after Kalman filtering is used as the denoised data output for subsequent data analysis and processing. For instance, the denoised soil moisture value is 61.6875% and the rate of change is 0.625% / hour, providing more accurate data support for agricultural production decisions.

[0200] 4. Feature Extraction:

[0201] 4.1 Multispectral Image Feature Extraction

[0202] 4.1.1 Calculation of Vegetation Index

[0203] NDVI (Normalized Difference Vegetation Index) calculation: The formula is as follows:

[0204]

[0205] in For near-infrared reflectivity, Let be the red light band reflectance. Assume the near-infrared band reflectance of a pixel in a multispectral image. Red light band reflectivity ,but:

[0206]

[0207] The NDVI value ranges from -1 to 1, and a higher value indicates better vegetation coverage.

[0208] EVI (Enhanced Vegetation Index) calculation: using the formula

[0209]

[0210] in This refers to the reflectivity in the blue light band. If a certain pixel... , , ,but:

[0211]

[0212] EVI is better at reducing the impact of atmospheric and soil background than NDVI.

[0213] 4.1.2 Calculation of Leaf Area Index (LAI)

[0214] LAI can be calculated using empirical models or lookup table methods. For example, based on the NDVI-LAI empirical relationship model. ,in and The parameters are obtained by fitting experimental data, assuming , When the average of a certain area hour, The unit is , which represents the total area of ​​vegetation leaves on a unit of land area.

[0215] 4.11 Thermal Infrared Image Feature Extraction

[0216] 4.11.1 Temperature Inversion

[0217] Surface temperature inversion was performed using the atmospheric correction method. The formula is:

[0218]

[0219] Calculate the surface temperature, where Let be Planck's constant. At the speed of light, It is the wavelength of the thermal infrared band. Boltzmann constant, atmospheric transmittance (Assuming the value is 0.8 as measured by atmospheric parameters), surface emissivity (For farmland surface, the value is assumed to be 0.95). Given the radiance received by the sensor, the surface temperature of that area is... (Converted to approximately 27°C).

[0220] 4.12 Feature Extraction from LiDAR Point Cloud Data

[0221] 4.12.1 Extraction of structural features of pigsty buildings

[0222] Point cloud filtering and segmentation: Statistical filtering is used to remove outliers. This is done by calculating the distance between each point and its neighbors; if the distance exceeds the mean, additional points are added. Double the standard deviation (assumption) Points with outliers (e.g., 0) are considered outliers and removed. Then, the pigsty point cloud is segmented from the background using a region growing algorithm or an edge detection-based method.

[0223] Structural parameter calculation: Extract the geometric structural features of the pigsty, such as calculating the minimum bounding rectangle of the point cloud to obtain the length and width dimensions of the pigsty; calculate the average and maximum height of the pigsty through the height distribution of the point cloud. Assuming that the minimum bounding rectangle obtained from the segmented pigsty point cloud is 50 meters long, 20 meters wide, has an average height of 4 meters, and a maximum height of 5 meters, the point cloud density (the number of points per unit area, assumed to be 10 points / square meter) and other features can also be calculated.

[0224] 4.2 Data Integration

[0225] The extracted vegetation indices (NDVI, EVI), leaf area index (LAI) from multispectral images, temperature features from thermal infrared images, and structural features of pigsties extracted from lidar point cloud data are integrated. Relevant feature data are grouped into a single data record, organized by region (e.g., each field or each pigsty area), for example: [Region ID, NDVI, EVI, LAI, Surface Temperature, Pigsty Length, Pigsty Width, Average Pigsty Height, Maximum Pigsty Height, Point Cloud Density].

[0226] 4.3 Data Upload

[0227] The integrated data is packaged and compressed using algorithms such as ZIP compression to reduce its size. It is then uploaded to the cloud-based intelligent platform via a network transmission protocol (such as TCP / IP), with a data verification mechanism (such as MD5 checksum) in place during transmission to ensure data integrity. Assuming the packaged data size is 5MB, the theoretical transmission time is approximately 0.4 seconds with a network bandwidth of 100Mbps. The processed data is then packaged and uploaded to the cloud-based intelligent platform.

