Peanut yield intelligent prediction system and method based on multispectral image analysis

Abstract

The invention relates to the technical field of crop yield prediction, in particular to an intelligent peanut yield prediction system and method based on multispectral image analysis. According to the technical scheme, the system comprises a multi-spectral image acquisition module, a dynamic adaptive image processing unit, a multi-modal feature fusion analysis module and an adaptive yield prediction model, wherein the multi-spectral image acquisition module is configured to pass through an unmanned aerial vehicle group carrying multi-spectral imaging equipment; five-dimensional spectral data containing visible light, red edge, near-infrared, short-wave infrared and thermal infrared bands are synchronously obtained, wherein the thermal infrared band is used for capturing the plant transpiration dynamic state. According to the method, multi-spectral image analysis, multi-modal feature fusion and intelligent algorithm technologies are comprehensively utilized, precise monitoring of peanut growth and high-precision prediction of the yield are achieved, visual and effective decision support is provided for agricultural production, meanwhile, the system operation cost is optimized, ecological economic benefits are excavated, and development of precise agriculture and sustainable agriculture is powerfully promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of crop yield prediction, and in particular to a peanut yield intelligent prediction system and method based on multispectral image analysis. Background Art

[0002] Peanuts are an important cash crop in modern agricultural production, and their yield prediction is crucial for rationally arranging agricultural activities, optimizing resource allocation, and ensuring stable market supply. However, existing peanut yield prediction methods have many limitations and are unable to meet the needs of precision agriculture development. Traditional yield prediction methods rely on empirical judgment, simple statistical models, or manual field sampling surveys. These methods are not only labor-intensive, material-intensive, and time-consuming, but also significantly impact prediction accuracy due to subjective factors and sample representativeness. While some research using optical remote sensing imagery for yield prediction has made progress with the development of remote sensing technology, conventional optical images only contain visible light information and cannot fully reflect the complex changes in the physiological state of peanut plants and their growing environment.

[0003] The emergence of multispectral imagery technology has brought new opportunities for agricultural monitoring. It can acquire spectral information across multiple bands, revealing a rich tapestry of plant physiological and biochemical characteristics. However, the application of multispectral image analysis to peanut yield prediction faces a series of complex technical challenges. On the one hand, accurately and efficiently extracting features closely related to peanut yield from multispectral imagery and effectively integrating these features with environmental factors to construct an accurate yield prediction model remain key unresolved issues. Multispectral imagery data is large in volume and high in dimensionality, and complex correlations exist between information in different bands. Traditional feature extraction and fusion methods struggle to fully exploit the data's value. Furthermore, the peanut growth process is influenced by a combination of environmental factors, such as meteorological conditions and soil properties, which vary dynamically in time and space. Accurately and in real time in yield prediction, and adaptively adjusting the prediction model to improve accuracy and reliability, presents a technical bottleneck that urgently needs to be overcome. In summary, existing technologies for accurate, dynamic, and adaptive peanut yield prediction using multispectral image analysis are deficient, and innovative systems and methods are urgently needed to address these challenges.

[0004] Therefore, this application proposes a peanut yield intelligent prediction system and method based on multispectral image analysis. Summary of the Invention

[0005] The purpose of the present invention is to address the problem that the existing technology in the background technology is insufficient in using multispectral image analysis to achieve accurate, dynamic and adaptive prediction of peanut yield, and to propose an intelligent peanut yield prediction system and method based on multispectral image analysis.

[0006] In a first aspect, the present application provides an intelligent peanut yield prediction system based on multispectral image analysis, comprising:

[0007] The multispectral image acquisition module is configured to simultaneously acquire five-dimensional spectral data including visible light, red edge, near infrared, shortwave infrared, and thermal infrared bands through a swarm of drones equipped with multispectral imaging equipment. The thermal infrared band is used to capture the dynamics of plant transpiration.

[0008] A dynamic adaptive image processing unit, communicatively connected to the multispectral image acquisition module, includes:

[0009] Dynamic selector of band combinations based on growth stage identification;

[0010] 3D plant point cloud reconstruction engine integrating multi-view images;

[0011] A radiation correction device that incorporates real-time feedback of ambient light intensity;

[0012] Multimodal feature fusion analysis module, including:

[0013] The spatiotemporal feature associator performs convolutional fusion of spectral sequences and meteorological time series data;

[0014] Root-canopy coupling analyzer, which establishes a nonlinear mapping between soil parameters and canopy reflectance;

[0015] A core model for yield prediction based on deep reinforcement learning, integrating online learning and spatial deviation compensation mechanisms;

[0016] The adaptive yield prediction model is configured to generate an interactive visualization interface that integrates a three-dimensional plant model, a yield heat map, and an environmental factor tomography analysis, and outputs a yield prediction report with confidence intervals.

[0017] Optionally, the dynamic adaptive image processing unit further includes:

[0018] The leaf-level super-resolution reconstruction submodule uses a generative adversarial network (GAN) to enhance low-resolution images and generates high-resolution vein features guided by a pre-trained leaf texture library;

[0019] Shadow compensation algorithm, based on the solar azimuth calculation model and digital elevation data, automatically identifies and corrects the spectral values ​​of shadow areas in the image;

[0020] Leaf stomatal motion compensation submodule: Based on thermal infrared band imaging to invert stomatal conductance, a leaf transpiration-photosynthesis coupling model is established:

[0021]

[0022] Among them, gs is the stomatal conductance, T leaf is the leaf temperature, VPD is the vapor pressure difference, and a, b, and c are variety-specific parameters;

[0023] Multispectral polarization imaging unit: Integrated Stokes parameter measurement function, distinguishing healthy and diseased tissue by analyzing the polarization characteristics of the leaf surface, with the depolarization ratio of the diseased area reduced by 15%-25%;

[0024] Root acoustic wave sensing compensation mechanism: A piezoelectric sensor array deployed in the field captures the root water absorption acoustic emission signal (frequency range 1-10kHz) and performs time-frequency correlation analysis with the canopy spectral data.

[0025] Optionally, the multimodal feature fusion analysis module includes a canopy-soil coupling analysis unit, which is implemented by establishing the following correlation model:

[0026] Nonlinear mapping function based on near-infrared reflectance and soil moisture content;

[0027] Differential equation model of diurnal variation of canopy temperature and root water uptake rate;

[0028] The spatiotemporal coupling relationship between leaf nitrogen content spectral index and soil nitrate nitrogen concentration;

[0029] The canopy-soil coupling analysis unit further comprises:

[0030] Soil microbial activity inversion model: The nonlinear relationship between shortwave infrared spectral reflectance and soil respiration rate (R 2 >0.82), and a microbial carbon content predictor based on kernel ridge regression was constructed;

[0031] Root-mycorrhizal symbiosis analysis algorithm: Integrating X-ray fluorescence spectroscopy data, identifying mycorrhizal infection rates through calcium signaling characteristics, and establishing an optimization equation for phosphorus absorption efficiency:

[0032]

[0033] Wherein, EC is soil electrical conductivity, RAI is root activity index, MCR is mycorrhizal infection rate, and k is the constant term in the equation;

[0034] Modeling stress memory effect: A gated recurrent unit (GRU) is used to track the lagged impact of historical drought events on current leaf water potential. The weight matrix includes a time decay factor e -λt .

