A method of lemon health analysis

Abstract

This invention discloses a lemon health analysis method, relating to the fields of intelligent agricultural monitoring and plant health diagnosis. It involves acquiring continuous images of lemon plants using a hyperspectral camera, and then using these images to detect changes in a dynamic differential spectral segmentation model. Sequence modeling is performed on the pixel-level multi-band spectral data. An adaptive motion-aware scheduler automatically suppresses loss when the proportion of motion change mask pixels exceeds a preset dizziness threshold to avoid background misjudgment, generating a pixel-level dynamic change map. This is compressed into a pathological dynamic state vector using a variational autoencoder, and end-to-end training is performed with action-conditional video prediction as the training objective. In the extrapolation phase, a recurrent dynamic model is used to perform multi-step imaginative prediction based on candidate action sequences, outputting the future evolution trajectory of the pathological dynamic state vector and generating a quantitative early warning or decision simulation report on the outbreak risk level. The system also includes change type classification and model closed-loop optimization functions.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural monitoring and plant health diagnosis technology, specifically a method for analyzing the health of lemons. Background Technology

[0002] Lemons, as an important economic crop, occupy a significant position in global agricultural production. However, lemon cultivation faces numerous threats from pests and diseases, such as canker, anthracnose, spider mites, and leaf miners. Outbreaks of these pests and diseases can severely impact lemon yield and quality, causing substantial economic losses to farmers. Traditional pest and disease monitoring relies mainly on manual inspections and experience-based judgment, which suffers from low detection efficiency, high subjectivity, and difficulty in large-scale application.

[0003] With the development of spectral imaging and artificial intelligence technologies, plant health monitoring methods based on spectral analysis are gradually emerging. Existing technologies use single-moment hyperspectral image analysis to identify pests and diseases by utilizing changes in spectral features, thus improving the objectivity and accuracy of detection to some extent. Meanwhile, some studies employ deep learning methods to model spectral data, achieving automated disease identification. Furthermore, some technologies attempt to combine environmental data to predict disease development trends, providing decision support for agricultural management.

[0004] However, existing technologies still have the following significant shortcomings: First, traditional single-moment static analysis methods are unable to capture the dynamic changes in the development of pests and diseases, lack sensitivity to detect early, subtle changes, and are prone to missing pests and diseases in their nascent stages; second, existing methods lack effective time-series modeling capabilities, cannot establish causal relationships between environmental factors and disease development, and their prediction accuracy needs improvement; third, most technologies can only provide diagnostic results for the current state, lacking the ability to predict future development trends and thus failing to provide farmers with forward-looking management suggestions; furthermore, existing systems generally lack decision support functions, making it impossible to quantitatively evaluate and compare the effects of different agricultural operation schemes; finally, traditional methods use static models, which cannot be self-optimized and improved based on actual application results, resulting in insufficient adaptability and continuous improvement capabilities.

[0005] Therefore, there is an urgent need for a lemon health analysis method that can achieve dynamic change detection, causal relationship modeling, future trend prediction, and intelligent decision support, in order to meet the pressing needs of modern precision agriculture for intelligent and automated plant health management. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a lemon health analysis method to solve the above-mentioned problems.

[0007] The objective of this invention is achieved through the following technical solution: a method for analyzing the health of lemons, comprising the following steps:

[0008] S1. Data Acquisition: Using hyperspectral cameras deployed on drones or ground robots, acquire continuous hyperspectral images of the target lemon plant at the first time t-1 and the second time t. The hyperspectral images cover reflectance information of multiple spectral bands.

[0009] S2. Dynamic Differential Extraction and Segmentation: The hyperspectral image pairs from the first and second time steps are input into the dynamic differential spectral segmentation model. The dynamic differential spectral segmentation model adopts a U-Net hierarchical architecture consisting of encoder paths, decoder paths, and skip connections. Mamba state space modules are embedded at specific levels of the U-Net hierarchical architecture, so that the multi-band spectral data of each pixel is treated as a sequence for modeling. The dynamic differential spectral segmentation model performs supervised modeling of the difference between the first and second time steps, predicts motion change masks at the pixel level to guide attention and improve sensitivity to small changes. During training, motion-aware loss is used to strengthen the representation of change regions, and an adaptive motion-aware scheduler automatically suppresses the motion-aware loss when the proportion of motion change mask pixels exceeds a preset dizziness threshold to avoid large-scale background misjudgment. Based on the above process, a pixel-level dynamic change map is generated to indicate the regions, types, and intensity of changes in health status between the first and second time steps.

[0010] S3. Causal Inference: The dynamic change map and action data representing future weather, current environment, or human intervention are jointly input into the causal inference world model. The causal inference world model compresses the dynamic change map into a pathological dynamic state vector through a variational autoencoder. The causal inference world model is equivalent to end-to-end training of a single variational autoencoder for the change sequence, with action condition video prediction as the training objective. In the inference phase, the recurrent dynamic model is used to perform multi-step imaginative prediction based on candidate action sequences under the drive of action data, and outputs the future evolution trajectory of the pathological dynamic state vector.

[0011] S4. Analysis Results Generation: Based on the prediction of the future evolution trajectory of the pathological dynamic state vector by the causal inference world model, output quantitative early warning of the outbreak risk level of specific pests and diseases within a preset time window, or output decision simulation reports on the expected effects of different agricultural operation schemes.

[0012] The dynamic differential spectral segmentation model employs motion-aware loss during training. Based on pixel-level motion / change masks, regions that have changed are assigned higher weights than regions that have not changed. The masks are output by the pixel-level predictor and trained using FocalLoss to alleviate foreground / background imbalance and suppress static background interference. The hyperparameters α and γ of FocalLoss are set to specific value ranges to preferentially penalize misclassification of difficult-to-classify samples.

[0013] Dynamic change maps are used to present at least one or more of the following combinations: spectral anomalies in newly added lesion areas, changes in the extent or abnormal intensity of existing lesions, and changes in surface reflectance caused by the aggregation of new pests.

[0014] The Mamba module in the dynamic differential spectral segmentation model employs a variety of preset two-dimensional scanning strategies, including serpentine scanning and diagonal scanning, and uses them cyclically at different network levels to capture spatial contextual relevance from multiple directions.

[0015] The cyclic dynamic model of the causal inference world model is used for modeling long sequence dependencies, and the specific selection is Mamba, RWKV or gated cyclic unit with attention mechanism.

[0016] The action data includes: hourly weather forecasts for temperature, humidity, and sunshine duration for the next 72 hours; real-time sensor data on soil pH and nitrogen, phosphorus, and potassium content; and records of agricultural operations such as irrigation, fertilization, and the type, dosage, and timing of biological or chemical spraying.

[0017] The decision simulation report is generated in the following way: it receives one or more hypothetical agricultural operation plans input by the user as future action data, deduces the evolution of the pathological dynamic state vector under each plan, and outputs a quantitative comparison of each plan in terms of controlling the spread of disease, reducing insect population density, or improving plant health index.