[0228] IV. Data Fusion and Analysis

[0229] (I) Construction of Spatiotemporal Data Cube

[0230] The spatiotemporal data cube construction unit of the cloud-based intelligent platform integrates UAV remote sensing data, ground sensor data, and meteorological data according to a three-dimensional structure of "field-time-indicator". For example, it integrates data such as NDVI value (0.65), soil moisture (60%), and temperature (25℃) of banana field A at 10:00 on XX / XX / XXXX to form a multi-dimensional data record for that time point.

[0231] (II) Bayesian Network Association Analysis

[0232] 1. Data Preparation

[0233] Agricultural production data from the past three years were collected, covering pig manure emissions (unit: tons / day, range 0-5 tons, recorded daily), pig manure nutrient content (nitrogen content range 1%-5%, phosphorus content 0.5%-3%, potassium content 0.3%-2%, measured weekly), banana growth indicators (plant height unit: cm, range 50-300cm, measured weekly; leaf area index (LAI), range 0-5, measured weekly; fruit weight unit: kg / plant, recorded daily during harvest), and fishpond water quality parameters (dissolved oxygen range 3-12 mg / L, pH value 6-9, ammonia nitrogen content 0-2 mg / L, measured every 2 hours). All data were acquired through sensor collection and laboratory testing to ensure accuracy.

[0234] 2. Variable Definition and Network Structure Construction

[0235] 2.1 Variable Definition

[0236] Determining the variable nodes in a Bayesian network includes:

[0237] Pig manure emissions (A) are categorized into three levels: low (<1 ton / day), medium (1-3 tons / day), and high (>3 tons / day).

[0238] The nitrogen content (B) in pig manure is divided into three states: low (<2%), medium (2%-3%), and high (>3%).

[0239] The phosphorus content (C) in pig manure is classified into three states: low (<1.5%), medium (1.5%-2%), and high (>2%).

[0240] The potassium content (D) in pig manure is classified into three states: low (<1%), medium (1%-1.5%), and high (>1.5%).

[0241] Banana plant height (E) is divided into three categories: short (<100cm), medium (100-200cm), and tall (>200cm).

[0242] Banana leaf area index (F) is divided into three categories: small (<2), medium (2-3.5), and large (>3.5).

[0243] Banana fruit weight (G) is divided into three categories: light (<1kg / plant), medium (1-2kg / plant), and heavy (>2kg / plant);

[0244] Dissolved oxygen (H) in fishponds is classified into three states: low (<5mg / L), medium (5-8mg / L), and high (>8mg / L).

[0245] The pH value (I) of fish ponds is divided into three states: low (<7), medium (7-8), and high (>8).

[0246] 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).

[0247] 2.2 Network Structure Construction

[0248] Based on agricultural production knowledge and experience, the causal relationships between variable nodes are determined, and a Bayesian network structure is constructed. For example, the amount of pig manure discharged affects the nutrient content of pig manure, which in turn affects banana growth indicators and fishpond water quality parameters. There is also a certain correlation between banana growth indicators and fishpond water quality parameters, thus forming a directed acyclic graph.

[0249] 3.1 Calculate conditional probability

[0250] By statistically analyzing historical data, the conditional probability of each variable node under different states is calculated. For example, to calculate the probability that the nitrogen content of pig manure is "high" when the amount of pig manure discharge is "high": In the past 3 years, there are 200 records of "high" pig manure discharge, of which 120 records show "high" nitrogen content in pig manure, so the probability is 0.6. Similarly, the conditional probabilities between other variable nodes are calculated.

[0251] 4. Model Training and Optimization

[0252] 4.1 Training Methods

[0253] The maximum likelihood estimation method is used to train the Bayesian network by adjusting the parameters in the conditional probability table to maximize the likelihood function of the model on the training data. In other words, it finds a set of conditional probability values ​​that, given these probabilities, maximize the probability of observing the training data.

[0254] 4.2 Optimization Process

[0255] Historical data is divided into a training set (80%) and a test set (20%). The model is trained on the training set, and then the prediction accuracy of the model is evaluated on the test set. The conditional probability table parameters are adjusted multiple times, such as by 0.05 each time, until the prediction accuracy of the model on the test set no longer improves significantly (e.g., the improvement is less than 1%), resulting in the optimized Bayesian network model.