[0035] Optionally, the adaptive yield prediction model adopts a staged training strategy:

[0036] Pre-training phase: Use cross-regional historical datasets to train the basic network and establish global feature associations;

[0037] Fine-tuning stage: Using transfer learning technology, a small amount of sample data from the target plot is used to adjust the network parameters;

[0038] Online update phase: The model is continuously optimized through new data uploaded in real time by field IoT devices, and an elastic weight solidification algorithm is used to prevent catastrophic forgetting;

[0039] Knowledge distillation compression technology: In the pre-training phase, a teacher network (ResNet-152) is used to generate soft labels. In the fine-tuning phase, a student network (MobileNetV3) learns feature distribution using the KL divergence loss function.

[0040] Adversarial sample enhancement mechanism: Fourier domain perturbation noise is injected during the online update phase, and the perturbation amplitude is dynamically adjusted according to the model confidence:

[0041] δ=η·(1-p max )

[0042] Among them, p max is the maximum probability value predicted by the model, and η is the disturbance coefficient;

[0043] Neural Architecture Search (NAS) optimization: A population-based training (PBT) algorithm is used to automatically evolve the network depth and convolution kernel size to adapt to the morphological differences of different peanut varieties.

[0044] Optionally, it also includes edge-cloud collaborative computing architecture:

[0045] Edge computing nodes: deployed on smart gateways in the field, performing lightweight computations for image preprocessing and feature extraction;

[0046] Cloud computing center: runs deep neural network model training and complex spatiotemporal analysis;

[0047] Dynamic task allocator: adjusts the distribution strategy of processing tasks in real time based on network bandwidth and computing load;

[0048] The edge-cloud collaborative computing architecture further includes:

[0049] Quantum computing acceleration module: Deploy the quantum convolution layer in the cloud computing center to encode the spectral feature matrix into quantum states:

[0050]

[0051] where x i is the pixel coordinate, y i is the spectral intensity, N is the total number of image pixels, is the tensor product operator;

[0052] |x i>: pixel position quantum state, encoding coordinate information;

[0053] |y i >: spectral intensity quantum state, mapping multi-spectral values ​​into superposition state amplitudes;

[0054] Bionic communication protocol: Drawing on the electrical signal transmission mechanism of plants, a pulse code modulation strategy is designed to prioritize the transmission of data in areas where the NDVI gradient change rate is greater than 5% when bandwidth is limited;

[0055] Energy-aware task scheduler: Dynamically adjusts image sampling density based on the remaining battery power of the drone and establishes an energy consumption model:

[0056]

[0057] Among them, E total Total energy consumption, D is the flight distance, N is the number of shots, R is the resolution parameter, and α and β are coefficients.

[0058] On the other hand, the present application provides a method for intelligently predicting peanut yield based on multispectral image analysis, comprising the following steps:

[0059] S1. Dynamic spectrum acquisition stage:

[0060] According to the daily physiological changes of plants, automatic aerial photography is carried out from 09:00 to 11:00 every day;

[0061] Use multiple drones in formation for collaborative filming, and achieve centimeter-level track overlap through RTK positioning;

[0062] S2. Multi-scale feature fusion stage:

[0063] Macroscale: Calculate the spatial distribution matrix of NDVI, PSRI and 10 other spectral indices;

[0064] Mesoscopic scale: segmenting individual tree canopies and extracting leaf inclination distribution characteristics;

[0065] Microscopic scale: identifying the morphological fractal dimensions of leaf lesions;

[0066] S3. Modeling phase of production formation process:

[0067] Establish a light energy utilization efficiency model:

[0068] Y=ε×Σ(PAR×fAPAR)×WUE

[0069] Wherein, Y is yield, ε is variety characteristic parameter, PAR is photosynthetically active radiation, fAPAR is light energy absorption rate, and WUE is water use efficiency;

[0070] Construct a dynamic model of the source-sink relationship to simulate the distribution of photosynthetic products between stems, leaves and pods;

[0071] S4. Dynamic calibration of prediction results:

[0072] Introducing field-measured biomass as the observation value of Kalman filtering to correct the predicted value in real time;

[0073] Establish a feedback adjustment mechanism between yield forecast error and meteorological anomaly.

[0074] Optionally, in step S2, the multi-scale feature fusion adopts a feature pyramid network weighted by an attention mechanism:

[0075] The underlying network extracts texture detail features;

[0076] The middle-level network captures plant morphological characteristics;

[0077] The high-level network learns the spatial distribution pattern at the plot level;

[0078] Cross-scale feature fusion is achieved through a learnable attention weight matrix;

[0079] The multi-scale feature fusion adopts a biologically inspired feature pyramid network:

[0080] Leaf vein fractal feature extraction layer: calculate the box dimension of the leaf image and construct the fractal feature vector F fractal =[D0,D1,D2], corresponding to the fractal dimensions at 0°, 45°, and 90° respectively;

[0081] Canopy topology optimization analysis: Apply persistent homology theory to identify canopy pore structure and calculate Betti number distribution to characterize three-dimensional spatial connectivity;

[0082] Metabolic flux dynamic modeling: combining 13 C isotope labeling experimental data, and the partial differential equation for photosynthate transport was established:

[0083]

[0084] Where D is the diffusion coefficient, v is the phloem flow velocity, S is the source-sink intensity function, and C(x,t) is the concentration of photosynthetic products (mg / cm 3 ), S(x,t): source library function, t is time, x is spatial position.

[0085] Optionally, it also includes establishing an anti-interference training mechanism:

[0086] Injecting simulated noise data during model training includes:

[0087] a) Spectral distortion in simulated rainy weather;

[0088] b) Plant loss caused by crushing by generating equipment;

[0089] c) Create characteristic variations under different fertilization levels;

[0090] Adapting adversarial training strategies to improve model robustness.

