Information detection active resampling method and system based on intelligent perception and source-destination decoupling assimilation

Abstract

The application discloses a kind of information detection active resampling methods and systems based on intelligent perception and source and sink decoupling assimilation, it is related to agricultural information processing technical field, method includes: obtaining unmanned vehicle multimodal observation data, position attitude parameter, sampling time stamp and canopy structure parameter;Based on the modal time lag parameter of phenological phase and canopy structure parameter constraint constructs observation operator, generates disease risk intensity field;Establish the propagation model including spatial heterogeneity resistance parameter field and time-varying infection source intensity field, carry out inversion assimilation, forward prediction and uncertainty quantification;Determine resampling trajectory in combination with propagation blocking boundary, cognitive uncertainty and expected observation quality, so that sampling is preferentially used in the area that can reduce cognitive uncertainty, and new observation is returned to update.Through the technical scheme of the application, the prediction of the spatiotemporal evolution trend of the disease is realized, and the accuracy of risk assessment, the pertinence of resampling and the stability of closed-loop updating in complex farmland environment are improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural information processing and intelligent agricultural machinery sensing and control technology, and in particular to an active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation, and an active resampling system for information detection based on intelligent sensing and source-destination decoupling assimilation. Background Technology

[0002] Crop disease monitoring is a crucial component of smart agriculture. Current technologies typically employ drones equipped with visible light, multispectral, or thermal infrared sensors to inspect farmland and then use the acquired images to identify, classify, or determine the distribution of diseases. These solutions mostly utilize visual recognition combined with path coverage to achieve disease inspection, which can improve farmland inspection efficiency and reduce the cost of manual inspections to some extent.

[0003] However, most existing disease monitoring schemes rely on image recognition results at the current observation time as the basis for judgment. Disease identification is typically treated as a relatively isolated static image classification or target recognition problem, focusing only on extracting and judging the crop's appearance at the current moment. Essentially, it remains a static identification or phenomenon record, lacking the ability to model the spatiotemporal evolution of diseases in farmland. Existing technologies typically struggle to use existing observational information to predict the subsequent development trend of diseases, reflect the dynamic laws of disease propagation within the canopy, and characterize the impact of implicit states such as crop spatial heterogeneity and resistance on disease transmission. Therefore, existing technologies can usually only passively identify diseases that have already occurred or manifested, making it difficult to achieve forward-looking inspections and active resampling in high-risk areas.

[0004] Furthermore, existing schemes typically assume that crop growth status and disease resistance are relatively uniform at the farmland scale, lacking effective characterization and inversion mechanisms for spatial heterogeneity caused by varietal differences, local fertility gradients, soil conditions, and microclimate disturbances. In actual farmland environments, crop sensitivity to diseases and their transmission responses are usually not uniform. Without modeling and updating these implicit state parameters, high false alarm rates, high false negative rates, and insufficient prediction stability are likely to occur under complex plot conditions.

[0005] Furthermore, existing disease prediction methods often struggle to distinguish whether a high-risk state stems from increased local infection sources or decreased host resistance. This leads to a discriminative coupling between infection source location, infection pressure, and host vulnerability, limiting the ability of subsequent resampling to pinpoint truly critical areas. Even when incorporating diffusion models, existing disease transmission retrieval schemes often parameterize transmission drivers holistically, failing to separately constrain latent variables of two distinct natures: spatial heterogeneity of host resistance and spatiotemporal enhancement of infection sources. In particular, they lack mechanisms to spatially constrain host resistance parameter fields using prior farmland structures such as soil moisture distribution and crop growth distribution. Consequently, existing schemes typically only yield overall transmission fitting results, failing to establish a decoupled update process for identifying agricultural transmission mechanisms.

[0006] Meanwhile, different observation modalities exhibit stage-specific time lags in their responses to disease progression, and crop phenological changes cause drift in the mapping relationship between multimodal observations and disease state variables. Simply performing static calibration or simple drift correction can easily lead to systematic biases in the construction of the risk field. Existing technologies typically treat visible light, multispectral, and thermal infrared information as parallel observations at the same time and semantic level, fusing them together. At most, they only perform static adjustments to modal weights, failing to explicitly model the response time lags of each modality relative to disease state variables at different phenological stages. Therefore, while existing solutions can achieve multimodal collaborative identification, they struggle to ensure temporal alignment of each modality in the state space, easily leading to systematic temporal biases in the construction of disease risk state quantities.

[0007] Moreover, if the resampling path planning is based solely on risk values ​​and uncertainties without considering the impact of flight altitude, ground speed, attitude stability, and positioning errors on observation quality, a situation may arise where the theoretical information gain on the path is high but the actual sampling quality is low, affecting the stability of subsequent assimilation.

[0008] Furthermore, while some existing technologies attempt to trigger re-shooting or additional sampling when recognition confidence is low, their triggering criteria are usually quite simple, often relying on image differences, classification confidence, or single recognition results as the basis for judgment. They fail to further decompose prediction uncertainty and cannot distinguish between accidental uncertainty caused by environmental noise and measurement noise, and cognitive uncertainty caused by insufficient model knowledge and inadequate parameter estimation. As a result, existing solutions are prone to redundant sampling in high-noise areas, leading to wasted sampling resources. Moreover, they struggle to prioritize flight and sampling capabilities under limited endurance conditions to areas with truly high information gain, thus limiting the resampling's relevance and sampling efficiency.

[0009] Furthermore, existing inspection path planning technologies often employ preset routes, fixed waypoints, or geometric full coverage methods, typically failing to incorporate disease risk prediction results and uncertainty distributions for constrained optimization. This makes it difficult to dynamically generate resampling trajectories based on resampling value. Simultaneously, the algorithm decision-making process and flight control execution process in existing inspection systems are often disconnected. Path planning results struggle to be timely mapped to flight control commands and sensor trigger signals, resulting in insufficient registration accuracy between sampling timestamps, position, and attitude information. This leads to mapping errors between observation data and spatial position, consequently affecting the stability of subsequent assimilation updates.

[0010] Therefore, existing technologies have at least the following three core problems: First, there is a lack of an observation mechanism that can uniformly map multimodal observations with asynchronous responses at different phenological stages into disease state quantities, resulting in a time mismatch in the construction of the disease risk field; Second, the lack of a decoupling mechanism for the transmission mechanism that distinguishes between spatial differences in host resistance and the spatiotemporal enhancement contribution of the source of infection under the a priori constraint of farmland structure makes it difficult to identify the causes of high-risk areas; Third, the lack of a closed-loop resampling mechanism that simultaneously applies observation quality to information gain assessment and actual sampling triggering makes it difficult to convert theoretically high-value sampling points into practically usable observations. Summary of the Invention

[0011] To address the aforementioned issues, this invention provides an active resampling method and system for information detection based on intelligent sensing and source-sink decoupling assimilation. First, a phenologically conditional time-delay observation operator is used to uniformly map visible light, multispectral, and thermal infrared observations with asynchronous response characteristics at different phenological stages into a disease risk intensity field. The modal time-delay parameters of this phenologically conditional time-delay observation operator are also constrained by canopy structure parameters. Second, the host resistance spatial heterogeneity parameter field is updated under prior constraints of farmland structure, and the infection source time-varying intensity field is updated under local enhancement constraints and propagation blocking boundary constraints, achieving source-sink decoupling in the agricultural propagation mechanism. Third, resampling is performed only in areas that can reduce the joint posterior uncertainty of the parameter field under quality constraints. Finally, the same observation quality factor is simultaneously introduced into information gain evaluation and sampling trigger control to determine whether candidate observations are sufficient to form effective update constraints on the parameter field, ensuring that new observations theoretically have parameter contraction value and meet effective sampling conditions during execution, thus forming a closed-loop active resampling framework for propagation mechanism identification.