[0256] 5. Model Decision Recommendations

[0257] Compare the posterior probabilities of different pig manure emission levels and select the pig manure input strategy corresponding to the state with the highest posterior probability. If the calculated posterior probability of "high" pig manure emission level is the highest, and the current pig manure input level is medium, then it is recommended to increase the pig manure input level by 10% (assuming that, based on experience and model calculations, the input level change from "medium" to "high" is 10%). Simultaneously, this decision recommendation is fed back to the user terminal for farmers to refer to and implement.

[0258] (III) Blockchain-based evidence storage

[0259] The blockchain-based evidence storage unit generates hash values ​​for records such as pig manure treatment (time and amount), fishpond fertilization time, and banana fertilization data using the SHA-256 encryption algorithm, and stores these hash values ​​on the consortium blockchain. Each time data is updated, a timestamp and operation log are automatically recorded to ensure data immutability. For example, if 2 tons of pig manure were added to a banana field at 14:00 on [Date], this record is immediately stored on the blockchain.

[0260] (iv) Strengthen learning and optimize resource circulation

[0261] Using a reinforcement learning algorithm (PPO), with banana yield, fish growth, and resource utilization rate as reward functions, the fermentation cycle of pig manure and the frequency of fertilization in fish ponds are dynamically adjusted. For example, if the initial fermentation cycle of pig manure is 7 days, and the PPO algorithm learns that adjusting it to 6 days can increase banana yield by 5%, then the fermentation cycle parameters are automatically updated, and instructions are issued through a smart contract execution unit.

[0262] V. Model Prediction

[0263] (a) LSTM Neural Network for Predicting Algal Outbreak Risk

[0264] Historical data on algae density in fishponds (one data point per hour for the past 7 days), water temperature, and light intensity are input into an LSTM neural network. After model training (learning rate 0.001, 1000 iterations), the probability of a cyanobacterial bloom in the next 72 hours is predicted. If the predicted probability is ≥20% (threshold), an early warning mechanism is triggered.

[0265] (ii) Transformer hybrid model for predicting yield and growth rate

[0266] 1. Data Preparation

[0267] 1.1 UAV Multispectral Image Data

[0268] Multispectral drone imagery data of banana plantation and fishpond areas from the past two years were collected. 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. The calculation formula is as follows:

[0269]

[0270] in For near-infrared reflectivity, Let be the reflectance in the red light band. Assume that in a multispectral image of a certain region, a certain pixel... , Then the pixel:

[0271]

[0272] Simultaneously, other spectral features, such as the ratio vegetation index (RVI), are extracted to reflect vegetation growth status, including:

[0273] .

[0274] 1.2 Ground sensor data

[0275] Banana growing areas: Soil moisture sensors collect data every 30 minutes, with a measurement range of 0-100% and an accuracy of ±2%. For example, a soil moisture value of 65% is collected in a certain instance. Soil nutrient sensors (detecting nitrogen, phosphorus, and potassium content) are tested once a week, such as nitrogen content of 1.2%, phosphorus content of 0.8%, and potassium content of 1.0%.

[0276] Fishpond area: pH / EC probe collects data every 15 minutes, pH measurement range 0-14, accuracy ±0.01, pH value measured in a certain measurement is 7.5; dissolved oxygen sensor monitors in real time, measurement range 0-20mg / L, accuracy ±0.1mg / L, such as the current dissolved oxygen content is 8.2mg / L; algae density sensor collects data in real time, unit is cells / mL.

[0277] 2. Model Input

[0278] The input sequence contains 7 days of data. The data at each time step includes UAV multispectral image features (such as NDVI, RVI, etc.), ground sensor data (soil moisture, pH value, etc.), and relevant features of 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 (containing 5 multispectral image features, 3 soil sensor features, and 2 historical yield-related features).

[0279] 3. Results Output

[0280] After model calculations, predicted values ​​for banana yield per plant and fish weekly growth rate are obtained. For example, the predicted banana yield per plant is 16.4 kg; the predicted fish weekly growth rate is 56 g. The prediction results are output for use.

[0281] (III) Identification of Banana Diseases

[0282] Banana leaf spot disease identification and early warning process

[0283] 1. Data Collection

[0284] 1.1 UAV Hyperspectral Image Acquisition

[0285] Images of banana-growing areas were captured using drones equipped with hyperspectral imagers (such as Headwall Photonics' hyperspectral cameras). Data collection was conducted between 9 and 11 AM (when lighting conditions are stable), at a flight altitude of 60 meters, covering all banana fields. The hyperspectral images covered the 400-1000 nm band, with a spectral resolution of 5 nm and a spatial resolution of 0.2 meters. The acquired image data was stored as radiance values, in units of... .