[0091] The anti-interference training mechanism further includes:

[0092] Multiphysics coupled noise generation:

[0093] Electromagnetic interference simulation: superimposing random Lorentz noise spectrum in the visible light band;

[0094] Atmospheric turbulence distortion: Zernike polynomials are used to generate phase screen disturbances;

[0095] Mechanical vibration artifacts: Construct a 6-DOF rigid body motion blur model;

[0096] Biomorphological Generative Adversarial Network: The generator network embeds the L-system plant growth rules to generate adversarial examples that conform to the morphological laws of peanuts;

[0097] Brain-like spiking neural network verification: The test samples are input into the spiking neural network (SNN), and the model robustness is evaluated using the firing rate stability indicator.

[0098] Optionally, the output forecast result is displayed using augmented reality (AR) technology:

[0099] Use mobile terminal cameras to overlay predicted production information in real time;

[0100] Support gesture interaction to query plant growth parameters;

[0101] Provide agricultural operation simulation training function based on virtual reality;

[0102] The augmented reality (AR) technology specifically includes:

[0103] Photonic crystal chromaticity matching algorithm: Analyzes the ambient light spectrum distribution and dynamically adjusts the color coordinates of virtual information display to ensure that the color difference ΔE is less than 3 under different lighting conditions;

[0104] Tactile feedback enhancement mechanism: integrated piezoelectric tactile actuator that triggers a specific vibration pattern (frequency 120Hz, amplitude 0.3mm) when the user selects a low-yield area;

[0105] Olfactory decision-making assistance module: Deploys an electronic nose sensor array to detect the concentration of volatile organic compounds (VOCs) and triggers an AR warning mark when the C6 aldehyde content exceeds the standard.

[0106] Optionally, it also includes the establishment of a carbon sequestration assessment subsystem:

[0107] Based on the yield prediction results, the amount of biomass carbon sequestration can be inferred:

[0108] C=Y×CF×(1-MC)

[0109] Among them, Y is the yield, CF is the carbon content coefficient, and MC is the moisture content;

[0110] Calculate the spatial distribution of carbon sinks by combining plant coverage from drone aerial photography;

[0111] Generate visual certification reports that comply with carbon trading standards;

[0112] The carbon sink assessment subsystem further includes:

[0113] Root carbon deposition model: Combined with neutron imaging data, a three-dimensional root turnover model was established:

[0114]

[0115] Among them, C root is the root carbon deposition, ρ is the root density, r is the root radius, k is the decomposition rate, t is the time, T is the total length of the growth cycle, and z is the coordinate in the soil depth direction;

[0116] Methane flux inversion algorithm: By using the difference between thermal infrared and shortwave infrared bands, a CH4 emission estimation model for corn and peanut intercropping systems is constructed with an accuracy of ±15%;

[0117] Carbon credit smart contract: Based on blockchain technology, carbon sink data is encoded into non-fungible tokens (NFTs), which contain triple verification information of timestamp, geohash and spectral fingerprint.

[0118] Compared with the prior art, this application has at least one of the following beneficial technical effects:

[0119] A swarm of drones equipped with multispectral imaging equipment can capture five-dimensional spectral data, capturing comprehensive information about peanut plants. Thermal infrared bands can monitor plant transpiration dynamics. A dynamic adaptive image processing unit intelligently processes images for different growth stages, correcting for radiation and compensating for shadows and stomatal movement. This improves image quality and analysis accuracy, providing a reliable data foundation for subsequent predictions.

[0120] The multimodal feature fusion analysis module integrates spectral, meteorological, and soil data from multiple sources to explore the complex relationships between plants and their environment. The core yield prediction model, based on deep reinforcement learning and combined with a phased training strategy for adaptive yield prediction models, equips the model with powerful learning and adaptive capabilities, compensates for spatial deviations, generates accurate and visual yield forecast reports, and reduces prediction errors.

[0121] The intelligent prediction method works closely together at all stages. Dynamic spectral collection follows the physiological laws of plants to ensure data validity. Multi-scale feature fusion obtains information from macro, meso and micro levels, and uses an attention mechanism network to improve feature quality. Yield modeling considers light energy utilization and sink-source relationships. The prediction result calibration combines measured biomass and meteorological feedback to make the prediction results closer to reality.

[0122] In the edge-cloud collaborative computing architecture, edge nodes preprocess data, cloud computing centers handle complex computations, and a dynamic task allocator optimizes task distribution. Technologies such as quantum computing acceleration, biomimetic communications, and energy-aware scheduling improve computing, communication, and energy efficiency, ensuring stable system operation and reducing resource waste.

[0123] The forecast results are presented using augmented reality technology, providing an intuitive and interactive experience, allowing farmers to easily query information and simulate farming operations. The carbon sink assessment subsystem infers biomass carbon sequestration based on yield forecasts, calculates carbon sink distribution, and generates visual reports, helping farmers participate in carbon trading and achieve win-win results for both ecological and economic benefits.

[0124] This invention comprehensively utilizes multispectral image analysis, multimodal feature fusion, and intelligent algorithm technology to achieve precise monitoring of peanut growth and high-precision prediction of peanut yield, providing intuitive and effective decision-making support for agricultural production. At the same time, it optimizes system operating costs, taps into ecological and economic benefits, and effectively promotes the development of precision agriculture and sustainable agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0125] Figure 1 This is the principle block diagram of the peanut yield intelligent prediction system based on multispectral image analysis. DETAILED DESCRIPTION

[0126] The technical solution of the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0127] Example 1

[0128] like Figure 1 As shown, the peanut yield intelligent prediction system based on multispectral image analysis proposed in the present invention includes a multispectral image acquisition module, a dynamic adaptive image processing unit, a multimodal feature fusion analysis module, and an adaptive yield prediction model. Each part is described in detail below.

[0129] In this embodiment, the multispectral image acquisition module is configured to synchronously acquire five-dimensional spectral data including visible light, red edge, near infrared, short-wave infrared and thermal infrared bands through a group of drones equipped with multispectral imaging equipment, wherein the thermal infrared band is used to capture the dynamics of plant transpiration, specifically including multispectral image acquisition of peanut planting areas at different stages of the peanut growth cycle. The multispectral image acquisition module includes a multispectral camera and a drone or ground mobile platform equipped with the multispectral camera. The multispectral camera can simultaneously acquire at least five Spectral images of different bands include visible light (450-680nm), red edge (710-730nm), near infrared (780-1100nm), short-wave infrared (1550-1750nm) and thermal infrared (10.5-12.5μm). The thermal infrared is used to monitor the transpiration intensity of plants. The multispectral image acquisition module uses the multispectral imaging equipment carried by the drone swarm to synchronously acquire five-dimensional spectral data including visible light, red edge, near infrared, short-wave infrared and thermal infrared bands. The thermal infrared band can capture the dynamics of plant transpiration and provide key information for studying the physiological status of plants. The coordinated shooting of multiple drones and the realization of centimeter-level track overlap through RTK positioning ensure the comprehensiveness and high precision of data acquisition, so that the acquired image data can more accurately reflect the growth status of peanuts.