[0012] To achieve the above objectives, this invention provides an active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation, comprising: 1. Construction of unified disease status based on phenological conditional time-delay observation operator: acquire multimodal observation data collected by UAV during the inspection of the target farmland area, as well as the position parameters, attitude parameters and sampling timestamps corresponding to the multimodal observation data, and acquire the canopy structure parameters corresponding to the multimodal observation data; The multimodal observation data are time-synchronized, spatially registered, and phenological stage marked. An observation operator is constructed based on the modal time delay parameters constrained by the phenological stage marking results and the canopy structure parameters. Based on the multimodal observation data, a disease risk intensity field for the target farmland area is constructed after mapping by the observation operator, and a convection-diffusion-response model for the evolution of the disease risk intensity field over time is established. The convection-diffusion-response model includes a spatial heterogeneous resistance parameter field characterizing the spatial heterogeneous resistance of crops and a time-varying infection source intensity field characterizing the spatiotemporal changes of pathogen pressure. 2. Source-sink dual latent variable decoupling inversion based on farmland structure prior: The spatial heterogeneity resistance parameter field and the time-varying infection source intensity field are jointly inverted and assimilated, and the cross-regional expansion of the time-varying infection source intensity field is restricted by the propagation blocking boundary constraint in the joint inversion and assimilation, so as to obtain the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field. Based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field, the convection-diffusion-response model is used for forward prediction to obtain the disease risk intensity field prediction results; 3. Candidate region screening based on parameter posterior uncertainty contraction: Based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field, the uncertainty of the disease risk intensity field prediction results is quantified, and the total prediction uncertainty is decomposed into accidental uncertainty and cognitive uncertainty; 4. Observation quality participates in information gain gating for resampling path planning: A cognitive uncertainty trigger threshold is set, and regions with cognitive uncertainty higher than the cognitive uncertainty trigger threshold are identified as resampling candidate regions. A scoring function is constructed based on the disease risk intensity field prediction results, the cognitive uncertainty, the expected observation quality, and the flight sampling cost. The expected observation quality is used to characterize whether the candidate observations are sufficient to form effective update constraints on the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. The decrease in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the candidate trajectory is used as the information gain index. Constrained optimal experimental design path planning is performed for the resampling candidate regions to obtain the resampling trajectory. The complex sampling trajectory is converted into flight control commands and sent to the UAV flight controller to control the UAV to perform complex sampling. When the UAV reaches the target sampling position and meets the preset sampling trigger conditions, a trigger signal is output to control the sensor to perform complex sampling. At the same time, the sampling timestamp, attitude parameters and position parameters corresponding to the complex sampling are recorded simultaneously. 5. Sampling triggering and closed-loop update for consistent observation quality: The new observation data obtained from resampling is fed back to the joint inversion assimilation process to incrementally update the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. Based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field, the disease risk intensity field is re-predicted to enter the next round of forward prediction, uncertainty quantification, resampling candidate region determination, path planning and resampling execution closed-loop iteration.

[0013] In the above technical solution, preferably, the observation operator for constructing the disease risk intensity field is a phenological conditional time-delay observation operator. The phenological conditional time-delay observation operator is configured with weight parameters and response time-delay parameters for lesion proportion characteristics, normalized red edge index characteristics, and normalized canopy temperature difference characteristics for different phenological stages. The weight parameters and response time-delay parameters are jointly constrained by the phenological stage labeling results and at least one canopy structure parameter among canopy leaf area index, canopy closure degree, ground surface exposed ratio, and canopy shading degree, and the modal characteristics after time-delay compensation are mapped to a unified disease risk intensity value.

[0014] In the above technical solution, preferably, the joint inversion assimilation includes: A target functional is constructed that includes observation residuals, prior constraints of spatial heterogeneity resistance parameter fields, and sparse constraints of time-varying infection source intensity fields. The prior constraints of spatial heterogeneity resistance parameter fields are constrained by root zone soil moisture gradient, canopy leaf area index, and propagation blocking boundary. The sparse constraints of time-varying infection source intensity fields are used to limit the unconstrained expansion of infection sources in non-locally enhanced regions. Construct the adjoint equation corresponding to the convection-diffusion-response model and solve the adjoint variables in reverse time to calculate the gradient of the target functional with respect to the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field, respectively. The spatial heterogeneity resistance parameter field and the time-varying infection source intensity field are iteratively updated based on the gradient, and the cross-regional expansion of the time-varying infection source intensity field is restricted at the propagation blocking boundaries formed by plot boundaries, field ridges, ditches, or bare soil isolation zones.

[0015] In the above technical solution, preferably, the time-varying infection source intensity field is expressed by expanding using a time basis function as follows: in, For the preset time basis function, Let be the spatial coefficient field to be estimated.

[0016] In the above technical solution, preferably, the step of decomposing the total prediction uncertainty into accidental uncertainty and cognitive uncertainty includes: By jointly perturbing and sampling the posterior distribution of the spatially heterogeneous resistance parameter field and the observation noise, multiple sets of disease risk intensity field prediction results are obtained, and the total prediction uncertainty is determined accordingly. By fixing the updated spatial heterogeneity resistance parameter field, and only perturbing the observation noise and environmental random terms, random uncertainty is obtained; Cognitive uncertainty is determined by the difference between the total prediction uncertainty and the accidental uncertainty, and by treating regions with negative differences as zero.

[0017] In the above technical solution, preferably, the step of setting the cognitive uncertainty trigger threshold and constructing the scoring function includes: Regions where cognitive uncertainty exceeds the cognitive uncertainty trigger threshold are included in the resampling candidate set; A candidate region scoring function is constructed based on the risk prediction value corresponding to the disease risk intensity field prediction result, the cognitive uncertainty, the observation quality factor, and the comprehensive cost of flight and sampling. The observation quality factor is determined by flight altitude deviation, ground speed deviation, attitude angular velocity, viewpoint deviation, and positioning uncertainty, and is used to characterize whether the current observation is sufficient to form an effective update constraint on the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. The information gain of the corresponding candidate region will be included in the path planning objective function only when the observation quality factor is greater than the preset quality gating threshold. The candidate region scoring function is used as the target driving force for the constrained optimal experimental design path planning; The process of performing constrained optimal experimental design path planning on the resampling candidate region to obtain the resampling trajectory includes: The reduction in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field is used as the optimization objective, and a path planning objective function is constructed by combining the observation quality factor, flight energy consumption cost and sensor activation duration cost. Under constraints of remaining battery power, flight altitude, flight speed, turning maneuver, and sampling trigger, a trajectory search is performed on the resampling candidate region to obtain the resampling trajectory.

[0018] In the above technical solution, preferably, the GPIO trigger signal is output when the following conditions are met: the distance between the UAV and the target sampling position is not greater than the preset neighborhood radius, the ground speed is not greater than the preset ground speed threshold, the attitude angular velocity is not greater than the preset stability threshold, and the observation quality factor is not lower than the preset quality threshold. The observation quality evaluation model used in the sampling triggering stage is consistent with the observation quality evaluation model used for information gain gating in the path planning stage.