[0286] 1.2 Ground spore capture data collection

[0287] Five spore traps were evenly distributed 50 meters apart in the banana growing area. Samples were collected from the traps every morning at 8:00 AM, and the number of spores per cubic meter of air was counted using a microscope in the laboratory, with the data recorded. For example, one test showed 80 spores per cubic meter in a certain area.

[0288] 1.3 Meteorological Parameter Collection

[0289] Meteorological data is collected in real time by weather stations installed in banana growing areas, including air humidity (measurement range 0-100%, accuracy ±1%, e.g., current humidity 75%) and wind speed (measurement range 0-30m / s, accuracy ±0.1m / s, e.g., current wind speed 2.5m / s), with a data collection interval of 15 minutes.

[0290] 2. Identification of Banana Leaf Spot Disease

[0291] 2.1 Spectral Feature Extraction

[0292] Spectral reflectance data of banana leaf pixels were extracted from the preprocessed hyperspectral image, and reflectance values ​​in the 400-1000nm band were selected as features. For example, the reflectance values ​​of a certain leaf pixel in the 450nm, 550nm, 650nm, 750nm, 850nm, and 950nm bands were 0.2, 0.3, 0.25, 0.4, 0.5, and 0.45, respectively.

[0293] 2.2 YOLOv8 Model Recognition

[0294] Model Training: The YOLOv8 model was trained using a labeled hyperspectral image dataset of banana leaf spot disease (containing 2000 images, of which 1600 were used for training and 400 for validation). The training parameters were set as follows: learning rate 0.001, batch size 16, number of iterations 300, and the Adam optimizer was used.

[0295] Model inference: The extracted spectral feature data of banana leaves is input into the trained YOLOv8 model. The model outputs the judgment result of whether the leaf has leaf spot disease and the location of the lesion (represented by bounding box coordinates, e.g., ...). The test results include a confidence score (range 0-1). A confidence score ≥ 0.95 indicates an early symptom of leaf spot disease.

[0296] 3. Establishment of a disease transmission probability model

[0297] 3.1 Data Integration

[0298] The information on leaf spot disease outbreaks identified by the YOLOv8 model, ground spore capture data (spore count per cubic meter), and meteorological parameters (humidity and wind speed) are integrated to form a dataset containing disease outbreak area identifiers, spore counts, humidity values, wind speed values, and whether the disease is present (1 indicates disease, 0 indicates no disease). For example, a record might be [Area A, 80, 75%, 2.5 m / s, 1].

[0299] 3.2 Logistic Regression Model Construction

[0300] Variable definition: Let the dependent variable be... The independent variable represents the incidence of banana leaf spot disease (1 for diseased, 0 for unaffected). Number of spores per cubic meter For air humidity, This refers to wind speed.

[0301] Model Formula: The formula for the logistic regression model is:

[0302] .

[0303] in Given independent variables The probability of developing the disease at that time. , , , These are the model parameters.

[0304] Parameter estimation: The model parameters are estimated using the maximum likelihood estimation method, calculated using the training dataset (assuming it contains 1000 records). , , , , .

[0305] 3.3 Disease transmission probability prediction

[0306] By inputting the weather forecast data (humidity, wind speed) for the next 72 hours and the predicted spore count (estimated based on historical data trends and current data) into a logistic regression model, the probability of disease transmission in each region can be calculated. For example, if a region is expected to have 100 spores, 80% humidity, and a wind speed of 3 m / s in the next 72 hours, the following calculation can be performed using the model:

[0307] .

[0308] 4. Early Warning Issuance

[0309] Based on the calculated disease transmission probability, risk levels are classified as follows:

[0310] Low risk: probability < 0.3;

[0311] Medium risk: 0.3 ≤ probability < 0.6;

[0312] High risk: Probability ≥ 0.6.

[0313] Early warning information is sent to farmers via SMS, APP push notifications, etc., including the affected areas, risk levels, and prevention and control recommendations (such as immediate spraying of fungicides in high-risk areas).