[0130] The dynamic adaptive image processing unit is in communication with the multispectral image acquisition module and includes a band combination dynamic selector based on growth stage identification, a 3D plant point cloud reconstruction engine that integrates multi-view images, and a radiation correction device that combines real-time feedback of ambient light intensity. Specifically, it includes:

[0131] The band selector based on the growth stage automatically activates the preset band combination according to the current growth cycle, enabling the visible light + red edge + near infrared combination in the seedling stage, and the near infrared + short wave infrared + thermal infrared combination in the pod setting stage;

[0132] The 3D point cloud reconstruction module generates a 3D plant model using multi-view UAV aerial images, accurately calculating the plant volume and spatial distribution density;

[0133] A real-time radiation correction module dynamically adjusts image exposure parameters using data from an onboard light intensity sensor. In this embodiment, the dynamic adaptive image processing unit also includes:

[0134] The leaf-level super-resolution reconstruction submodule uses a generative adversarial network to enhance low-resolution images, generating high-resolution vein features guided by a pre-trained leaf texture library. The shadow compensation algorithm automatically identifies and corrects the spectral values ​​of shadow areas in the image based on a solar azimuth calculation model and digital elevation data.

[0135] Leaf stomatal motion compensation submodule: Based on thermal infrared band imaging to invert stomatal conductance, a leaf transpiration-photosynthesis coupling model is established:

[0136]

[0137] Among them, g s is the stomatal conductance, T leaf is the leaf temperature, VPD is the vapor pressure difference, and a, b, and c are variety-specific parameters;

[0138] Multispectral polarization imaging unit: Integrated Stokes parameter measurement function, distinguishing healthy and diseased tissue by analyzing the polarization characteristics of the leaf surface, with the depolarization ratio of the diseased area reduced by 15%-25%;

[0139] Root acoustic wave sensing compensation mechanism: A piezoelectric sensor array deployed in the field captures the root water absorption acoustic emission signal and performs time-frequency correlation analysis with the canopy spectral data.

[0140] In addition, the dynamic adaptive image processing unit also includes:

[0141] The leaf-level super-resolution reconstruction submodule uses a generative adversarial network (GAN) to enhance low-resolution images and generates high-resolution vein features guided by a pre-trained leaf texture library;

[0142] Shadow compensation algorithm, based on the solar azimuth calculation model and digital elevation data, automatically identifies and corrects the spectral values ​​of shadow areas in the image;

[0143] Leaf stomatal motion compensation submodule: Based on thermal infrared band imaging to invert stomatal conductance, a leaf transpiration-photosynthesis coupling model is established:

[0144]

[0145] Among them, g s is the stomatal conductance, T leaf is the leaf temperature, VPD is the vapor pressure difference, and a, b, and c are variety-specific parameters;

[0146] Multispectral polarization imaging unit: Integrated Stokes parameter measurement function, distinguishing healthy and diseased tissue by analyzing the polarization characteristics of the leaf surface, with the depolarization ratio of the diseased area reduced by 15%-25%;

[0147] Root Acoustic Sensor Compensation Mechanism: A piezoelectric sensor array deployed in the field captures root water absorption acoustic emission signals (frequency range 1-10kHz) and performs time-frequency correlation analysis with canopy spectral data. The dynamic adaptive image processing unit is rich in functions. A dynamic band combination selector based on growth stage recognition selects the most representative band combinations based on the needs of different peanut growth stages, improving data processing efficiency and analysis accuracy. A 3D plant point cloud reconstruction engine that integrates multi-view imaging constructs accurate 3D plant models and intuitively displays plant morphology and structure. A radiation correction device with real-time feedback on ambient light intensity effectively eliminates interference from changes in ambient light intensity on image data, ensuring stable and reliable image quality. Furthermore, a leaf-level super-resolution reconstruction submodule, a shadow compensation algorithm, a leaf stomatal motion compensation submodule, a multispectral polarization imaging unit, and a root acoustic sensor compensation mechanism optimize image processing from various perspectives, further enhancing the accuracy and reliability of image analysis.

[0148] Multimodal feature fusion analysis module, including:

[0149] The spatiotemporal feature association unit convolves the spectral features with the time series of meteorological data to construct a composite feature vector that includes the temperature accumulation effect and the photoperiod.

[0150] Root development prediction unit, which inverts underground root biomass based on a nonlinear regression model of canopy spectral reflectance and soil electrical conductivity data;

[0151] The pest and disease evolution model uses an LSTM neural network to analyze the spatiotemporal diffusion patterns of abnormal spectral features in multi-temporal images;

[0152] The multimodal feature fusion analysis module includes a canopy-soil coupling analysis unit, which is implemented by establishing the following correlation model:

[0153] a) Nonlinear mapping function based on near-infrared reflectance and soil moisture content;

[0154] b) Differential equation model of diurnal variation of canopy temperature and root water uptake rate;

[0155] c) The spatiotemporal coupling relationship between leaf nitrogen content spectral index and soil nitrate nitrogen concentration;

[0156] The canopy-soil coupling analysis unit further comprises:

[0157] Soil microbial activity inversion model: The nonlinear relationship between shortwave infrared spectral reflectance and soil respiration rate (R 2 >0.82), and a microbial carbon content predictor based on kernel ridge regression was constructed;

[0158] Root-mycorrhizal symbiosis analysis algorithm: Integrating X-ray fluorescence spectroscopy data, identifying mycorrhizal infection rates through calcium signaling characteristics, and establishing an optimization equation for phosphorus absorption efficiency:

[0159]

[0160] Wherein, EC is soil electrical conductivity, RAI is root activity index, MCR is mycorrhizal infection rate, and k is the constant term in the equation;

[0161] Modeling stress memory effect: A gated recurrent unit (GRU) is used to track the lagged impact of historical drought events on current leaf water potential. The weight matrix includes a time decay factor e -λt The multimodal feature fusion analysis module convolutionally fuses the spectral sequence with the meteorological time series data through the spatiotemporal feature associator, comprehensively considering the relationship between spectral changes and meteorological factors during the growth process of peanuts; the root-canopy coupling analyzer establishes a nonlinear mapping between soil parameters and canopy reflectance, and deeply explores the intrinsic connection between soil and plants; the canopy-soil coupling analysis unit uses a series of correlation models, such as near-infrared reflectance and soil moisture content, diurnal changes in canopy temperature and root water absorption rate, leaf nitrogen content spectral index and soil nitrate nitrogen concentration, as well as soil microbial activity inversion model, root-mycorrhizal symbiosis analysis algorithm and stress memory effect modeling, etc., to reveal the complex relationship between plants, soil and environment during the growth process of peanuts from multiple levels, providing more comprehensive and accurate feature information for yield prediction.