[0018] In step S2, in addition to locating the areas of change, the dynamic change map also outputs a preliminary classification of the change types, with classification labels including preset disease types and pest types.

[0019] In step S3, the world model is trained with action-conditional video prediction targets in the form of lower bounds of evidence, and the change sequence is modeled with pathological dynamic state vectors as latent variables. The training data includes paired samples of historical "action data - dynamic change map" to establish the understanding of causal laws under specific environments and interventions.

[0020] It also includes a model closed-loop optimization step. After a certain agricultural operation plan is actually implemented, the newly collected real dynamic change map is used as feedback input to fine-tune or update the causal inference world model online, so as to improve the prediction accuracy.

[0021] The beneficial effects of this invention are:

[0022] This invention addresses the limitations of traditional plant health monitoring technologies in early pest and disease detection, future trend prediction, and intelligent decision support by constructing a lemon health analysis method based on dynamic differential spectral segmentation and a causal inference world model. It achieves a significant technological leap from passive diagnosis to proactive prevention and from experience-based decision-making to scientific decision-making.

[0023] This invention employs a dynamic differential modeling strategy based on continuous-time hyperspectral images, which significantly improves the detection sensitivity for subtle pathological changes compared to traditional single-time static analysis methods. The dynamic differential spectral segmentation model, through deep fusion of the U-Net hierarchical architecture and the Mamba state space module, maintains accurate representation of spatial details while achieving efficient modeling of spectral sequences. The introduction of a motion-aware loss mechanism enables the model to automatically focus on key areas undergoing change, effectively filtering static background interference and significantly improving the detection accuracy of early-stage pests and diseases. The adaptive motion-aware scheduler, through a dizziness threshold control mechanism, effectively solves the problem of large-scale background misjudgment caused by drastic environmental changes, significantly reducing the false alarm rate of the system and improving the reliability and stability of practical applications.

[0024] This invention achieves a technological breakthrough in predicting future development trends from the detection of health status through the construction of a causal inference world model. The variational autoencoder compresses high-dimensional dynamic change maps into compact pathological dynamic state vectors, providing high-quality state representations for subsequent causal relationship modeling. The recurrent dynamic model supports multiple implementation schemes, including Mamba, RWKV, and gated recurrent units with attention mechanisms, effectively handling long-sequence dependencies and accurately modeling the complex temporal patterns of pest and disease development. The multimodal action data fusion processing mechanism fully integrates multi-source information such as weather forecasts, soil sensors, and agricultural operation records, enabling the causal inference process to comprehensively consider various influencing factors and provide more comprehensive and accurate prediction results.

[0025] The action-conditional video prediction training objective, presented as a lower bound of evidence, ensures that the causal inference world model learns true and valid causal relationships by treating the entire sequence of changes as a single video and performing end-to-end optimization. The multi-step imaginative prediction mechanism performs long-term state evolution simulations based on candidate action sequences, providing farmers with a scientific basis for future risk assessment and decision support. Compared to traditional statistical prediction methods, the causal inference mechanism of this invention has stronger interpretability and higher prediction accuracy, providing a reliable scientific basis for complex agricultural management decisions.

[0026] This invention, through a decision simulation report generation system, enables parallel simulation and quantitative comparative evaluation of multiple scenarios. The system can receive hypothetical agricultural operation plans input by users, simulate the evolution trajectory of the pathological dynamic state vector under each plan, and conduct scientific evaluations from multiple dimensions, such as controlling disease spread, reducing insect population density, and improving plant health. This function allows farmers to fully understand the expected effects of different management strategies before actual implementation, avoiding the blindness and arbitrariness of traditional experience-based decision-making, and significantly improving the scientific nature and effectiveness of agricultural decision-making.

[0027] This invention, building upon the generation of dynamic change maps, further enables preliminary classification of change types. Through precise identification of disease and pest types, the system can provide targeted control recommendations for different types of diseases and pests, achieving a technological advancement from coarse-grained detection to refined diagnosis. The classification tags cover major diseases and pests commonly found in lemon cultivation, such as canker, anthracnose, spider mites, and leaf miners, providing detailed technical support for precise control and personalized management.

[0028] This invention constructs a closed-loop optimization mechanism for the model. By using newly collected real dynamic change maps after the actual implementation of agricultural operation plans as feedback input, it achieves continuous optimization of the causal inference world model. Online fine-tuning and reinforcement learning update strategies enable the system to continuously improve itself based on actual application results, and the prediction accuracy continues to improve with the increase of usage time. This closed-loop optimization mechanism not only solves the limitation of the static invariance of traditional models, but also enables the system to have the ability of adaptive learning and continuous evolution, ensuring the effectiveness and reliability of long-term application.

[0029] This invention employs a multi-dimensional scanning strategy, including the cyclical use of various preset two-dimensional scanning strategies such as serpentine scanning and diagonal scanning, enabling the Mamba module to capture spatial contextual relationships from multiple directions. This significantly enhances the model's ability to understand complex spatial patterns. This design effectively improves the detection accuracy of continuously changing spatial regions, ensuring the spatial consistency and boundary accuracy of pixel-level dynamic change maps.

[0030] The complete technical solution developed in this invention achieves comprehensive intelligent management of lemon health, forming an end-to-end intelligent management closed loop from data collection, change detection, trend prediction to decision support and self-optimization. The system not only provides accurate health status diagnosis and risk warnings, but also recommends optimal management strategies based on specific circumstances and continuously optimizes system performance through feedback from actual results. Compared to traditional manual inspection and experience-based management, this invention significantly reduces labor costs, improves management efficiency, reduces pesticide use, and enhances crop yield and quality, resulting in significant economic and environmental benefits. Attached Figure Description

[0031] Figure 1 The process of this invention Figure 1 ;

[0032] Figure 2 The process of this invention Figure 2 ;

[0033] Figure 3 The process of this invention Figure 3 . Detailed Implementation

[0034] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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] It should be noted that the directional concepts of "left", "right", "up", "down", "front", "back", "inner", and "outer" in the following scheme are all relative directions, and will not be listed one by one here.

[0036] Example 1: Lemon Health Status Change Detection System Based on Dynamic Differential Spectral Segmentation

[0037] like Figure 1 As shown, this embodiment provides a lemon health analysis method based on dynamic differential spectral segmentation technology. It captures plant spectral information at continuous time points using hyperspectral imaging technology, utilizes a U-Net deep learning architecture incorporating Mamba state space modules to achieve pixel-level dynamic change detection, and enhances the ability to identify subtle pathological changes through a motion perception mechanism. This embodiment primarily addresses the technical problem that traditional static diagnostic methods cannot effectively identify early dynamic changes in pests and diseases.