[0019] This invention also proposes an information detection active resampling system based on intelligent perception and source-destination decoupling assimilation, including an edge execution platform, an observation operator construction module, an edge computing module, a cloud-based joint inversion module, an observation quality assessment module, and a control and triggering module; The end-side execution platform is used to perform inspection flights in the target farmland area and collect multimodal observation data; The observation operator construction module is connected to the end-side execution platform and is used to perform time synchronization, spatial registration and phenological stage marking on the multimodal observation data. Based on the phenological stage, canopy structure parameters and modal time delay parameters, a phenological conditional time delay observation operator is constructed, and the multimodal observation features after time delay compensation are mapped into a disease risk intensity field. The edge computing module is connected to the observation operator construction module and is used to perform forward prediction on the convection-diffusion-response model based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field issued by the cloud joint inversion module, to obtain the disease risk intensity field prediction result, to quantify the uncertainty of the disease risk intensity field prediction result, to decompose the total prediction uncertainty into accidental uncertainty and cognitive uncertainty, and to determine the resampling candidate region based on the cognitive uncertainty; The cloud-based joint inversion module is connected to the edge computing module and is used to establish a convection-diffusion-response model of the disease risk intensity field evolving over time. Based on the adjoint state method, it performs joint inversion and assimilation of the spatial heterogeneous resistance parameter field and the time-varying infection source intensity field to obtain the updated spatial heterogeneous resistance parameter field and the updated time-varying infection source intensity field. The updated spatial heterogeneous resistance parameter field and the updated time-varying infection source intensity field are then sent to the edge computing module. The observation quality assessment module is connected to the edge computing module and the control and triggering module. It is used to calculate the observation quality factor based on the flight altitude deviation, ground speed deviation, attitude angular velocity, viewpoint deviation and positioning uncertainty, and to provide the observation quality factor to the complex sampling path planning and sampling triggering control. The observation quality factor is used to characterize whether the current observation is sufficient to form an effective update constraint on the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. The edge computing module is also used to construct a scoring function based on the disease risk intensity field prediction results, the cognitive uncertainty, the observation quality factor and the flight sampling cost, and to use the decrease in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the candidate trajectory as the information gain index, and to perform constrained optimal experimental design path planning for the resampling candidate region to obtain the resampling trajectory. The edge computing module is also used to include the information gain of the corresponding candidate region in the path planning objective function when the observation quality factor is greater than the preset quality gate threshold. The control and triggering module is connected to the edge computing module and the edge execution platform respectively. It is used to convert the complex sampling trajectory into MAVLink flight control commands and send them to the flight control of the edge execution platform to control the edge execution platform to perform complex sampling. When the UAV reaches the target sampling position and meets the preset sampling trigger conditions, it outputs a trigger signal through GPIO to control the sensors in the edge execution platform to perform complex sampling. At the same time, it synchronously records the sampling timestamp, attitude parameters and position parameters corresponding to the complex sampling. The cloud-based joint inversion module is also used to receive new observation data obtained from resampling, incrementally update the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field, and feed back the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field to the edge computing module, so as to form a closed-loop iteration of forward prediction, uncertainty quantification, resampling candidate region determination, path planning, resampling execution and incremental update.

[0020] In the above technical solution, preferably, the end-side execution platform includes a flight controller, a positioning module, a lidar or ToF module, and a multimodal sensor assembly; The multimodal sensor assembly includes a visible light camera, a multispectral camera, and a thermal infrared sensor; The positioning module is used to output position parameters corresponding to the multimodal observation data; The flight controller is used to output attitude parameters corresponding to the multimodal observation data; The lidar or ToF module is used to output flight altitude-related information and provide ranging input for terrain-following control; The flight controller is also used to control the end-side execution platform to fly along the complex sampling trajectory according to the MAVLink flight control commands issued by the control and triggering module.

[0021] In the above technical solution, preferably, the control and triggering module includes a MAVLink control submodule, a GPIO triggering submodule, and a synchronization recording submodule; The MAVLink control submodule is used to encapsulate the complex sampled trajectory into an MAVLink message containing waypoint control information and send it to the flight control of the end-side execution platform; The GPIO trigger submodule is used to output a trigger pulse when the end-side execution platform enters the neighborhood of the target waypoint, the ground speed is not greater than a preset ground speed threshold, the attitude angular velocity is not greater than a preset stability threshold, and the observation quality factor is not lower than a preset quality threshold, so as to trigger the sensors in the end-side execution platform to perform synchronous sampling. The synchronous recording submodule is used to synchronously record the sampling timestamp, attitude parameters, and position parameters when the GPIO is triggered, and to associate and store the sampling timestamp, attitude parameters, and position parameters with the new observation data obtained by resampling.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By constructing a phenological conditional time-delay observation operator, visible light, multispectral and thermal infrared observations with asynchronous responses under different phenological stages are first compensated by stage conditional time delay and then uniformly mapped into a disease risk intensity field. This makes the state variables received by the subsequent propagation model no longer affected by cross-modal time mismatch, thereby improving the stability and identifiability of source and destination dual latent variable decoupling inversion, rather than just improving the accuracy of single image recognition.

[0023] (2) By splitting the spatial differences in host resistance and the spatiotemporal enhancement of infection sources in disease transmission into two latent variables subject to different agricultural constraints, and conducting joint inversion under the constraints of farmland structure prior and local enhancement, it is possible to distinguish between high-risk areas caused by crop vulnerability and high-risk areas caused by external infection pressure, so that the resampling decision is oriented towards the identification of transmission mechanism, rather than just the level of risk value.

[0024] (3) Uncertainty was quantified in the prediction results of the disease risk intensity field based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field. The total prediction uncertainty was decomposed into accidental uncertainty and cognitive uncertainty, thus distinguishing between reducible and non-reducible uncertainty. By determining the resampling candidate region based solely on cognitive uncertainty, redundant sampling of high-noise and low-value regions was reduced, thereby improving the resampling targeting and sampling resource utilization efficiency.

[0025] (4) By setting the observation quality factor as a gate condition for whether the information gain is valid, and further participating in path scoring within the candidate region that meets the gate condition, this invention makes the resampling trajectory no longer oriented towards geometric coverage or simply high-risk areas, but towards high-value areas that can form effective assimilation constraints, thereby increasing the contribution of unit sampling to the reduction of the joint posterior variance of the host resistance parameter field and the infection source intensity field under limited endurance conditions.

[0026] (5) By continuing to use the same observation quality evaluation model as the path planning stage as the triggering condition during the sampling execution stage, invalid sampling caused by insufficient observation quality even though the trajectory is in place is avoided. This allows theoretical planning benefits to be transformed into actual usable observation benefits, thereby improving the stability of closed-loop updates rather than just improving flight control execution accuracy.

[0027] (6) By introducing canopy structure parameters into the phenological conditional time-delay observation operator, and introducing farmland structure priors, root zone moisture gradient, canopy connectivity and propagation blocking boundary into the source-sink decoupling assimilation process, this invention can express the inter-plot propagation blocking effect, canopy connectivity enhancement effect and local jump propagation phenomenon that are difficult to characterize by general uniform diffusion models, thereby improving the physical consistency and interpretability of disease propagation prediction in complex farmland scenarios. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating an active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation, as disclosed in one embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0030] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 As shown, according to the present invention, an active resampling method for information detection based on intelligent perception and source-destination decoupling and assimilation is provided. The present invention does not regard multimodal acquisition, propagation model update, resampling path planning and flight execution as independent serial modules. Instead, it forms a closed-loop technical solution around the main line of constructing a unified disease state quantity, decoupling propagation driving factors, implementing resampling under quality constraints only in areas where the posterior uncertainty of parameters can be reduced, and making the same observation quality factor act simultaneously on information gain evaluation and sampling triggering.

[0031] First, the UAV performs initial inspections according to the preset inspection area. It collects multimodal observation data of the target farmland area using onboard multimodal observation payloads (such as visible light cameras, multispectral cameras, and thermal infrared imagers), and simultaneously acquires the corresponding position parameters, attitude parameters, and sampling timestamps. It also acquires the canopy structure parameters corresponding to the multimodal observation data. Position parameters characterize the spatial location of the observation data, attitude parameters characterize the UAV's flight attitude during sampling, and the sampling timestamp characterizes the temporal attribution of the observation data. This provides a unified spatiotemporal reference for subsequent disease risk intensity field construction, model prediction, and observation mapping. Canopy structure parameters include at least one of the following: canopy leaf area index, canopy closure, percentage of bare soil, and canopy shading degree. Within the same phenological stage, the contribution of different modal observations to the disease risk intensity field can be adjusted based on differences in canopy observability, avoiding modal mapping deviations caused by canopy shading, bare soil background, or differences in canopy closure status. The onboard multimodal observation payloads include visible light cameras, multispectral cameras, and thermal infrared sensors.