[0162] The adaptive yield prediction model is built using a deep reinforcement learning framework and includes:

[0163] Online learning mechanism, which dynamically adjusts model weights through real-time field sensor data received;

[0164] Regional yield compensator, which builds a spatial autocorrelation matrix based on historical plot yield differences to correct local deviations in the forecast results;

[0165] Uncertainty quantification module, which uses the Monte Carlo Dropout method to calculate the confidence interval of the predicted value;

[0166] Visualization decision support module, which generates interactive 3D maps containing the following elements:

[0167] A virtual field roaming interface based on a 3D point cloud model of plants;

[0168] Overlay analysis function of yield prediction heat map and soil fertility layer;

[0169] Animation simulation of the spatiotemporal evolution of pest and disease risk areas.

[0170] The adaptive yield prediction model adopts a staged training strategy:

[0171] Pre-training phase: Use cross-regional historical datasets to train the basic network and establish global feature associations;

[0172] Fine-tuning stage: Using transfer learning technology, a small amount of sample data from the target plot is used to adjust the network parameters;

[0173] Online update phase: The model is continuously optimized through new data uploaded in real time by field IoT devices, and an elastic weight solidification algorithm is used to prevent catastrophic forgetting;

[0174] Knowledge distillation compression technology: In the pre-training phase, a teacher network (ResNet-152) is used to generate soft labels. In the fine-tuning phase, a student network (MobileNetV3) learns feature distribution using the KL divergence loss function.

[0175] Adversarial sample enhancement mechanism: Fourier domain perturbation noise is injected during the online update phase, and the perturbation amplitude is dynamically adjusted according to the model confidence:

[0176] δ=η·(1-p max )

[0177] Among them, p max is the maximum probability value predicted by the model, and η is the disturbance coefficient;

[0178] Neural Architecture Search (NAS) optimization: A population-based training (PBT) algorithm is used to automatically evolve network depth and convolution kernel size to adapt to morphological differences among different peanut varieties. The core yield prediction model based on deep reinforcement learning integrates online learning and spatial bias compensation mechanisms. This allows it to continuously learn from new data and adjust the prediction model in real time, effectively compensating for prediction bias caused by spatial differences between different plots and improving prediction accuracy. The adaptive yield prediction model employs a phased training strategy. The pre-training phase utilizes cross-regional historical datasets to establish global feature associations. The fine-tuning phase optimizes network parameters using a small sample of target plot data. The online update phase continuously optimizes the model using real-time data uploaded from field IoT devices. An elastic weight fixation algorithm is used to prevent catastrophic forgetting, enabling the model to rapidly adapt to changes across plots and growth stages. Furthermore, knowledge distillation compression techniques, adversarial sample enhancement mechanisms, and neural architecture search optimization methods are combined to further enhance the model's generalization, robustness, and adaptability. The model ultimately generates an interactive visualization interface that integrates a three-dimensional plant model, yield heatmaps, and environmental factor tomography analysis. The model also outputs a yield forecast report with confidence intervals, providing users with intuitive, accurate, and valuable yield forecast information.

[0179] In this embodiment, an edge-cloud collaborative computing architecture is also included:

[0180] Edge computing nodes: deployed on smart gateways in the field, performing lightweight computations for image preprocessing and feature extraction;

[0181] Cloud computing center: runs deep neural network model training and complex spatiotemporal analysis;

[0182] Dynamic task allocator: adjusts the distribution strategy of processing tasks in real time based on network bandwidth and computing load;

[0183] The edge-cloud collaborative computing architecture further includes:

[0184] Quantum computing acceleration module: Deploy the quantum convolution layer in the cloud computing center to encode the spectral feature matrix into quantum states:

[0185]

[0186] where x i is the pixel coordinate, y i is the spectral intensity, N is the total number of image pixels, is the tensor product operator;

[0187] |x i >: Pixel position quantum state, encoding coordinate information;

[0188] |y i >: spectral intensity quantum state, mapping multi-spectral values ​​into superposition state amplitudes;

[0189] Bionic communication protocol: Drawing on the electrical signal transmission mechanism of plants, a pulse code modulation strategy is designed to prioritize the transmission of data in areas with NDVI gradient change rates greater than 5% when bandwidth is limited;

[0190] Energy-aware task scheduler: Dynamically adjusts image sampling density based on the remaining battery power of the drone and establishes an energy consumption model:

[0191]

[0192] Among them, E total Total energy consumption, where D is the flight distance, N is the number of frames captured, R is the resolution parameter, and α and β are coefficients. In the edge-cloud collaborative computing architecture, edge computing nodes are deployed on field smart gateways to perform lightweight computations for image preprocessing and feature extraction, reducing data transmission and alleviating network burden. The cloud computing center runs deep neural network model training and complex spatiotemporal analysis, fully utilizing powerful computing resources for complex calculations. The dynamic task allocator adjusts the distribution strategy of processing tasks in real time based on network bandwidth and computing load to achieve optimal configuration of computing resources. In addition, quantum computing acceleration modules, bionic communication protocols, and energy-aware task schedulers further improve the system's computing efficiency, data transmission efficiency, and energy utilization efficiency, ensuring stable and efficient operation of the entire system.

[0193] Example 2

[0194] This embodiment improves the peanut yield intelligent prediction method based on multispectral image analysis, including the following steps:

[0195] S1. Dynamic Spectral Collection Phase: Based on the diurnal physiological changes of the plants, automated aerial photography was conducted daily between 9:00 AM and 11:00 AM. This phase employed collaborative photography by multiple drones, with RTK positioning achieving centimeter-level track overlap. Based on the diurnal physiological changes of the plants, automated aerial photography was conducted daily between 9:00 AM and 11:00 AM. During this time, when plants are most active, the acquired spectral data better reflects their true growth status. This phase employed collaborative photography by multiple drones, with RTK positioning achieving centimeter-level track overlap, ensuring comprehensive and accurate data collection, providing a reliable basis for subsequent analysis.