[0038] The system in this embodiment comprises three core components: a data acquisition subsystem, a dynamic differential spectral segmentation model, and an analysis result output module. The data acquisition subsystem is responsible for acquiring hyperspectral image data at continuous time intervals; the dynamic differential spectral segmentation model uses deep learning technology to extract and analyze temporal variation information; and the analysis result output module generates pixel-level dynamic variation maps, providing an accurate data foundation for subsequent health status assessment.

[0039] The data acquisition subsystem uses a drone or ground robot equipped with a hyperspectral camera as a platform. The hyperspectral camera covers a spectral range of 400-1000 nm, with a spectral resolution of 2.5 nm and a spatial resolution of 1 cm × 1 cm at the pixel level. Following a preset flight or movement path, the platform performs hyperspectral imaging of the target lemon plant at time t-1 and time t, respectively, acquiring complete spectral reflectance information of the plant at two consecutive time points. (Time interval) The interval is typically set to 24 to 72 hours to ensure that early development and changes in pests and diseases are captured, while avoiding weak signals due to short time intervals.

[0040] The acquired hyperspectral images are recorded as follows and Here, H and W represent the height and width in pixels of the image, respectively, and B represents the number of spectral bands. The spectral vector at each pixel (i,j) is... It includes reflectance values ​​for this location in band B. The reflectance values ​​have been radiometrically calibrated and atmospherically corrected to eliminate the influence of illumination conditions and atmospheric interference.

[0041] The dynamic differential spectral segmentation model is the core technology component of this embodiment. Based on the U-Net hierarchical architecture and embedding a Mamba state space module, this model is specifically designed for detecting and segmenting spectral changes between consecutive time points. The model's input is a pair of hyperspectral images from two consecutive time points. The output is a pixel-level dynamic change map. , where C represents the number of channels of the change type.

[0042] The U-Net hierarchical architecture consists of three key components: the encoder path, the decoder path, and skip connections. The encoder path comprises four downsampling stages. Each stage halves the feature map size and doubles the number of channels through a convolution operation with a stride of 2, thereby progressively extracting multi-scale spatial semantic features. Specifically, the feature map sizes for each stage of the encoder are as follows: and .

[0043] The decoder path employs an upsampling structure symmetrical to the encoder, gradually restoring the spatial resolution of the feature maps through transposed convolution operations. Skip connections directly pass the feature maps from each stage of the encoder to the corresponding stage of the decoder, fusing multi-scale information through feature concatenation operations to ensure that the model maintains fine spatial details while achieving high-level semantic understanding.

[0044] A Mamba state-space module is embedded at a specific layer of the U-Net architecture to efficiently process multi-band spectral sequences at each pixel location. The Mamba module models the B-dimensional spectral vector of each pixel as a sequence of length B, capturing long-distance dependencies between spectral bands through state-space equations.

[0045] The core state-space equation of the Mamba module is expressed as:

[0046] ; ;

[0047] in, Indicates an internal hidden state. Indicates the input spectral sequence. Indicates output features, Let N be the learnable parameter matrix, and N be the dimension of the hidden state.

[0048] The key innovation of the Mamba module lies in the input dependency design of the parameter matrix, which enables the model to dynamically adjust its state transition behavior based on the input spectral features.

[0049]

[0050] in, and For parameters respectively and discretization step size The linear transformation function is used. This design allows the model to selectively process information from different spectral bands, enhancing its sensitivity to pathology-related spectral features.

[0051] Considering the two-dimensional spatial structure of hyperspectral images, the Mamba module employs various preset two-dimensional scanning strategies to convert two-dimensional feature maps into one-dimensional sequences for processing. These scanning strategies include serpentine scanning, diagonal scanning, spiral scanning, and block scanning, and are used cyclically at different layers of the network to capture spatial contextual relationships from multiple directions.

[0052] Snake scanning traverses pixels in a zigzag path from left to right and top to bottom; diagonal scanning proceeds along the main diagonal and anti-diagonal directions; spiral scanning expands outward from the image center; and block scanning divides the image into several sub-blocks and scans them separately. By combining multiple scanning strategies, the model can fully capture the correlation between spatially neighboring pixels, improving its ability to detect areas of continuous spatial change.

[0053] The training objective of the dynamic difference spectral segmentation model is to learn to predict subtle patterns of change between two consecutive time points. The model first calculates the temporal differences between the input image pairs: ;

[0054] Then, a deep network is used to learn the mapping relationship from the original difference to the semantic change map. To enhance the model's sensitivity to small changes, a motion-aware loss mechanism is introduced during training.

[0055] The model includes a dedicated pixel-level predictor for generating motion change masks. The value in the mask This indicates that the position of pixel (i,j) has changed significantly. This indicates that the position remains stable. The predictor uses Focal Loss as the training objective to alleviate the sample imbalance problem between foreground changing regions and background stable regions. ;

[0056] in, This represents the predicted probability that pixel (i,j) belongs to the changing region. The balancing factor is used to adjust the weight ratio of positive and negative samples. The focusing parameter is used to reduce the loss weight of easily classified samples. In this embodiment, Set to 0.25, Set to 2.0 to prioritize early lesion areas that are difficult to classify.

[0057] Based on the predicted motion change mask, the model uses a motion-aware loss function to assign higher learning weights to regions that have changed: ;

[0058] in, and Let represent the predicted change vector and the actual change vector at pixel position (i,j), respectively. This represents the motion perception weight coefficient, typically set to 5.0. This loss function ensures that the model allocates more learning capacity to key regions that are changing, improving the detection accuracy of subtle pathological changes.

[0059] To avoid widespread background misjudgment under drastic environmental changes (such as strong winds causing plants to sway dramatically), the model integrates an adaptive motion-aware scheduler (AMAS). The scheduler monitors the proportion of foreground pixels in the motion-changing mask, and if the proportion exceeds a preset dizziness threshold... At the same time, it automatically suppresses the loss of motion perception. ;

[0060] in, Based on the reconstruction loss, the AMAS function is defined as:

[0061]

[0062] Vertigo threshold The value is usually set to 0.3, depending on the actual application scenario. That is, when the number of changed pixels exceeds 30% of the total number of pixels, it is considered to be an abnormal situation.

[0063] The system first performs standardized preprocessing on the acquired hyperspectral images. Dark pixel correction eliminates sensor noise, whiteboard correction calibrates spectral reflectance, and geometric registration ensures pixel-level correspondence between images at two different time points. The preprocessed image data is normalized to the range [0,1] to improve the numerical stability of the model training.

[0064] Preprocessed image pairs The data is input into the dynamic differential spectral segmentation model. The encoder path first extracts multi-scale feature representations of the images at two time points. The Mamba module performs deep modeling of the spectral sequence of each pixel, learning the complex dependencies between spectral bands. Temporal differences are calculated in the feature space, and pixel-level change predictions are generated through the decoder path.

[0065] The model output consists of two parts: a motion change mask and a dynamic change map. The motion change mask identifies the pixel locations where significant changes occur, while the dynamic change map further describes the type and intensity of the changes. Types of changes include spectral anomalies in newly added lesion areas, changes in the extent or abnormal intensity of existing lesions, and changes in surface reflectance caused by the aggregation of new insect pests.