[0032] After obtaining multimodal observation data, the data is first synchronized in time, registered in space, and labeled with phenological stages. Then, an observation operator is constructed based on the modal time delay parameters constrained by the phenological stage labeling results and canopy structure parameters.

[0033] In this embodiment, the edge side first performs time synchronization and spatial registration on the multimodal observation data to form observation input under a unified coordinate system. Furthermore, phenological stage labeling is applied to the multimodal observation data. Specifically, the crop phenological stage label g(t) corresponding to the current observation time can be determined based on accumulated temperature index, growing season days, leaf age information, fertility records, or historical agronomic data, providing stage condition information for the construction of subsequent observation operators.

[0034] An observation operator is constructed based on modal time delay parameters constrained by phenological stage marking results and canopy structure parameters. This observation operator is a phenological conditional time delay observation operator, which includes not only modal weight parameters under different phenological stages, but also response time delay parameters of different modes relative to disease state variables.

[0035] It should be noted that the phenological conditional time-delay observation operator does not directly fuse multiple modal features in parallel at the same time. Instead, it generates or selects corresponding response time-delay parameters and stage weight parameters for different modes based on the phenological stage labels. It first performs stage conditional time-delay compensation on each modal observation and then uniformly maps it into a disease risk intensity field, so that the disease risk intensity field becomes the only state quantity for subsequent propagation model assimilation and parameter update.

[0036] Specifically, assuming the current phenological stage of the crop is labeled g(t), the phenological conditional time-delay observation operator is configured with weight parameters and response time-delay parameters for the lesion proportion feature, normalized red edge index feature, and normalized canopy temperature difference feature for different phenological stages. The weight parameters and response time-delay parameters are jointly constrained by the phenological stage labeling results and at least one canopy structure parameter among the canopy leaf area index, canopy closure degree, ground surface exposure ratio, and canopy shading degree.

[0037] In one embodiment, the phenological conditional time-delay observation operator can be expressed as: in, Characteristics of the proportion of lesions, For the normalized red-edge index, To normalize the canopy temperature difference, , , These represent the response time delay parameters corresponding to the visible light, multispectral, and thermal infrared modes under phenological stage g, respectively. , , These represent the stage-conditional weights corresponding to each mode under phenological stage g. This is a bounded mapping function. By first compensating for time delays in each modal feature, and then performing a unified mapping according to the weight parameters under the corresponding phenological stage, the disease risk intensity value corresponding to each spatial location is obtained, thus forming a disease risk intensity field. The phenological conditionalized time-delay observation operator does not perform a general weighted fusion of multimodal features, but rather generates or selects corresponding response time delay parameters and stage weight parameters for different modes based on the phenological stage labels. After performing stage-conditionalized time delay compensation on each modal observation, it is then uniformly mapped into a disease risk intensity field, thereby making the disease risk intensity field a unified state variable for subsequent propagation model assimilation and parameter inversion.

[0038] In this embodiment, the disease risk intensity field C(x,t)∈[0,1] is used to characterize the degree of disease risk of the target plot at location x and time t. Its physical meaning is the continuous risk characterization corresponding to the ratio of infected biomass to total biomass per unit area. By introducing the above-mentioned phenological conditional time-delay observation operator, the systematic bias caused by phenological drift and asynchronous modal response can be reduced, so that the disease risk intensity field can simultaneously reflect the information of overt lesions, spectral physiological degradation, and canopy temperature difference anomalies, thereby providing a unified and stable state observation input for subsequent joint inversion and assimilation.

[0039] After obtaining multimodal observation data, a disease risk intensity field for the target farmland area is constructed based on the multimodal observation data after mapping with observation operators. This disease risk intensity field is then used as a state variable to establish a convection-diffusion-response model of its evolution over time. This model is not a simple record of current observation results, but rather abstracts the disease propagation process into a continuous evolutionary process with spatiotemporal propagation mechanisms. It introduces a spatially heterogeneous resistance parameter field and a time-varying infection source intensity field. The spatially heterogeneous resistance parameter field characterizes the differences in endogenous resistance to disease propagation in crops at different farmland locations, while the time-varying infection source intensity field characterizes the temporal and spatial changes in pathogen pressure. Therefore, subsequent predictions no longer rely solely on apparent image features, but incorporate both spatial differences in crop resistance and spatiotemporal changes in infection sources into the propagation model.

[0040] In this embodiment, the enhanced disease risk directly characterized by image observations may originate from either decreased host resistance or increased infection source activity, with these two factors overlapping at the apparent level. If only parameter fitting is performed on the overall transmission model, it is impossible to distinguish the contribution sources of crop vulnerability and increased external pathogen pressure. Therefore, this embodiment decomposes the transmission driving factors into two latent variables: a spatial heterogeneity parameter field of host resistance and a time-varying intensity field of infection sources. Prior constraints on farmland structure and local enhancement constraints are applied to these variables respectively to achieve source-sink decoupling in the agricultural transmission mechanism.

[0041] The spatially heterogeneous resistance parameter field and the time-varying infection source intensity field are jointly inverted and assimilated using the adjoint state method. In this joint inversion and assimilation, propagation blocking boundary constraints are incorporated to limit the cross-regional expansion of the time-varying infection source intensity field, resulting in updated spatially heterogeneous resistance parameter fields and updated time-varying infection source intensity fields. The inversion and assimilation does not directly measure crop resistance or infection source intensity; instead, it updates the implicit parameters through the dynamic residuals between the observed results and the propagation model. Specifically, to obtain the endogenous resistance parameter field κ(x) and the time-varying infection source intensity field q(x,t), the adjoint state method is used to construct the adjoint equation corresponding to the forward model. The adjoint variables are solved inversely over time, and the assimilation and inversion are performed using the dynamic residuals between the observed values ​​of the disease risk intensity field and the predicted values ​​of the forward model. This process does not rely on destructive physiological and physicochemical testing of the crop; instead, it uses the dynamic residuals of spatiotemporal sequence images to infer the implicit variables, resulting in the assimilated and updated parameters. and .

[0042] After the parameters are updated, the convection-diffusion-response model is forward-predicted based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field to obtain the disease risk intensity field prediction results. In this way, the model includes both existing observation information and parameter inversion results, thus forming a risk prediction output for future time.

[0043] Uncertainty quantification is performed on the disease risk intensity field prediction results based on the updated spatial heterogeneous resistance parameter field and the updated time-varying infection source intensity field, decomposing the total prediction uncertainty into accidental uncertainty and cognitive uncertainty. Specifically, in the uncertainty quantification stage, the posterior distributions of the spatial heterogeneous resistance parameter field and the time-varying infection source intensity field are jointly perturbed and sampled with observation noise to calculate the total prediction variance; [The remaining text appears to be incomplete and requires further context.] and Subsequently, only observation noise and environmental random terms are sampled to obtain the random uncertainty; the cognitive uncertainty field σ is obtained through difference. 2 epi (x,t). Random uncertainty is used to characterize the uncertainty caused by random factors such as environmental noise, light flicker, and environmental fluctuations. It cannot be reduced by increasing sampling. Cognitive uncertainty is used to characterize the prediction blind spots caused by insufficient model knowledge, insufficient parameter estimation, or insufficient identification of infection sources. It can be significantly reduced by adding observations in the corresponding areas. The effective decoupling of the two avoids a large amount of invalid and redundant sampling caused by the confusion between the two types of uncertainty in the existing technology.

[0044] Based on this decomposition result, a cognitive uncertainty trigger threshold is set, and regions with cognitive uncertainty exceeding the threshold are identified as candidate regions for resampling. This ensures that resampling decisions are driven only by regions where adding new observations can reduce model errors. A scoring function is constructed based on the disease risk intensity field prediction results, cognitive uncertainty, expected observation quality, and flight sampling cost. Expected observation quality characterizes whether candidate observations are sufficient to effectively update the spatially heterogeneous resistance parameter field and the time-varying infection source intensity field. Using this as the objective driving force, constrained optimal experimental design path planning is performed under endurance constraints, flight altitude constraints, maneuver constraints, and sampling trigger constraints. This generates the optimal resampling trajectory π that maximizes information gain under limited endurance conditions. The information gain metric for path planning is the decrease in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing candidate trajectories. Therefore, limited flight resources are preferentially allocated to regions that are both high-risk and high-information-gain. Cognitive uncertainty corresponds to the knowledge blind spots formed by insufficient estimation of the host resistance spatial heterogeneity parameter field and the time-varying infection source intensity field. Therefore, the determination of resampling candidate regions is essentially a candidate region selection process aimed at shrinking the joint posterior uncertainty of the two latent variables, rather than a re-shooting of simply high-risk regions.