[0196] S2. Multi-scale feature fusion stage:

[0197] Macroscale: Calculate the spatial distribution matrix of NDVI, PSRI and 10 other spectral indices;

[0198] Mesoscopic scale: segmenting individual tree canopies and extracting leaf inclination distribution characteristics;

[0199] Microscopic scale: identifying the morphological fractal dimensions of leaf lesions;

[0200] The multi-scale feature fusion adopts the feature pyramid network weighted by the attention mechanism:

[0201] The underlying network extracts texture detail features;

[0202] The middle-level network captures plant morphological characteristics;

[0203] The high-level network learns the spatial distribution patterns at the plot level;

[0204] Cross-scale feature fusion is achieved through a learnable attention weight matrix;

[0205] The multi-scale feature fusion adopts a biologically inspired feature pyramid network:

[0206] Leaf vein fractal feature extraction layer: calculate the box-counting dimension of the leaf image and construct the fractal feature vector F fractal =[D0,D1,D2], corresponding to the fractal dimensions at 0°, 45°, and 90° respectively;

[0207] Canopy topology optimization analysis: Apply persistent homology theory to identify canopy pore structure and calculate Betti number distribution to characterize three-dimensional spatial connectivity;

[0208] Metabolic flux dynamic modeling: combining 13 C isotope labeling experimental data, and the partial differential equation for photosynthate transport was established:

[0209]

[0210] Where D is the diffusion coefficient, v is the phloem flow velocity, S is the source-sink intensity function, and C(x,t) is the concentration of photosynthetic products (mg / cm 3 ), S(x,t): source-storage function, t is time, x is spatial position. The multi-scale feature fusion stage extracts features from three scales: macro, meso, and micro. At the macro scale, the spatial distribution matrix of multiple spectral indices is calculated to grasp the growth status of the plot as a whole; at the meso scale, the canopy of a single plant is segmented and the distribution characteristics of leaf inclination are extracted to deeply analyze the growth characteristics of a single plant; at the micro scale, the morphological fractal dimensions of leaf spots are identified to accurately monitor the health status of the plant. At the same time, the feature pyramid network with weighted attention mechanism and the biologically inspired feature pyramid network are adopted to extract features at different levels through the bottom, middle, and high-level networks respectively, and the cross-scale feature fusion is realized through the learnable attention weight matrix, which fully mines the rich information in the multispectral image and provides more comprehensive and accurate feature support for yield prediction.

[0211] S3. Modeling phase of production formation process:

[0212] Establish a light energy utilization efficiency model:

[0213] Y=ε×∑(PAR×fAPAR)×WUE

[0214] Wherein, Y is yield, ε is variety characteristic parameter, PAR is photosynthetically active radiation, fAPAR is light energy absorption rate, and WUE is water use efficiency;

[0215] A dynamic sink-source relationship model was constructed to simulate the distribution of photosynthates between stems, leaves, and pods. During the yield formation modeling phase, a light use efficiency model and a dynamic sink-source relationship model were established. The light use efficiency model comprehensively considers factors such as cultivar characteristics, photosynthetically active radiation, light energy absorption rate, and water use efficiency to scientifically quantify the process of converting light energy into yield. The dynamic sink-source relationship model simulates the distribution of photosynthates between stems, leaves, and pods, providing a deeper understanding of the yield formation process from a physiological perspective and providing a solid theoretical foundation and model support for yield prediction.

[0216] S4. Dynamic calibration of prediction results:

[0217] Introducing field-measured biomass as the observation value of Kalman filtering to correct the predicted value in real time;

[0218] A feedback adjustment mechanism is established to balance yield forecast errors and meteorological anomalies. During the dynamic calibration phase of forecast results, field-measured biomass is used as an observation in the Kalman filter to correct forecast values ​​in real time. This approach also establishes a feedback adjustment mechanism to balance yield forecast errors and meteorological anomalies. This approach fully utilizes actual measurement data and meteorological information, allowing for timely adjustments to the forecast model, effectively improving the accuracy and reliability of forecast results and ensuring they are more consistent with actual production conditions.

[0219] In this embodiment, an anti-interference training mechanism is also established:

[0220] Injecting simulated noise data during model training includes:

[0221] a) Spectral distortion of simulated rainy weather

[0222] b) Plant loss caused by crushing by generating equipment

[0223] c) Create characteristic variations under different fertilization levels;

[0224] Adapting adversarial training strategies to improve model robustness.

[0225] The anti-interference training mechanism further includes:

[0226] Multi-physics coupled noise generation: a) Electromagnetic interference simulation: Random Lorentz noise spectra are superimposed on the visible light band; b) Atmospheric turbulence distortion: Phase screen perturbations are generated using Zernike polynomials; c) Mechanical vibration artifacts: A 6-DOF rigid-body motion blur model is constructed. An anti-interference training mechanism is established, injecting various simulated noise data during model training, including spectral distortion caused by rainy weather, plant loss caused by equipment crushing, and feature variation under different fertilization levels. Adversarial training strategies are also used to enhance model robustness. Furthermore, multi-physics coupled noise generation, biomorphic generative adversarial networks, and brain-inspired spiking neural networks for verification further enhance the model's adaptability to complex environments and interference factors, ensuring its stability and reliability in practical applications.

[0227] Biomorphological Generative Adversarial Network: The generator network embeds the L-system plant growth rules to generate adversarial examples that conform to the morphological laws of peanuts;

[0228] Brain-like spiking neural network verification: The test samples are input into the spiking neural network (SNN), and the model robustness is evaluated using the firing rate stability indicator.

[0229] In this embodiment, the yield forecast results are displayed using augmented reality (AR) technology: predicted yield information is superimposed in real time through the mobile terminal camera, gesture interaction is supported to query plant growth parameters, and a virtual reality-based farming operation simulation training function is provided;

[0230] The AR technology implementation details include:

[0231] Photonic crystal chromaticity matching algorithm: Analyzes the ambient light spectrum distribution and dynamically adjusts the color coordinates of virtual information display to ensure that the color difference ΔE is less than 3 under different lighting conditions;

[0232] Tactile feedback enhancement mechanism: integrated piezoelectric tactile actuator that triggers a specific vibration pattern (frequency 120Hz, amplitude 0.3mm) when the user selects a low-yield area;

[0233] Olfactory Decision-Assisted Module: This module deploys an electronic nose sensor array to detect volatile organic compound (VOC) concentrations, triggering an AR warning marker when C6 aldehyde levels exceed the standard. The yield forecast results are displayed using augmented reality technology, overlaying predicted yield information in real time via the mobile terminal camera. It supports gesture-based interactive querying of plant growth parameters and provides virtual reality-based agricultural operation simulation training. A photonic crystal chromaticity matching algorithm, a tactile feedback enhancement mechanism, and an olfactory decision-assisted module make the augmented reality experience more realistic and convenient, providing farmers and agricultural managers with an intuitive and interactive information display method, enabling them to keep abreast of peanut growth conditions and yield forecasts, and make scientific and reasonable agricultural decisions.