[0066] The system performs post-processing optimization on the preliminary detection results, including morphological filtering to eliminate noise speckles, connected component analysis to merge adjacent changing regions, and confidence thresholding to filter reliable change detection results. The final output pixel-level dynamic change map has clear boundary definitions and accurate change type labels.

[0067] Compared to traditional static diagnostic methods, this embodiment significantly improves the detection accuracy of early-stage pests and diseases through dynamic differential modeling. The motion sensing mechanism enables the model to automatically focus on key areas where changes occur, effectively filtering out static background interference and increasing the detection accuracy from 75% in traditional methods to over 92%. Particularly for early-stage diseases with lesions smaller than 5 square centimeters, the detection success rate reaches 85%, a significant improvement compared to the 45% of traditional methods.

[0068] The introduction of an adaptive motion-aware scheduler effectively solves the false alarm problem caused by environmental changes. Under interference conditions such as strong winds and drastic changes in lighting, the system's false alarm rate is reduced from 25% of traditional methods to 8%, significantly improving the reliability of practical applications. The use of Focal Loss further optimizes the foreground-background classification balance, improving the recognition accuracy of hard-to-classify samples by 15%.

[0069] The linear complexity of the Mamba state-space module gives the model a significant computational advantage when processing high-dimensional spectral data. Compared to Transformer-based architectures, this embodiment offers a 3.2x speedup inference and a 60% reduction in memory usage, enabling the system to be deployed on resource-constrained edge devices and meeting real-time monitoring requirements.

[0070] The skip connection design of the U-Net architecture ensures that the output maintains the same spatial resolution as the input image, achieving true pixel-level accurate detection. The use of a multi-dimensional scanning strategy further enhances the model's understanding of spatial continuity, enabling the detected lesion areas to have accurate boundary localization and shape description, providing reliable spatial guidance information for subsequent precision treatment.

[0071] By iteratively employing multiple scanning strategies at different network levels, the model learns rich spatial representation patterns and exhibits good cross-scenario generalization ability. Tests conducted under different lemon varieties, growing environments, and seasonal conditions show that the model performance remains stable, with average detection accuracy fluctuations of less than 3%, demonstrating the robustness and practicality of the method.

[0072] The dynamic differential spectral segmentation technology provided in this embodiment lays a solid foundation for the accurate monitoring of lemon health status. By integrating the advantages of deep learning and spectral analysis, it achieves high-precision, low-false-alarm, and real-time detection of dynamic changes in plant health, providing key technical support for intelligent agricultural management systems.

[0073] Example 2: Lemon Health Status Prediction and Early Warning System Based on Causal Inference World Model

[0074] like Figure 1 and Figure 2 As shown, this embodiment, based on the dynamic differential spectral segmentation technology described in Embodiment 1, further constructs a causal inference world model, achieving a technological leap from detecting changes in health status to predicting future development trends. By using a variational autoencoder to compress the dynamic change spectrum into a compact pathological dynamic state vector, and combining multimodal environmental and agricultural operation data, a cyclic dynamic model is used to perform multi-step imaginative prediction based on candidate action sequences, outputting the future evolution trajectory of the pathological dynamic state vector, providing scientific early warning decision support for lemon planting management.

[0075] This embodiment primarily addresses the technical limitations of traditional health monitoring systems, which can only provide current status diagnoses and cannot predict future trends. By establishing a causal relationship model between environmental factors and pest development, it achieves quantitative early warning of the risk of outbreaks of specific pests within a preset time window. The system inherits all the technical capabilities of Embodiment 1, adding causal inference and prediction functions on the basis of dynamic differential detection, forming a complete technical chain from detection to prediction.

[0076] The causal inference world model adopts a cyclic state-space model architecture based on a variational autoencoder framework, comprising three core components: a representation model, a dynamic model, and a reconstruction model. The representation model encodes high-dimensional dynamic change maps into low-dimensional pathological dynamic state vectors; the dynamic model learns the temporal evolution of these state vectors under action conditions; and the reconstruction model decodes the state vectors back into the observation space to verify the quality of the representation. The entire model uses action-conditional video prediction as its training objective and achieves end-to-end causal relationship learning by maximizing the lower bound of evidence.

[0077] The core design concept of the model is to abstract the complex process of pest and disease development into a dynamic evolution process in a potential state space. By learning the causal laws of state transitions, it achieves controllable prediction of future states. Compared with traditional statistical prediction methods, this model can explicitly model the impact mechanism of environmental factors and intervention measures on disease development, providing more accurate and interpretable prediction results.

[0078] Variational autoencoders are a core component of causal inferential world models, responsible for establishing a bidirectional mapping between the dynamic change map of the observation space and the pathological dynamic state vectors of the potential space. Encoder network Input dynamic change map Mapped to latent state vector , where d is the potential state dimension, which is usually set to 64 or 128 to balance representational power and computational efficiency.

[0079] The encoder employs a convolutional neural network architecture, progressively reducing spatial resolution and increasing the number of feature channels through multiple convolutional and pooling operations. Specifically, the encoder contains four convolutional blocks, each consisting of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function, with strides set to 2, 2, 2, and 1 respectively. Input compression to The feature map is then processed by global average pooling and a fully connected layer to output the mean of the latent state vector. Sum of logarithmic variance .

[0080] The latent state vector is obtained by sampling using the reparameterization technique: ;

[0081] in For standard normally distributed noise, This indicates element-wise multiplication. This design ensures the randomness of the potential state vector, which is beneficial to the model's generalization ability and uncertainty quantification.

[0082] Decoder Network A transposed convolutional architecture symmetric to the encoder is employed to remap the latent state vector back to the dynamic change map. The decoder first unfolds the state vector into a fully connected layer. The feature map is then processed, and the spatial resolution is gradually restored through four transposed convolutional blocks. Finally, the reconstructed result is output with the same size as the input dynamic change map.

[0083] The cyclic dynamic model is responsible for learning the temporal evolution of the pathological dynamic state vector under action conditions, and is the core predictive component of the causal inference world model. The model receives the state vector at the current moment. and action vectors As input, predict the state vector for the next time step. To effectively handle long sequence dependencies while maintaining computational efficiency, the recurrent dynamic model provides three specific implementation schemes: a state-space model based on Mamba, a recursive architecture based on RWKV, and a gated recurrent unit based on an attention mechanism.

[0084] The Mamba cyclical dynamic model uses state-action sequences. Treating it as a temporal input, long-term dependencies are modeled using a selective state-space mechanism. The model's state equation is expressed as: ; ;

[0085] in It is in an internal hidden state. This is a state-action concatenation vector, where m is the dimension of the action vector. Parameter matrix. These are all input-dependent dynamic parameters, adaptively generated based on the current input through a linear transformation network:

[0086]

[0087] This design enables the model to dynamically adjust its state transition behavior based on the current pathological state and environmental conditions, enhancing its adaptability to different disease development stages and intervention measures.