[0045] The complex sampling trajectory π The commands are converted into MAVLink flight control commands and sent to the UAV flight controller to control the UAV to perform resampling. When the UAV reaches the target sampling position and meets the preset sampling trigger conditions, a trigger signal (TTL trigger pulse) is output via GPIO to control the sensor to perform resampling. At the same time, the sampling timestamp, attitude parameters, and position parameters corresponding to the resampling are recorded. This allows the path planning results to be directly converted into executable flight control behavior and strictly binds the sampling action with spatiotemporal parameters. Furthermore, the GPIO hard trigger mechanism controls the trigger latency to the millisecond level, effectively reducing sampling space misalignment errors and ensuring the projection accuracy of the observation operator and the quality of the assimilated input.

[0046] The newly observed data obtained from resampling is fed back to the parameter inversion and assimilation process to incrementally update the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. Based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field, the disease risk intensity field is re-predicted to enter the next round of forward prediction, uncertainty quantification, resampling candidate region determination, path planning and resampling execution closed-loop iteration, realizing the deep binding of algorithm decision logic and flight control underlying layer.

[0047] In this implementation, a closed-loop logic is used to transform disease monitoring from static identification to spatiotemporal evolution prediction and active resampling. This is achieved through risk intensity field construction, propagation model establishment, parameter inversion and assimilation, uncertainty decomposition, cognitive uncertainty-driven resampling, MAVLink / GPIO hard-triggered execution, and incremental updates. By introducing a time-varying infection source intensity field and expected observation quality, the resampling decision-making is simultaneously geared towards propagation-driven identification and high-quality observation acquisition. This allows for the triggering of high-risk area warnings through physical field prediction before lesions are visible to the naked eye, significantly advancing the warning time. Through precise scheduling driven by cognitive uncertainty, sampling resources are avoided from being wasted in high-noise, low-value areas, increasing the information gain per unit energy under the same endurance conditions. After multiple rounds of closed-loop assimilation, the estimation accuracy of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field continuously improves with each iteration, enhancing both assimilation accuracy and closed-loop stability. This improves the foresight of risk prediction, the targeting of resampling, and the stability of closed-loop updates in complex farmland environments.

[0048] In the above embodiments, preferably, lesion proportion features are extracted from visible light images, red edge index features are extracted from multispectral images, and the red edge index features are normalized to obtain normalized red edge index features; canopy temperature difference features are extracted from thermal infrared images, and the canopy temperature difference features are normalized to obtain normalized canopy temperature difference features. Further, the observation operator for constructing the disease risk intensity field is a phenologically conditional time-delay observation operator. The phenologically conditional time-delay observation operator configures weight parameters and response time-delay parameters for lesion proportion features, normalized red edge index features, and normalized canopy temperature difference features for different phenological stages.

[0049] Specifically, based on the current phenological stage of the crop, the stage weight and response time lag of the corresponding modal features are determined. After time lag compensation is performed on each modal feature, the lesion proportion feature, normalized red edge index feature, and normalized canopy temperature difference feature are uniformly mapped to obtain the disease risk intensity value at the corresponding location.

[0050] In this invention, the stage weight parameter and response delay parameter are not fixed empirical constants, but are generated or constrained based on phenological stage labels, canopy cover status, temperature difference sensitivity, and spectral degradation degree. Specifically, the response delay parameter of at least one mode takes different values ​​in different phenological stages, and the stage weight ranking of at least two modes differs in different phenological stages. Preferably, in the canopy unclosed stage, thermal infrared mode observation is suppressed by the proportion of exposed ground; in the canopy closed stage, the multispectral red-edge degradation signal is more sensitive to the precursor state of disease, and its response delay is less than that of the thermal infrared mode; in the lesion manifestation stage, the visible light mode is the dominant observation item.

[0051] Phenological stages are used not only to divide multimodal characteristics into stages, but also to determine the response time lag of each modal characteristic relative to the disease state variable. That is, under different phenological stages, the response times of the visible light mode, multispectral mode, and thermal infrared mode to disease progression are not entirely consistent. Therefore, it is necessary to set corresponding time lag parameters for each phenological stage and perform time lag compensation before mapping. The lesion proportion feature, normalized red edge index feature, and normalized canopy temperature difference feature, after time lag compensation, are then fused according to the weight parameters under the corresponding phenological stage to form a unified disease risk intensity value.

[0052] Through the above implementation methods, the differences in the response of different modes to the disease process under different phenological stages can be uniformly corrected, thereby improving the consistency and accuracy of the disease risk intensity field construction.

[0053] In the above implementation, the endogenous resistance parameter field κ(x) is a spatial distribution function characterizing the crop's ability to suppress diseases per unit time. Its value exhibits significant heterogeneity in space due to varietal differences, soil moisture gradients, local fertility, and microclimate disturbances. The time-varying infection source intensity field q(x,t) is used to characterize the changes in pathogen pressure in time and space. Its value dynamically evolves with changes in pathogen spread, local environmental conditions, and the activity level of the infection source. This parameter cannot be obtained through direct measurement but can only be obtained through the dynamic residual inversion of spatiotemporal sequence images. In the agricultural disease transmission scenario, the risk enhancement directly characterized by image observation may come from either a decrease in host resistance or an increase in infection source activity, and the two are overlapping at the apparent level. If only the overall parameters of the transmission model are estimated, it is impossible to distinguish the contribution sources of crop vulnerability and increased external pathogen pressure, which will lead to target bias in resampling. To this end, this invention decomposes the transmission driving factors into two latent variables: the host resistance spatial heterogeneity parameter field and the infection source time-varying intensity field, and applies different types of agricultural constraints to achieve source-sink decoupling in the agricultural transmission mechanism. Preferably, the joint inversion and assimilation of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field based on the adjoint state method includes: A target functional J is constructed, comprising observation residual terms, parameter prior constraints, and a sparse constraint term for the time-varying infection source intensity field. The observation residual term characterizes the deviation between the output of the observation operator and the prediction results of the propagation model. It is weighted by the observation error covariance matrix R, where each component is dynamically weighted by the sensor laboratory noise figure and the flight environment (e.g., light intensity fluctuations), ensuring that the contribution weights of observation data of different qualities to the residuals match their reliability. The update of the host resistance spatial heterogeneity parameter field is constrained by the farmland structure prior, which is jointly generated by the soil moisture distribution M(x) and the crop growth distribution G(x). At abrupt boundaries of soil moisture content or crop growth index, corresponding gradient jumps in the host resistance parameter field are allowed to ensure consistency between the inversion results and the farmland physical topology. The parameter prior constraint term is regularized using the structural prior covariance Q, which is constructed based on the soil moisture distribution M(x) and crop growth distribution G(x). Q introduces gradient jump constraints at the corresponding boundaries to guide the inversion results to converge within a physically reasonable spatial topology. This constraint limits the update direction and spatial smoothness of the parameter field, preventing non-physical smooth transitions in the inversion results and improving the consistency between the inversion results and the heterogeneous distribution of actual farmland, ensuring that the parameter field update is consistent with the physical topology of actual farmland. The time-varying infection source intensity field sparsity constraint term characterizes the agricultural scenario characteristic that infection sources typically enhance only in local areas, thereby suppressing large-scale unconstrained diffusion of the infection source intensity field in space. Its update is subject to local enhancement sparsity constraints to reflect the agricultural disease transmission characteristics where pathogen pressure typically enhances in local areas and time periods rather than spreading uniformly globally. The prior constraints of the spatial heterogeneity resistance parameter field are constrained by the root zone soil moisture gradient, canopy leaf area index, and propagation blocking boundary. The sparse constraint of the time-varying infection source intensity field is used to limit the unconstrained expansion of the infection source in non-locally enhanced regions.