[0234] Example 3

[0235] This embodiment, based on embodiment 2, further includes establishing a carbon sink assessment subsystem:

[0236] Based on the yield prediction results, the amount of biomass carbon sequestration can be inferred:

[0237] C=Y×CF×(1-MC)

[0238] Among them, Y is the yield, CF is the carbon content coefficient, and MC is the moisture content;

[0239] Calculate the spatial distribution of carbon sinks by combining plant coverage from drone aerial photography;

[0240] Generate visual certification reports that comply with carbon trading standards;

[0241] The carbon sink assessment subsystem further comprises

[0242] Root carbon deposition model: Combined with neutron imaging data, a three-dimensional root turnover model was established:

[0243]

[0244] Among them, C root is the root carbon deposition, ρ is the root density, r is the root radius, k is the decomposition rate, t is the time, T is the total length of the growth cycle, and z is the coordinate in the soil depth direction;

[0245] Methane flux inversion algorithm: By using the difference between thermal infrared and shortwave infrared bands, a CH4 emission estimation model for corn and peanut intercropping systems is constructed with an accuracy of ±15%;

[0246] Carbon Credit Smart Contract: Based on blockchain technology, carbon sink data is encoded into non-fungible tokens (NFTs) that include triple verification information: timestamp, geohash, and spectral fingerprint. A carbon sink assessment subsystem has been established to infer biomass carbon sequestration based on yield forecasts. This system calculates the spatial distribution of carbon sinks using drone-based aerial photography of plant cover and generates a visual certification report that meets carbon trading standards. Furthermore, root carbon deposition models, methane flux inversion algorithms, and carbon credit smart contracts have further improved the carbon sink assessment system. This not only helps farmers understand the ecological benefits of peanut cultivation but also provides data support and technical assurance for participation in carbon trading, promoting sustainable agricultural development.

[0247] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art may make various alternative improvements and combinations to the above specific embodiments.

Claims

1. The peanut yield intelligent prediction system based on multispectral image analysis is characterized by: include: The multispectral image acquisition module is configured to simultaneously acquire five-dimensional spectral data including visible light, red edge, near infrared, shortwave infrared, and thermal infrared bands through a swarm of drones equipped with multispectral imaging equipment. The thermal infrared band is used to capture the dynamics of plant transpiration. A dynamic adaptive image processing unit, communicatively connected to the multispectral image acquisition module, includes: Dynamic selector of band combinations based on growth stage identification; 3D plant point cloud reconstruction engine integrating multi-view images; A radiation correction device that incorporates real-time feedback of ambient light intensity; Multimodal feature fusion analysis module, including: The spatiotemporal feature associator performs convolutional fusion of spectral sequences and meteorological time series data; Root-canopy coupling analyzer, which establishes a nonlinear mapping between soil parameters and canopy reflectance; A core model for yield prediction based on deep reinforcement learning, integrating online learning and spatial deviation compensation mechanisms; The adaptive yield prediction model is configured to generate an interactive visualization interface that integrates a three-dimensional plant model, a yield heat map, and an environmental factor tomography analysis, and outputs a yield prediction report with confidence intervals.

2. The peanut yield intelligent prediction system based on multispectral image analysis according to claim 1 is characterized in that: The dynamic adaptive image processing unit also includes: The leaf-level super-resolution reconstruction submodule uses a generative adversarial network (GAN) to enhance low-resolution images and generates high-resolution vein features guided by a pre-trained leaf texture library; Shadow compensation algorithm, based on the solar azimuth calculation model and digital elevation data, automatically identifies and corrects the spectral values ​​of shadow areas in the image; Leaf stomatal motion compensation submodule: Based on thermal infrared band imaging to invert stomatal conductance, a leaf transpiration-photosynthesis coupling model is established: Among them, g s is the stomatal conductance, T leaf is the leaf temperature, VPD is the vapor pressure difference, and a, b, and c are variety-specific parameters; Multispectral polarization imaging unit: Integrated Stokes parameter measurement function, distinguishing healthy and diseased tissue by analyzing the polarization characteristics of the leaf surface, with the depolarization ratio of the diseased area reduced by 15%-25%; Root acoustic wave sensing compensation mechanism: A piezoelectric sensor array deployed in the field captures the root water absorption acoustic emission signal and performs time-frequency correlation analysis with the canopy spectral data.

3. The peanut yield intelligent prediction system based on multispectral image analysis according to claim 1 is characterized in that: The multimodal feature fusion analysis module includes a canopy-soil coupling analysis unit, which is implemented by establishing the following correlation model: Nonlinear mapping function based on near-infrared reflectance and soil moisture content; Differential equation model of diurnal variation of canopy temperature and root water uptake rate; The spatiotemporal coupling relationship between leaf nitrogen content spectral index and soil nitrate nitrogen concentration; The canopy-soil coupling analysis unit further comprises: Soil microbial activity inversion model: A microbial carbon prediction model based on kernel ridge regression was constructed using the nonlinear relationship between shortwave infrared spectral reflectance and soil respiration rate; Root-mycorrhizal symbiosis analysis algorithm: Integrating X-ray fluorescence spectroscopy data, identifying mycorrhizal infection rates through calcium signaling characteristics, and establishing an optimization equation for phosphorus absorption efficiency: Wherein, EC is soil electrical conductivity, RAI is root activity index, MCR is mycorrhizal infection rate, and k is the constant term in the equation; Modeling stress memory effect: A gated recurrent unit is used to track the lagged impact of historical drought events on current leaf water potential, and the weight matrix includes a time decay factor e -λt .

4. The peanut yield intelligent prediction system based on multispectral image analysis according to claim 1 is characterized in that: The adaptive yield prediction model adopts a staged training strategy: Pre-training phase: Use cross-regional historical datasets to train the basic network and establish global feature associations; Fine-tuning stage: Using transfer learning technology, a small amount of sample data from the target plot is used to adjust the network parameters; Online update phase: The model is continuously optimized through new data uploaded in real time by field IoT devices, and an elastic weight solidification algorithm is used to prevent catastrophic forgetting; Knowledge distillation compression technology: In the pre-training phase, a teacher network is used to generate soft labels. In the fine-tuning phase, the student network learns the feature distribution through the KL divergence loss function. Adversarial sample enhancement mechanism: Fourier domain perturbation noise is injected during the online update phase, and the perturbation amplitude is dynamically adjusted according to the model confidence: δ=η(1-p max ) Among them, p max is the maximum probability value predicted by the model, and η is the disturbance coefficient; Neural architecture search optimization: A population-based training algorithm is used to automatically evolve the network depth and convolution kernel size to adapt to the morphological differences of different peanut varieties.