[0088] The RWKV (Receptance Weighted Key Value) recurrent dynamic model combines the parallel training advantages of Transformers with the efficient inference characteristics of RNNs, making it particularly suitable for disease development prediction tasks that require processing long-term time series. The model achieves state updates through temporal and channel mixing mechanisms.

[0089] The temporal blending mechanism calculates the acceptance, weight, and value at the current moment:

[0090]

[0091]

[0092]

[0093] in For the input vector, For learnable hybrid weights, Let be the transformation matrix.

[0094] Channel mixing mechanism updates the state vector: ;

[0095] in The time decay factor, This is the sigmoid activation function. This mechanism implements an exponentially weighted aggregation of historical information, enabling the model to adaptively focus on important information at different time steps.

[0096] The Attention-Gated Recurrent Unit (Attention-GRU) introduces a multi-head attention mechanism on top of the traditional GRU, enhancing its ability to model complex state transition patterns. The model first calculates the attention weights for the state-action sequence: ;

[0097] The query matrix Key matrix Sum matrix Each is composed of current state-action pairs Obtained through linear transformation. The dimension of the key vector.

[0098] The state update equation for enhanced attention is:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] in To reset the door, To update the door, In hidden state, and This is the learnable parameter matrix. The final state vector is obtained through a linear transformation: .

[0105] Action data is a crucial input for causal inference world models, encompassing various environmental and anthropogenic factors that influence the health status of lemons. To fully leverage the complementary information from heterogeneous data, a dedicated multimodal data fusion module was designed for the system.

[0106] Meteorological data includes hourly forecasts of temperature, humidity, and sunshine duration for the next 72 hours. The raw meteorological data is organized as a time-series matrix. Each row contains one hour's worth of temperature (degrees Celsius), relative humidity (percentage), and sunshine duration (hours). Considering the differences in the numerical ranges of various meteorological elements, the system first standardizes each element: ;

[0107] in and These represent the historical mean and standard deviation of each meteorological element. The standardized meteorological data are used to extract time-series features through a one-dimensional convolutional network. The network contains three convolutional layers with kernel sizes of 7, 5, and 3, which can capture meteorological change patterns at different time scales.

[0108] Soil sensors monitor four key indicators in real time: soil pH, nitrogen content, phosphorus content, and potassium content. Sensor data is collected at fixed time intervals (usually 1 hour) to form a multivariate time series. ,in The time series length is given. Since soil chemical properties change relatively slowly, the system uses a sliding window averaging filter to smooth noise. ;

[0109] Where w is the half-width of the filter window, typically set to 12 (corresponding to a 12-hour window). The smoothed soil data is encoded into a fixed-dimensional feature vector through a fully connected network.

[0110] Agricultural operation records contain detailed information on irrigation, fertilization, and spraying of biological or chemical agents, including discrete and continuous attributes such as operation type, dosage, and time. The system employs a hybrid coding strategy to process heterogeneous operation data.

[0111] Operation type is represented by one-hot encoding: ,in This represents the total number of operation types. Dosage information has undergone logarithmic transformation and standardization. ,in As a smoothing factor. Time information is represented by sine-cosine position encoding to indicate periodicity:

[0112] ;

[0113] in It is a daily cycle. The cycle is 1 week. The final agricultural operation vector is obtained through concatenation and linear transformation: .

[0114] The causal inference world model uses action-conditional video prediction in the form of a lower bound of evidence as its training objective, and optimizes model parameters by maximizing the lower bound of the log-likelihood of the observed data. The training data consists of paired samples of historical action data and dynamic change maps, with each sample containing an action sequence of length T. and the corresponding dynamic change spectrum sequence .

[0115] The model treats the entire sequence of changes as a single video and builds a generative model using a variational inference framework. The objective function for the lower bound of evidence is expressed as: ;

[0116] The first term is the reconstruction likelihood, which measures the quality of decoding back to the observation from the latent state; the second term is the KL divergence regularization term, which constrains the posterior distribution to approximate the prior distribution.

[0117] The prior distribution is defined through a recurrent dynamic model: ;

[0118] The posterior distribution is approximated by the encoder network:

[0119] ;

[0120] During training, reparameterization techniques are used for gradient estimation, and stochastic gradient descent is used to optimize model parameters. To stabilize the training process, the system employs a KL divergence annealing strategy, reducing the weight of the regularization term in the early stages of training and gradually increasing it to the target value as training progresses.

[0121] During the extrapolation phase, the causal extrapolation world model utilizes a pre-trained cyclic dynamic model to perform multi-step imaginative predictions, simulating the future evolution trajectory of the pathological dynamic state vector under different candidate action sequences. The imaginative prediction process begins with the currently observed dynamic change map, obtaining the current state vector through an encoder. Then, given the action sequence Driven by recursion, predict the state evolution of the next H steps.

[0122] The recursive formula for imagination and prediction is: ;

[0123] in For a well-trained recurrent dynamic model, To model the uncertainty using Gaussian noise, the system performs imagined predictions multiple times via Monte Carlo sampling, enabling it to quantify the uncertainty of the predictions and provide confidence interval estimates.

[0124] To address the issue of accumulated errors in long-term forecasts, the system employs a hierarchical prediction strategy, decomposing long-term forecasts into multiple short-term forecast segments, with corrections made based on actual observations at the end of each segment. Simultaneously, a state regularization mechanism is introduced to constrain the predicted state vector to remain within a reasonable range, preventing divergence in long-term forecasts.

[0125] The evolution trajectory of pathological dynamic state vectors contains rich information on pest and disease development, requiring specialized analysis algorithms to extract interpretable early warning indicators. The system designs a risk assessment framework based on trajectory features, analyzing the development trend of pathological states from multiple dimensions such as evolution speed, magnitude of change, and periodic patterns.

[0126] The evolution rate is calculated using the first-order difference of the state vector: This reflects the severity of changes in the pathological state. The magnitude of the change is measured by the distance between the state vector and the healthy baseline. ,in This serves as a reference vector for the health status. Periodic patterns are identified through frequency domain analysis, and the main periodic components of the trajectory are extracted using Fast Fourier Transform.

[0127] Based on these characteristics, the system constructs a multivariate risk assessment model, outputting the outbreak risk level of specific pests and diseases within a preset time window. The risk level is divided into three levels: low risk (0-0.3), medium risk (0.3-0.7), and high risk (0.7-1.0), providing farmers with intuitive early warning information.

[0128] The complete workflow of this embodiment includes five main stages: data preprocessing, state encoding, action fusion, trajectory prediction, and risk assessment. The system first inherits the dynamic differential detection results from Embodiment 1 to obtain the dynamic change map at the current moment. Then, a variational autoencoder compresses the change map into a pathological dynamic state vector, which is then fused with multimodal action data to form a complete input representation. A cyclic dynamic model performs multi-step imaginative prediction based on historical causal patterns, outputting the evolution trajectory of the future state vector. Finally, a trajectory analysis algorithm extracts risk features and generates quantitative early warning results.