[0054] Specifically, prior knowledge of farmland structure is introduced to guide the inversion results to conform to the actual physical topology: Where M(x) is soil moisture and G(x) is growth index, this structural prior makes the inverted resistance field κ have a reasonable gradient jump at the abrupt change in soil moisture content or growth.

[0055] An adjoint equation corresponding to the convection-diffusion-response model is constructed, and the adjoint variables are solved in reverse time to calculate the gradient of the objective functional with respect to the spatially heterogeneous resistance parameter field and the time-varying infection source intensity field. In a preferred embodiment, the gradient in the joint inversion process of source and host hidden variables can be solved using the adjoint state method, and iterative updates are performed using L-BFGS. The cross-regional expansion of the time-varying infection source intensity field is restricted at the propagation blocking boundaries formed by plot boundaries, field ridges, ditches, or bare soil isolation zones. The adjoint state method and L-BFGS are only preferred solution methods and do not constitute the only technical feature that distinguishes this invention from the prior art. The key to this invention does not lie in the specific gradient solution algorithm used, but in constraining the update of the host resistance parameter field through prior constraints of farmland structure, constraining the update of the infection source intensity field through local enhancement, and achieving source-host decoupling in the agricultural propagation mechanism under the drive of propagation residuals. The adjoint state method does not directly perform numerical difference differentiation on the target functional with respect to κ(x) and q(x,t) (the computational cost is a linear multiple of the parameter dimension). Instead, it constructs the conjugate equation of the forward equation (i.e., the adjoint equation), solves for the adjoint variable λ(x,t) by inverse time integration, and then efficiently calculates the gradient by convolving λ with the forward state. and The core feature of the adjoint state method is its inverse-time solution, which makes the computational cost of gradient calculation independent of the parameter space dimension, making it suitable for the inversion requirements of high-dimensional continuously distributed parameter fields in this implementation. The advantage of using the adjoint state method is that it can efficiently obtain gradient information in a high-dimensional distributed parameter space without explicitly perturbing the parameter field at each spatial location, thus adapting to the computational requirements of parameter assimilation in continuous farmland fields.

[0056] Specifically, we define the minimization functional: Using the adjoint state method, λ is solved in reverse time by constructing the adjoint equation, and the gradient is calculated as follows: The L-BFGS (Limited-Memory Quasi-Newton) iterative algorithm is used to iteratively update the spatial heterogeneous resistance parameter field and the time-varying infection source intensity field based on gradients, resulting in updated spatial heterogeneous resistance parameter fields and updated time-varying infection source intensity fields. The L-BFGS iterative algorithm approximates the inverse of the Hessian matrix by storing a limited amount of historical gradient information, achieving an iterative speed close to second-order convergence without significantly increasing memory consumption. It can achieve fast convergence while maintaining low storage overhead, making parameter inversion both computable and meeting the requirements of near-real-time assimilation updates. Thus, the parameter field update results driven by multiple rounds of observation residuals can gradually approximate the true distribution of spatial heterogeneous crop resistance and the spatiotemporal variation distribution of infection source intensity in the target farmland, completing the near-real-time assimilation update of the parameter field within a limited number of iterations.

[0057] Farmland structure priors are used not only to constrain the update direction and smoothness of the spatially heterogeneous resistance parameter field, but also to constrain the cross-regional propagation capabilities of pathogen transmission pathways at root zone water content abrupt change zones, canopy sparse zones, plot boundaries, field ridges, ditch zones, and bare soil isolation zones. Therefore, the propagation barrier effect, canopy connectivity enhancement effect, and local jump propagation effect in agricultural scenarios can be expressed in the propagation model.

[0058] The spatially heterogeneous resistance parameter field is not a freely updating abstract parameter field, but a constrained parameter field associated with the root zone soil moisture gradient, canopy leaf area index, canopy connectivity, and propagation barrier boundaries. At abrupt changes in root zone water content, canopy sparse zones, and plot boundaries, discontinuous gradient jumps in the spatially heterogeneous resistance parameter field are permitted to characterize the propagation barrier effect. The time-varying infection source intensity field is only allowed to update within locations and time periods that meet local enhancement conditions, and cross-regional expansion is restricted at propagation barrier boundaries. When local pathogen hotspots overlap with areas of high canopy leaf area index, a jump-like enhancement of the time-varying infection source intensity field is permitted, rather than updating in a uniform diffusion manner across the entire field.

[0059] In this implementation, by combining structural prior covariance, inverse-time solution of the adjoint equation, and L-BFGS iterative update, the inverted κ(x) spatially matches the actual farmland soil-crop physical topology, and the inverted q(x,t) can characterize the local enhancement region of the infection source in space and time. This achieves joint and efficient inversion assimilation of the spatially heterogeneous resistance parameter field and the time-varying infection source intensity field, significantly reducing inversion error. It enables the dynamic estimation of latent variables of crop resistance and infection sources that are difficult to measure directly, thereby improving the interpretability, adaptability, and predictive stability of the propagation model. The combination of the adjoint state method and L-BFGS achieves near real-time and efficient assimilation of the high-dimensional parameter field, providing continuously updated high-quality parameter field input for forward prediction and uncertainty quantification at the edge.

[0060] In the above embodiments, preferably, the total prediction uncertainty is decomposed into accidental uncertainty and cognitive uncertainty, including: Joint perturbation sampling is performed on the posterior distribution of the spatially heterogeneous resistance parameter field (i.e., the distribution near κ^(x) after assimilation), the posterior distribution of the time-varying infection source intensity field (i.e., the distribution near q^(x,t) after assimilation), and observation noise. Under the same boundary conditions, a forward model is driven to obtain multiple sets of disease risk intensity field prediction results. The spatial pointwise variance of N sets of prediction fields is calculated, and the total prediction uncertainty is determined accordingly. This total prediction uncertainty includes multiple factors such as model parameter uncertainty, measurement noise perturbation, and environmental fluctuations.

[0061] The estimation of random uncertainty employs a single-source perturbation sampling method with fixed parameter fields: the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field are fixed, and perturbation sampling is performed only on observation noise (such as sensor thermal noise, image blurring introduced by motor vibration) and environmental random terms (such as light flicker, airflow disturbances) to obtain Na sets of prediction fields. The pointwise variance of these fields is then calculated to obtain the random uncertainty. Since the parameter field remains unchanged at this point, this result mainly reflects the impact of external random perturbations on the prediction results. Physically, random uncertainty reflects the degree of prediction dispersion caused by unavoidable random environmental factors under the condition that the current parameter field estimation is known; it cannot be significantly reduced regardless of the number of new observations.

[0062] Cognitive uncertainty is determined based on the difference between total prediction uncertainty and random uncertainty. In one embodiment, regions where the difference between total prediction uncertainty and random uncertainty is negative are treated as zero. Thus, cognitive uncertainty physically reflects the model knowledge blind spot caused by insufficient estimation of the parameter field κ(x) and the time-varying infection source intensity field q(x,t). Incorporating new observations into the assimilation update in this region can separate the uncertainty caused by insufficient parameter estimation and model knowledge blind spot from the overall uncertainty, significantly reducing the posterior uncertainty of κ(x) and q(x,t), thereby reducing prediction error.

[0063] Specifically, in the implementation process, under the same boundary conditions, the i-th prediction field C(i)(x,t) is obtained through multiple random samplings, where i=1,2,…,N. The prediction mean field is defined as: Define the total prediction variance as: Total uncertainty Decomposed into: in, Due to chance and uncertainty, To understand uncertainty.