5. The peanut yield intelligent prediction system based on multispectral image analysis according to claim 1 is characterized in that: It also includes edge-cloud collaborative computing architecture: Edge computing nodes: deployed on smart gateways in the field, performing lightweight computations for image preprocessing and feature extraction; Cloud computing center: runs deep neural network model training and complex spatiotemporal analysis; Dynamic task allocator: adjusts the distribution strategy of processing tasks in real time based on network bandwidth and computing load; The edge-cloud collaborative computing architecture further includes: Quantum computing acceleration module: Deploy the quantum convolution layer in the cloud computing center to encode the spectral feature matrix into quantum states: where x i is the pixel coordinate, y i is the spectral intensity, N is the total number of image pixels, is the tensor product operator; |x i >: pixel position quantum state, encoding coordinate information; |y i >: spectral intensity quantum state, mapping multi-spectral values ​​into superposition state amplitudes; Bionic communication protocol: Drawing on the electrical signal transmission mechanism of plants, a pulse code modulation strategy is designed to prioritize the transmission of data in areas with NDVI gradient change rates greater than 5% when bandwidth is limited; Energy-aware task scheduler: Dynamically adjusts image sampling density based on the remaining battery power of the drone and establishes an energy consumption model: Among them, E total Total energy consumption, D is the flight distance, N is the number of shots, R is the resolution parameter, and α and β are coefficients.

6. The peanut yield intelligent prediction method based on multispectral image analysis is characterized by: The following steps are involved: S1. Dynamic spectrum acquisition stage: According to the daily physiological changes of plants, automatic aerial photography is carried out from 09:00 to 11:00 every day; Use multiple drones in formation for collaborative filming, and achieve centimeter-level track overlap through RTK positioning; S2. Multi-scale feature fusion stage: Macroscale: Calculate the spatial distribution matrix of NDVI, PSRI and 10 other spectral indices; Mesoscopic scale: segmenting individual tree canopies and extracting leaf inclination distribution characteristics; Microscopic scale: identifying the morphological fractal dimensions of leaf lesions; S3. Modeling phase of production formation process: Establish a light energy utilization efficiency model: Y=ε×∑(PAR×fAPAR)×WUE Wherein, Y is yield, ε is variety characteristic parameter, PAR is photosynthetically active radiation, fAPAR is light energy absorption rate, and WUE is water use efficiency; Construct a dynamic model of the source-sink relationship to simulate the distribution of photosynthetic products between stems, leaves and pods; S4. Dynamic calibration of prediction results: Introducing field-measured biomass as the observation value of Kalman filtering to correct the predicted value in real time; Establish a feedback adjustment mechanism between yield forecast error and meteorological anomaly.

7. The method for intelligent peanut yield prediction based on multispectral image analysis according to claim 6, characterized in that: In step S2, the multi-scale feature fusion adopts a feature pyramid network with weighted attention mechanism: The underlying network extracts texture detail features; The middle-level network captures plant morphological characteristics; The high-level network learns the spatial distribution pattern at the plot level; Cross-scale feature fusion is achieved through a learnable attention weight matrix; The multi-scale feature fusion adopts a biologically inspired feature pyramid network: Leaf vein fractal feature extraction layer: calculate the box dimension of the leaf image and construct the fractal feature vector F fractal =[D0,D1,D2], corresponding to the fractal dimensions at 0°, 45°, and 90° respectively; Canopy topology optimization analysis: Apply persistent homology theory to identify canopy pore structure and calculate Betti number distribution to characterize three-dimensional spatial connectivity; Metabolic flux dynamic modeling: combining 13 C isotope labeling experimental data, and the partial differential equation for photosynthate transport was established: Where D is the diffusion coefficient, v is the phloem flow velocity, S is the source-sink intensity function, and C(x,t) is the concentration of photosynthetic products (mg / cm 3 ), S(x,t): source library function, t is time, x is spatial position.

8. The method for intelligent peanut yield prediction based on multispectral image analysis according to claim 6, characterized in that: It also includes establishing an anti-interference training mechanism: Injecting simulated noise data during model training includes: a) Spectral distortion in simulated rainy weather; b) Plant loss caused by crushing by generating equipment; c) Create characteristic variations under different fertilization levels; Adopt adversarial training strategies to improve model robustness; The anti-interference training mechanism further includes: Multiphysics coupled noise generation: Electromagnetic interference simulation: superimposing random Lorentz noise spectrum in the visible light band; Atmospheric turbulence distortion: Zernike polynomials are used to generate phase screen disturbances; Mechanical vibration artifacts: Construct a 6-DOF rigid body motion blur model; Biomorphological Generative Adversarial Network: The generator network embeds the L-system plant growth rules to generate adversarial examples that conform to the morphological laws of peanuts; Verification of brain-like spiking neural network: The test samples are input into the spiking neural network, and the robustness of the model is evaluated by the firing rate stability index.

9. The method for intelligent peanut yield prediction based on multispectral image analysis according to claim 6, characterized in that: The production forecast results are displayed using augmented reality technology, which specifically includes: Photonic crystal chromaticity matching algorithm: Analyzes the ambient light spectrum distribution and dynamically adjusts the color coordinates of virtual information display to ensure that the color difference ΔE is less than 3 under different lighting conditions; Tactile feedback enhancement mechanism: integrated piezoelectric tactile actuator that triggers a specific vibration pattern when the user selects a low-yield area; Olfactory decision-making support module: deploys an electronic nose sensor array to detect the concentration of volatile organic compounds and triggers an AR warning mark when the C6 aldehyde content exceeds the standard.

10. The method for intelligent peanut yield prediction based on multispectral image analysis according to claim 6, characterized in that: It also includes the establishment of a carbon sink assessment subsystem: Based on the yield prediction results, the amount of biomass carbon sequestration can be inferred: C=Y×CF×(1-MC) Among them, Y is the yield, CF is the carbon content coefficient, and MC is the moisture content; Calculate the spatial distribution of carbon sinks by combining plant coverage from drone aerial photography; Generate visual certification reports that comply with carbon trading standards; The carbon sink assessment subsystem further includes: Root carbon deposition model: Combined with neutron imaging data, a three-dimensional root turnover model was established: Among them, C root is the root carbon deposition, ρ is the root density, r is the root radius, k is the decomposition rate, t is the time, T is the total length of the growth cycle, and z is the coordinate in the soil depth direction; Methane flux inversion algorithm: By using the difference between thermal infrared and shortwave infrared bands, a CH4 emission estimation model for corn and peanut intercropping systems is constructed with an accuracy of ±15%; Carbon credit smart contract: Based on blockchain technology, carbon sink data is encoded into non-fungible tokens, which contain triple verification information of timestamp, geohash and spectral fingerprint.

Images