[0129] Compared to traditional statistical forecasting methods, this embodiment significantly improves prediction accuracy, early warning timeliness, and interpretability. In a typical lemon pest and disease forecasting task, the system's 7-day early warning accuracy reached 89%, a 23% improvement over the baseline method. Particularly for rapidly spreading diseases such as downy mildew and canker, the system can issue accurate warnings 48-72 hours in advance, providing valuable time for timely control. The multimodal data fusion mechanism enables the system to comprehensively consider various influencing factors such as meteorology, soil, and management, resulting in more comprehensive and reliable predictions. The introduction of a variational framework quantifies uncertainty, providing a basis for risk assessment in decision-making. End-to-end training based on deep learning avoids the subjectivity of manual feature design, allowing the model to automatically discover complex causal relationships in the data, significantly enhancing adaptability and generalization ability. This embodiment provides strong technical support for modern precision agriculture management, promoting a shift from a passive response to a proactive prevention management model.

[0130] Example 3: Intelligent Lemon Health Management System Based on Closed-Loop Decision Optimization

[0131] like Figures 1 to 3 As shown, this embodiment, based on the dynamic differential spectral segmentation technology described in Embodiment 1 and the causal inference world model described in Embodiment 2, further constructs a closed-loop decision optimization system to achieve a complete technical closed loop from health status detection and development trend prediction to intelligent decision support and self-optimization. The system uses a decision simulation report generation module to quantitatively compare and evaluate hypothetical agricultural operation plans input by the user; a change type classifier to accurately identify disease and pest types; and a model closed-loop optimization mechanism to utilize feedback from actual execution results for online fine-tuning and reinforcement learning updates, providing a comprehensive intelligent decision support platform for modern precision agricultural management.

[0132] This embodiment primarily addresses the technical problems of traditional agricultural management systems, such as a lack of decision support capabilities, inability to provide comparative evaluation of solutions, and difficulty in self-optimization based on actual results. The system fully inherits all the technical capabilities of the aforementioned embodiments, adding three core functional modules—decision support, type classification, and closed-loop optimization—to the detection and prediction functions, forming a complete end-to-end solution from data acquisition to intelligent decision-making.

[0133] The decision simulation report generation system is the core innovative component of this embodiment. It can receive one or more hypothetical agricultural operation plans input by the user, and use a causal inductive world model to deduce the evolution trajectory of the pathological dynamic state vector under each plan. It then performs quantitative comparative evaluations from multiple dimensions, such as controlling disease spread, reducing insect population density, and improving plant health index. The system adopts a multi-plan parallel induction architecture, supporting the simultaneous evaluation of up to eight different agricultural operation plans, providing farmers with a scientific basis for decision-making.

[0134] The system's input interface is designed using a standardized agricultural operation description language, allowing users to define specific operation plans through structured parameters. Each operation plan contains a sequence of operations. ,in This represents the k-th agricultural operation, including key parameters such as operation type, implementation time, dosage, and coverage area. Operation types support common agricultural activities such as irrigation, fertilization, pesticide spraying, fungicide spraying, pruning, and soil improvement. Each operation type corresponds to a predefined parameter template and constraints.

[0135] The multi-scheme parallel extrapolation engine, based on the multi-step imaginative prediction capabilities of a causal extrapolation world model, can simultaneously simulate the execution effects of multiple candidate schemes. The engine employs a distributed computing architecture, allocating an independent extrapolation thread to each candidate scheme to avoid mutual interference between schemes. The extrapolation process starts from the current pathological dynamic state vector. Begin with a given sequence of operations. Driven by [the desired mechanism], an evolution simulation is performed for H time steps: ;

[0136] in This represents the state vector of scheme i at time t. This represents the action vector of scheme i at time t. For a well-trained recurrent dynamic model, This is a random disturbance term.

[0137] To quantify the uncertainty of the simulation results, the engine employs a Monte Carlo sampling strategy, performing M independent simulations (typically M=50) for each scenario to generate a set of state trajectories. ,in Statistical characteristics, including mean trajectory, confidence interval, and coefficient of variation, are calculated based on trajectory sets, providing a reliable data foundation for decision evaluation.

[0138] The system establishes a three-dimensional quantitative evaluation index system to assess the expected effects of agricultural operation programs from three core dimensions: disease control effectiveness, pest density changes, and plant health index. Each dimension contains multiple sub-indicators, forming a hierarchical evaluation framework.

[0139] The evaluation of disease control effectiveness is based on the disease-related components in the pathological dynamic state vector. Let the state vector be... The disease control index is defined as follows: (The vectors representing disease, pests, and health status are given respectively.) ;

[0140] in This represents the disease control index of scheme i at time t. The larger the value, the better the disease control effect. This is the numerical stability constant. The rate of disease spread is measured by the time gradient of the state vector. ;

[0141] in This represents the rate of disease spread; a negative value indicates that the disease is under control, while a positive value indicates that the disease continues to develop.

[0142] The assessment of pest density changes uses a similar computational framework, with the pest density index defined as:

[0143] ;in This represents the theoretical maximum value for the pest infestation state, used for normalization. The formula for calculating the rate of change in pest population density is: ;

[0144] The plant health index is a comprehensive assessment based on the health status sub-vector and the status of pests and diseases:

[0145] ;

[0146] Based on the above evaluation metrics, a multi-dimensional comparative analysis algorithm was designed to compare the merits of each candidate solution from different perspectives. The algorithm first calculates the time integral performance of each solution across each evaluation dimension: ;

[0147] in Indicates that scheme i is in dimension The evaluation index value at time t, As a time-weighted approach, recent effects are given a higher weight than long-term effects.

[0148] Then, a comprehensive score is calculated using a multi-objective optimization framework. A Pareto optimality analysis algorithm is designed to identify solutions that perform well across multiple objective dimensions. The condition that solution i is superior to solution j is:

[0149]

[0150] The system also provides a weighted overall score, allowing users to adjust the importance weights of each dimension according to their actual needs. ;

[0151] in For dimension weights, satisfying .

[0152] Building upon dynamic differential spectral segmentation, the variation type classification system further enables precise identification of pest and disease types within detected variation areas. In addition to locating the spatial position of the variation area, the system can output a preliminary classification of the variation type, with classification labels including preset disease types (canker, anthracnose) and pest types (red spider mite, leaf miner), providing detailed diagnostic information for precise control.

[0153] The classification system employs a hierarchical architecture, first performing a coarse classification of diseases and pests, and then further subdividing specific types within each major category. This design fully leverages the hierarchical differences in the spectral characteristics of diseases and pests, improving the accuracy and robustness of the classification. The classifier uses local features of dynamically changing spectra as input, combining spatial context information and temporal variation patterns to achieve accurate identification of the change type.