[0064] With the parameter fields fixed, the optimal estimates κ^(x) and q^(x,t) are assimilated. Only observation noise and environmental random terms are sampled to obtain the Na-group prediction fields. : The estimated cognitive uncertainty is: And on Zero-truncation is performed on the region.

[0065] In practical applications, higher cognitive uncertainty indicates a greater lack of effective knowledge constraints in the region, making further sampling more likely to yield higher information gain. Conversely, regions with high random uncertainty but low cognitive uncertainty tend to reflect more random noise and are not suitable for excessive resampling resources. Therefore, the above decomposition results directly provide a basis for subsequent candidate region selection and resampling scheduling. By using only cognitive uncertainty as a trigger, it effectively avoids ineffective redundant sampling caused by environmental noise, achieving precise allocation of sampling resources to spatial regions that can truly improve model accuracy.

[0066] In this implementation, by decomposing the total prediction uncertainty into random uncertainty and cognitive uncertainty, it is possible to distinguish between high-noise regions and high-knowledge-blind-zone regions, avoiding the waste of sampling resources in regions dominated by random noise. The cognitive uncertainty field continuously shrinks as the parameter field estimation accuracy improves, where the parameter field estimation accuracy includes the estimation accuracy of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field, thereby improving the effectiveness of the resampling decision.

[0067] In the above embodiments, preferably, setting a cognitive uncertainty trigger threshold and constructing a scoring function includes: Uncertainty in overall perception Spatial filtering is performed to filter out individuals whose cognitive uncertainty exceeds the cognitive uncertainty trigger threshold. s th The region is included in the resampling candidate set; this threshold s th This serves as an initial screening threshold for candidate regions, used to exclude areas with low knowledge gaps, thereby narrowing down the scope for subsequent planning. Threshold s th The threshold setting strikes a balance between resampling coverage and sampling efficiency: too low a threshold will include a large number of low-value areas in the candidate set, increasing the computational burden of path planning and diluting information gain; too high a threshold may miss important parameter blind spots.

[0068] After forming the candidate set, a candidate region scoring function is constructed based on the risk prediction value corresponding to the disease risk intensity field prediction result, cognitive uncertainty, observation quality factor, and the comprehensive cost of flight and sampling. The risk prediction value measures the potential disease risk in the region, reflecting the current urgency of disease in the region. Cognitive uncertainty measures the knowledge gain that further sampling in the region may bring, reflecting the degree of knowledge blind spots in parameter estimation. The observation quality factor is used to suppress sampling tasks that, although possessing high theoretical information gain, are difficult to provide effective observation constraints due to imaging blur, attitude instability, positioning errors, or viewpoint deviations. The comprehensive cost of flight and sampling is used to suppress resampling tasks that are too far away, too energy-intensive, or have low repetition value, reflecting the flight energy consumption and sensor activation cost required to reach the region and complete sampling. Thus, the four types of driving quantities are weighted by coefficients. oh r , oh u , oh q and oh c The weighted combination scoring function actually establishes a balance between risk benefits, knowledge benefits, observation quality benefits, and execution costs, so that the scoring function simultaneously considers the urgency of the disease, information value, observation availability, and sampling costs.

[0069] The observation quality factor not only reflects imaging sharpness and registration accuracy but also characterizes whether the current observation is sufficient to provide effective update constraints for the disease risk intensity field, the spatially heterogeneous resistance parameter field, and the time-varying infection source intensity field. Only when the observation quality factor exceeds a preset quality gating threshold is the information gain of the corresponding candidate region included in the path planning objective function. Specifically, when the flight perspective causes canopy occlusion to exceed the threshold, or the flight altitude makes individual plant-scale thermal anomalies indistinguishable, the observation quality factor still determines that the observation lacks effective assimilation constraint capability, even though the sensor can acquire images.

[0070] Using the candidate region scoring function as the objective driving force for the constrained optimal experimental design path planning, it provides an objective evaluation basis for subsequent path planning, under the range constraint (remaining battery SOC not lower than SOC). min ), flight altitude constraints ( H min ≤ H ≤ H max Under constraints of maneuvering (minimum turning radius, maximum centripetal acceleration, maximum velocity) and sampling triggering constraints (attitude stability threshold, ground speed threshold, and observation quality threshold), perform constrained optimal experimental design path planning to generate the optimal complex sampling trajectory π. The corresponding sampling waypoints make path planning no longer a geometric shortest path search, but also consider the information gain of multi-point combinations, trajectory continuity and multiple flight constraints. It is an optimization problem oriented towards information gain and risk-reward.

[0071] The observation quality judgment model used in the sampling triggering phase is consistent with the observation quality evaluation model used for information gain gating in the path planning phase. Resampling is only allowed when the UAV reaches the target sampling position and meets the same observation quality criteria.

[0072] In this embodiment, the observation quality factor is not a typical flight cost, but a closed-loop variable used to characterize whether candidate observations can form effective assimilation constraints. The information gain of a candidate region is only included in the path planning objective function when the observation quality factor of the candidate region is greater than a preset quality gating threshold. For candidate regions that meet the quality gating conditions, a candidate region scoring function is constructed based on the risk prediction value, cognitive uncertainty, observation quality factor, and the combined cost of flight and sampling.

[0073] During implementation, a candidate region scoring function is defined as follows: in: This represents the average of risk predictions. For cognitive uncertainty; Q(x,t) is the observation quality factor; Cost(x,t) is the combined cost of flight and sampling; oh r , oh u , oh q , oh c >0 represents the weighting coefficient.

[0074] Furthermore, the observation quality factor is defined as: in, Let x be the working height of the trajectory π at point x; Ground speed; Angular velocity; For positional uncertainty; From a perspective; , , The desired sampling conditions; , , , , These are the corresponding weighting coefficients; It is a bounded mapping function.

[0075] Set a threshold to trigger cognitive uncertainty. Only when the following conditions are met Only then does region X enter the resampling candidate set. The observation quality evaluation model used in the sampling triggering phase is consistent with the observation quality evaluation model used for information gain gating in the path planning phase.

[0076] Furthermore, the constrained optimal experimental design path planning takes the reduction in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the candidate trajectory as the optimization objective, and constructs the path planning objective function by combining the observation quality factor, flight energy consumption cost and sensor activation duration cost.

[0077] In this implementation, by screening for cognitive uncertainty thresholds and constructing a scoring function, the dual weights of disease risk and knowledge blind spots are incorporated into path decision-making. Furthermore, observation quality constraints are incorporated into the scoring function and path planning objectives, elevating the resampling task from a rough judgment of whether to sample to a quantitative decision on where sampling is more valuable and whether effective data can be collected at that location. This improves the targeting of resampling candidate area determination and the foundation for subsequent path planning optimization.

[0078] In the above embodiments, preferably, constrained optimal experimental design path planning is performed on the resampling candidate region to obtain the resampling trajectory, including: The optimization objective is the reduction in the joint variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the candidate trajectory. A path planning objective function is constructed by combining the observation quality factor, flight energy cost, and sensor activation duration cost. The posterior variance reduction represents the decrease in parameter field uncertainty after executing the trajectory and obtaining corresponding observation data, reflecting the degree of reduction in model uncertainty after sampling. Specifically, the joint posterior variance reduction characterizes the degree to which the uncertainties of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field are reduced together after executing the trajectory. By targeting the posterior variance reduction, the path planning is ensured to directly address the core objective of improving the accuracy of parameter field estimation, rather than simply maximizing the number of observation points or coverage area. Based on this, the observation quality factor Q is introduced. meas Flight energy consumption cost E flight and sensor effective turn-on time T sensor As a constraint or penalty, the observation quality factor characterizes the probability of acquiring valid observation data at the sampling location corresponding to the candidate trajectory. The flight energy cost reflects the endurance required to execute the trajectory, and the sensor activation duration cost reflects the sampling execution cost, all expressed through the penalty coefficient λ. qλ1 and λ2 establish an explicit trade-off between information gain and resource consumption, constructing a path planning objective function. Combining these four factors ensures that path planning not only pursues sampling benefits but also considers UAV endurance, equipment burden, and actual observation availability.