[0314] VI. Command Output and Execution

[0315] Blue-green algae bloom prevention instructions: When the LSTM neural network predicts that the probability of a blue-green algae bloom exceeds a threshold, the smart contract execution unit automatically generates an instruction: turn on the fishpond aerator at 16:00 (for 14 hours continuously), and simultaneously push a suggestion to the fish farmer via the APP to add 1kg of probiotics per acre. After the fish farmer confirms, the system remotely controls the aerator to start and records the operation log.

[0316] 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 (error ±5%). The smart contract execution unit sends instructions to the fertilizer applicator, which completes the spreading of pig manure according to the set amount and feeds back the execution results to the cloud and user terminal.

[0317] 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 turning off the aerator in advance). The system records user operations in real time and updates production logs.

[0318] In summary, this invention utilizes UAV remote sensing modules and ground-based IoT sensor modules as data recording carriers to collect multi-source data from banana canopy areas, fishponds, and pigsties. This data is then integrated using a spatiotemporal data cube construction unit, achieving comprehensive and efficient collection and organization of agricultural production data. Edge computing node modules are used to preprocess the data, which is then combined with multimodal data fusion models and growth prediction models in the cloud-based intelligent platform module to perform in-depth analysis and prediction, providing precise data for agricultural production decisions. Blockchain-based evidence storage and smart contract execution units are used to drive automated production processes, while reinforcement learning algorithms optimize resource recycling based on data, enabling intelligent control of agricultural production. Based on UAV hyperspectral imagery, the YOLOv8 model, and multi-source data, a disease control model is constructed to achieve early identification and warning of banana leaf spot disease, reducing disease losses and ensuring stable agricultural production.

Claims

1. A crop growth prediction method based on agricultural unmanned aerial vehicle (UAV) remote sensing technology, characterized in that: Includes the following steps: Data acquisition steps: Use the drone remote sensing module to acquire multispectral images of banana canopy, thermal infrared data of fishpond, and 3D point cloud of pigsty; use ground sensors to collect pigsty environment, fishpond water quality and soil parameters. Preprocessing steps: The edge computing node module is used to perform noise reduction, feature extraction and outlier filtering on the collected data to generate a standardized dataset; Fusion processing steps: Input the preprocessed multi-source data into the spatiotemporal data cube of the cloud intelligent platform module, and establish the correlation between pig manure treatment data and crop growth and aquaculture parameters through Bayesian network; Model prediction steps: Use LSTM neural network to predict the risk of algal blooms in fish ponds, and use Transformer hybrid model to predict banana yield and fish growth rate; Command output steps: Generate control commands based on the prediction results and send them to the execution device through the user terminal module; The process involves establishing a correlation between pig manure treatment data and crop growth and aquaculture parameters using a Bayesian network. Variables include pig manure discharge volume, pig manure nutrient content, banana growth indicators, and fishpond water quality parameters. Historical data is used for statistical analysis to calculate the conditional probabilities between these variables, based on Bayes' theorem. ； Based on the currently known variable values, calculate the posterior probabilities of other variables to optimize the pig manure input strategy; The fusion processing step also includes: dynamically adjusting the fermentation cycle of pig manure and the frequency of fertilization in fish ponds using reinforcement learning algorithms; When deploying reinforcement learning algorithms, the fermentation cycle of pig manure and the frequency of fertilization in fish ponds are used as adjustable environmental variables; banana yield, fish growth, and resource utilization rate are used as environmental feedback indicators; the actions that the agent can perform include adjusting the fermentation cycle of pig manure and changing the frequency of fertilization in fish ponds; the reward function is: ； in The rate of change in banana yield. The rate of change in fish growth. This represents the rate of change in resource utilization; during training, the agent's strategy is continuously optimized based on the reward function.

2. The crop growth prediction method based on agricultural UAV remote sensing technology as described in claim 1, characterized in that: The fusion processing step also includes: storing the amount of pig manure input, fertilization time, and pond mud return records on the blockchain through a blockchain evidence storage unit to form an unalterable production log.

3. The crop growth prediction method based on agricultural UAV remote sensing technology as described in claim 1, characterized in that: In the instruction output step, when the LSTM neural network predicts that the probability of a cyanobacterial bloom exceeds the threshold, it generates an aerator start / stop instruction and a probiotic dosage suggestion.

Images