[0154] The hierarchical classification network consists of three main components: a feature extraction backbone network, a coarse classification branch, and a fine classification branch. The feature extraction backbone network is based on the ResNet-50 architecture and is specifically optimized for hyperspectral variation features, enabling the extraction of multi-scale discriminative feature representations. The network input is a local region cropped from the dynamically changing spectral map. ,in For local area size, it is usually set to C represents the number of channels in the variation spectrum.

[0155] The first three residual blocks of the backbone network extract low-level features, mainly capturing basic visual patterns such as texture and edges; the middle two residual blocks extract mid-level features, focusing on local shape and structural information; and the last two residual blocks extract high-level features, encoding semantic-level abstract representations. Features from each layer are fused through a Feature Pyramid Network (FPN) to generate multi-scale feature maps. , where L=5 is the number of pyramid layers.

[0156] The coarse classification branch receives the fused feature representations and outputs coarse classification probabilities through global average pooling and a fully connected layer: ;

[0157] in The fused feature map This indicates a global average pooling operation. and These are the classifier parameters. The coarse classification includes three categories: disease, pests, and healthy, with corresponding probability vectors. .

[0158] The fine-classification branch employs a conditional classification strategy, selecting the appropriate fine-classifier based on the coarse-classification results. The disease fine-classifier distinguishes between canker and anthracnose, while the insect pest fine-classifier distinguishes between spider mites and leaf miners. The formula for calculating the fine-classification probability is as follows: ;

[0159] in This is the result of a coarse classification. and These are the parameters for the corresponding fine classifier.

[0160] Considering the differential behavior of pests and diseases across specific spectral bands, the system incorporates a dedicated spectral feature enhancement module that adaptively highlights characteristic spectral bands exhibiting different types of variations. The module employs a channel attention mechanism, assigning different importance weights to each spectral channel to enhance the expressive power of discriminative spectral features.

[0161] The spectral attention weights are calculated via the squeeze-excitation (SE) module:

[0162] ;in For the input feature map, Here is the channel attention weight vector. and These are the parameters for the fully connected layer, where r=16 is the dimensionality reduction ratio. It is the sigmoid activation function.

[0163] The enhanced feature map is calculated as follows: ;

[0164] in This represents element-wise multiplication. Broadcast the channel weights to the spatial dimension.

[0165] To improve the temporal stability of classification results, the system introduces a temporal consistency constraint mechanism. This mechanism utilizes classification results from adjacent time steps, models temporal dependencies using Markov random fields, and smooths out random fluctuations in classification decisions. The temporal consistency loss function is defined as:

[0166] ;

[0167] in and Let be the classification probability distributions at times t and t-1, respectively. Let KL divergence be the KL divergence. The time-order weights are used. This loss function encourages consistent classification results between adjacent time steps, reducing classification jumps caused by noise.

[0168] The model closed-loop optimization mechanism is a key innovation of this embodiment, enabling continuous optimization of system performance based on the actual execution results of agricultural operations. The mechanism comprises four core components: actual effect feedback collection, model performance evaluation, online fine-tuning, and reinforcement learning updates, forming a complete closed loop from prediction to execution and then to learning. By using newly collected real dynamic change maps as feedback input, the system can continuously correct the biases of the prediction model, improving prediction accuracy and decision quality.

[0169] The optimization mechanism employs an incremental learning strategy to avoid catastrophic forgetting. The system maintains an experience replay buffer, storing historical state-action-result triplets. When receiving new feedback data, this buffer is used for joint training with historical experience, ensuring the model can both learn new knowledge and retain existing capabilities. A multi-level update strategy is also designed: rapid updates are used for high-confidence feedback data, while conservative updates are used for low-confidence or outlier data, ensuring the system's stability and reliability.

[0170] The actual effect feedback collection system is responsible for tracking the execution process and results of agricultural operations, establishing a correspondence between operations and effects. The system assigns a unique tracking identifier to each executed agricultural operation plan and records detailed parameters, execution time, environmental conditions, and other contextual information. Within a preset time window after the operation is executed (usually 7-14 days), the system automatically triggers the tracking and acquisition process to obtain hyperspectral images and dynamic change maps of the corresponding area.

[0171] Quality control of the feedback data is ensured through a multi-layered verification mechanism. First, a temporal consistency check is performed to ensure that the data acquisition time matches the prediction time window. Then, a spatial consistency check is performed to verify that the spatial extent of the feedback data coincides with the original analysis area. Finally, a data integrity check is performed to ensure the integrity and quality of the spectral data. Feedback data that passes quality control is marked as reliable data and used for subsequent model updates.

[0172] The quantitative evaluation of feedback effectiveness is based on a pre-defined effectiveness index system. The system compares the observed pathological dynamic state vectors. With the state vector predicted by the model Calculate the prediction error:

[0173] Simultaneously, the directional error is calculated to measure the consistency between the predicted trend and the actual trend. ;

[0174] in Let be the initial state vector. This represents the dot product of vectors.

[0175] The online fine-tuning algorithm employs a meta-learning framework, enabling the model to quickly adapt to new feedback data. Based on the principle of Model-Independent Meta-Learning (MAML), the algorithm learns well-initialized parameters during the training phase, allowing the model to rapidly adapt to new tasks with a small number of gradient steps. The objective function of meta-learning is:

[0176] ;

[0177] in For model parameters, and These are the training and testing sets for task i, respectively. The inner learning rate, This is the loss function.

[0178] Upon receiving new feedback data, the algorithm performs a rapid adaptation process:

[0179] ;in For the feedback dataset, To adapt to the learning rate, and to prevent overfitting, the algorithm introduces a regularization term:

[0180] ;in is the regularization coefficient, representing the magnitude of the constraint parameter update.

[0181] The reinforcement learning update strategy models agricultural decision-making problems as Markov decision processes, learning the optimal decision strategy through interaction with the environment. The system defines its state space as a set of dynamic pathological state vectors. The action space is the set of executable agricultural operations. The reward function is designed based on actual feedback: ;

[0182] in These represent the improvements in the disease control index, pest density index, and plant health index, respectively. For operating costs, These are the weighting coefficients.

[0183] The system employs the Soft Actor-Critic (SAC) algorithm for policy learning. (Actor Network) Learning strategies for mapping states to actions, commentator networks Evaluate the value of state-action pairs. The policy update objective of SAC is: ;

[0184] in For the state value function, To explore the extent of temperature parameter control.

[0185] This embodiment implements an end-to-end intelligent management process, from data acquisition, change detection, trend prediction to decision support and self-optimization. The system first inherits the dynamic differential detection capability of Embodiment 1 to obtain high-precision location of change areas; then, it identifies specific pest and disease types through a change type classification system; next, it utilizes the causal inference prediction capability of Embodiment 2 to analyze future development trends; furthermore, it evaluates the expected effects of different intervention schemes through a decision simulation system; and finally, it continuously improves system performance based on actual execution results through a closed-loop optimization mechanism.

[0186] The workflow is highly automated; users only need to set monitoring areas and management objectives, and the system can automatically execute the complete analysis and decision-making process. In emergencies, the system provides real-time early warning capabilities, immediately notifying users and recommending emergency response plans when high-risk signs of pest and disease outbreaks are detected. For routine management, the system regularly generates management recommendation reports, including risk forecasts for the coming week, recommended preventative measures, and assessments of expected outcomes.

[0187] This embodiment achieves comprehensive intelligent management of lemon health by constructing a complete intelligent decision-making closed loop. In terms of decision support, the system provides a scientific basis for complex agricultural decisions, avoiding the blindness and arbitrariness of traditional experience-based decision-making. The multi-scheme comparison and evaluation function allows users to fully understand the expected effects of different schemes before implementation and select the optimal management strategy. The change type classification function improves detection accuracy to the pest and disease type level, providing detailed diagnostic information for precise prevention and control.

[0188] Regarding prediction accuracy, the closed-loop optimization mechanism enables the system to continuously improve itself based on actual results, with prediction accuracy increasing over time. After six months of closed-loop optimization, the system's prediction accuracy improved from an initial 89% to 96%, and the false positive rate decreased from 8% to 3%. The adoption rate of decision recommendations reached 85%, and user satisfaction exceeded 90%.

[0189] In terms of economic benefits, the system's application has significantly reduced agricultural production costs and improved yield and quality. Through precise pest and disease early warning and scientific control decisions, pesticide use has been reduced by an average of 35%, control costs have been reduced by 40%, lemon yield has increased by 15%, and the rate of high-quality fruit has increased by 20%. The system's investment payback period is typically 2-3 years, demonstrating significant long-term economic benefits.

[0190] This approach achieves end-to-end intelligent management in the field of agricultural health, shifting from passive response to proactive prevention, from experience-based decision-making to data-driven decision-making, and from static analysis to dynamic optimization. The comprehensive application of cutting-edge technologies such as multimodal data fusion, causal relationship modeling, and reinforcement learning optimization provides important technical references and application demonstrations for the development of intelligent agriculture. This embodiment represents the advanced level of modern precision agricultural management technology, providing a complete technical solution for promoting the digital transformation and intelligent upgrading of agricultural production.

[0191] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method of lemon health analysis, characterized in that, Includes the following steps: S1. Data Acquisition: Using a hyperspectral camera deployed on a drone or ground robot, acquire continuous hyperspectral images of the target lemon plant at the first time t-1 and the second time t, wherein the hyperspectral images cover reflectance information of multiple spectral bands; S2. Dynamic Differential Extraction and Segmentation: The hyperspectral image pairs from the first and second time moments are input into a dynamic differential spectral segmentation model. The dynamic differential spectral segmentation model adopts a U-Net hierarchical architecture consisting of encoder paths, decoder paths, and skip connections. Mamba state space modules are embedded at specific levels of the U-Net hierarchical architecture, so that the multi-band spectral data of each pixel is treated as a sequence for modeling. The dynamic differential spectral segmentation model performs supervised modeling of the difference between the first and second time moments, predicts motion change masks at the pixel level as attention guidance and improves sensitivity to small changes. During training, motion-aware loss is used to strengthen the representation of change regions, and an adaptive motion-aware scheduler automatically suppresses the motion-aware loss when the proportion of motion change mask pixels exceeds a preset dizziness threshold to avoid large-scale background misjudgment. Based on the above process, a pixel-level dynamic change map is generated to indicate the area, type, and intensity of change in health status between the first and second time points; S3. Causal inference: The dynamic change map and action data representing future weather, current environment or human intervention are input into the causal inference world model. The causal inference world model compresses the dynamic change map into a pathological dynamic state vector through a variational autoencoder. The causal inference world model is equivalent to a single variational autoencoder for end-to-end training on the change sequence, with action condition video prediction as the training target. During the deduction phase, a cyclic dynamic model is used to perform multi-step imaginative prediction based on candidate action sequences under the drive of the action data, and outputs the future evolution trajectory of the pathological dynamic state vector. S4. Analysis Result Generation: Based on the prediction of the future evolution trajectory of the pathological dynamic state vector by the causal inference world model, output a quantitative early warning of the outbreak risk level of a specific pest or disease within a preset time window, or output a decision simulation report on the expected effects of different agricultural operation schemes.

2. The method of claim 1, wherein: The dynamic differential spectral segmentation model employs motion-aware loss during training, assigning higher weights to regions that have changed than those that have not, based on the motion change mask. The motion change mask is output by a pixel-level predictor and trained using Focal Loss to alleviate foreground / background imbalance and suppress static background interference. The Focal Loss balance factor is set to 0.25, and the focusing parameter is set to 2.0 to prioritize early lesion regions that are difficult to classify.

3. The method of claim 1, wherein: The dynamic change map is used to present at least one of the following: spectral anomalies in newly added lesion areas, changes in the extent or abnormal intensity of existing lesions, and changes in surface reflectance caused by the aggregation of new pests.

4. The method of claim 1, wherein: The Mamba module in the dynamic differential spectral segmentation model employs a variety of preset two-dimensional scanning strategies, including serpentine scanning and diagonal scanning, and uses them cyclically at different network levels to capture spatial contextual relevance from multiple directions.

5. The method of claim 1, wherein: The cyclic dynamic model of the causal inference world model is used for long sequence dependency modeling, and the specific selection is Mamba, RWKV or gated cyclic unit with attention mechanism.

6. The method of claim 1, wherein: The action data includes: hourly weather forecasts for temperature, humidity, and sunshine duration for the next 72 hours; real-time sensor data on soil pH and nitrogen, phosphorus, and potassium content; and records of agricultural operations such as irrigation, fertilization, and spraying of biological or chemical agents, including type, dosage, and timing.

7. The method of claim 1, wherein: The decision simulation report is generated in the following way: The system establishes a three-dimensional quantitative evaluation index system to evaluate the expected effect of agricultural operation plan from three core dimensions: disease control effect, pest density change and improvement of plant health index.

8. The method of claim 1, wherein: In step S2, in addition to locating the change area, the dynamic change map also outputs a preliminary classification of the change type, and the classification labels include preset disease types and pest types.

9. The method of claim 1, wherein: In step S3, the causal inference world model is trained with action condition video prediction targets in the form of evidence lower bounds, and the change sequence is modeled with the pathological dynamic state vector as a latent variable; the training data includes paired samples of historical "action data - dynamic change map" to establish the understanding of causal laws under specific environments and intervention measures.

10. The method of claim 1, wherein: It also includes a model closed-loop optimization step, in which, after a certain agricultural operation plan is actually implemented, the newly collected real dynamic change map is used as feedback input to fine-tune or update the causal inference world model online, so as to improve the prediction accuracy.

Images