[0079] Under constraints of remaining battery power, flight altitude, flight speed, turning maneuvering, and sampling triggering, a trajectory search is performed on the candidate region for resampling to obtain the resampling trajectory. The remaining battery power constraint requires that the remaining battery power (SOC) be no less than the total SOC throughout the entire flight. min (Recommended 20%), ensuring the drone has the ability to safely return after performing resampling tasks; flight altitude requirements: maintain the operating altitude at... H min to H max Within a certain range (recommended 0.4m to 1.5m), collisions with the canopy are avoided while ensuring the effective detection range of the sensor. Flight speed constraints are set with a maximum cruise speed limit (not exceeding 3.5m / s) to ensure imaging quality and flight stability during low-altitude operations and reduce the impact of motion blur on image quality. Turning maneuver constraints limit trajectory curvature changes to ensure attitude stability and prevent trajectories that are optimal but difficult to execute in practice. Sampling trigger constraints stipulate that sampling is only allowed when both attitude stability and ground speed thresholds are met, and require the observation quality factor to be no lower than a preset quality threshold to ensure that the planned trajectory matches the subsequent triggering sampling conditions. The resulting complex-sampled trajectory has both information gain significance and practical flight feasibility. The observation quality judgment model used in the sampling triggering phase is consistent with the observation quality evaluation model used in the path planning phase, ensuring that candidate sampling points considered in the path planning phase are considered to have high information value, and that sampling is only triggered in the actual execution phase when the same quality criteria are met, avoiding a disconnect between theoretical information gains and actual usable observations.

[0080] The trajectory search performs a local trajectory search within the feasible region defined by the above constraints, and outputs the optimal trajectory π. And the corresponding sampling waypoint sequence, for subsequent MAVLink instruction encapsulation and GPIO trigger scheduling.

[0081] During implementation, a constrained OED path planning model is established based on the candidate region set: in, , π represents the flight trajectory to be optimized; y π The expected observation data under trajectory π; Varκ(κ∣y π The posterior variance of the joint parameters of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the trajectory is the following.E flight For flight energy consumption; T sensor λ1 and λ2 are the effective start-up time of the sensor.

[0082] The remaining power limit is: in, SOC min The preferred value is 20%.

[0083] The safe flight altitude constraints are: Preferably, H min =0.4m, H max =1.5m.

[0084] In this implementation, by taking the reduction in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field as the core optimization objective, the path planning results are directly linked to the assimilation accuracy improvement objective. Furthermore, by introducing observation quality factors, energy consumption, sensor activation duration, and flight constraints, a resampling trajectory that balances information gain, observation quality, and flight executability can be generated under limited endurance and execution conditions, thereby improving the overall efficiency of active resampling.

[0085] In the above embodiment, preferably, the resampling trajectory is converted into MAVLink flight control commands and sent to the UAV flight controller to control the UAV to perform resampling. When the UAV reaches the target sampling position and meets the preset sampling trigger conditions, a trigger signal is output through GPIO to control the sensor to perform resampling. At the same time, the sampling timestamp, attitude parameters, and position parameters corresponding to the resampling are recorded synchronously, including: Each sampled waypoint in the resampled trajectory is encapsulated into a MAVLink message containing waypoint control information according to the MAVLink v2.0 protocol specification and sent to the UAV flight controller. The flight controller performs trajectory tracking at a control frequency of 400 Hz and combines it with LiDAR to achieve low-altitude terrain following. This realizes the direct mapping of the upper-level planning results to the flight control command layer, so that the path planning output can be directly recognized and executed by the flight control system.

[0086] During flight, when the UAV enters the neighborhood of the target waypoint and its horizontal ground speed is below a set threshold (ensuring image blur is within an acceptable range), its attitude angular velocity is below a stability threshold (ensuring image registration accuracy), and its observation quality factor is not below a preset quality threshold, a 3.3V TTL trigger pulse is output via GPIO to trigger the sensor to perform synchronous sampling. The rising edge triggers the camera shutter, the pulse width is not less than 5 ms, and the trigger delay is not more than 1 ms. The reason for setting the target waypoint neighborhood, attitude and ground speed stability conditions, and the observation quality factor threshold conditions is to ensure that the UAV is in a relatively stable motion state during sampling, and that the current flight altitude, viewing angle deviation, and positioning accuracy meet the requirements for effective observation, thereby reducing image blur, viewing angle shift, or multimodal data misalignment. Using GPIO hard triggering allows the sampling action to be strictly bound to the flight state, effectively ensuring high-precision registration between the observation point coordinates and the actual sampling position, compressing latency jitter from tens of milliseconds to milliseconds, and reducing the impact of latency jitter on spatial registration. The observation quality factor threshold is consistent with the observation quality evaluation model used in the information gain gating of the path planning stage, thus making the path planning stage a sampling point with high information value. In the actual execution stage, the same quality criteria must still be met before sampling can be triggered.

[0087] Upon GPIO triggering, the multispectral camera receives a hard trigger signal and completes synchronized shooting across all channels (with a synchronization error between channels not exceeding 1ms). Simultaneously, the thermal infrared imager triggers sampling, recording the sampling timestamp, attitude parameters (roll, pitch, and yaw angles), and position parameters (RTK-GNSS positioning coordinates) at the same moment the GPIO triggers. The sampling timestamp, attitude parameters, and position parameters are then associated with the newly obtained observation data obtained through resampling and stored, forming an aligned data packet integrating image, pose, and timestamp. This synchronized recording process allows subsequent assimilation processes to accurately determine when, where, and in what attitude the new observation data was acquired, thereby reducing projection errors in the observation operator and improving the stability of parameter inversion.

[0088] In this implementation, a strong coupling mechanism is established between planning results, flight execution, and observation sampling by issuing MAVLink flight control commands, GPIO hard-triggered sampling, and synchronous recording of spatiotemporal parameters. By incorporating the observation quality factor threshold into the triggering conditions, invalid sampling is avoided when the target position is reached but the observation quality is insufficient. This improves the registration accuracy of resampled data and spatiotemporal information and enhances the reliability of subsequent parameter assimilation updates.

[0089] According to the active resampling method for information detection based on intelligent perception and source-sink decoupling assimilation disclosed in the above embodiments, the steps in the implementation process include: (1) Based on RGB, multispectral and infrared data, after time synchronization, spatial registration, phenological stage marking and phenological conditional time delay compensation, the disease risk intensity field C(x,t) is constructed; (2) A forward model of disease transmission is established based on the convection-diffusion-response equation: the forward model includes both the spatially heterogeneous resistance parameter field κ(x) and the time-varying infection source intensity field q(x,t): (3) By jointly inverting the spatial heterogeneity resistance parameter field κ(x) and the time-varying infection source intensity field q(x,t) using the adjoint state method, the assimilated updated parameter field κ^(x) and infection source intensity field q^(x,t) are obtained; (4) Based on the Monte Carlo sampling decomposition of total uncertainty, the cognitive uncertainty field is obtained. The total uncertainty is determined by the combined perturbation of the spatial heterogeneity resistance parameter field, the time-varying infection source intensity field, and the observation noise. (5) with With risk forecast value Observation quality factor Q meas (x,π) jointly constructs a complex sampling scoring function and performs OED path planning, where the path planning takes the reduction of the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field as the optimization objective; (6) The optimal trajectory π The flight controller is injected via MAVLink, and the sensor is sampled via GPIO hard trigger at the target waypoint. The GPIO trigger conditions include ground speed threshold, attitude stability threshold, and observation quality factor threshold. (7) After the new observations are back-transmitted, incremental assimilation is performed to update κ(x) and q(x,t) and enter the next closed loop.

[0090] This invention also proposes an active resampling system for information detection based on intelligent perception and source-destination decoupling and assimilation. The system adopts a cloud-edge-device collaborative architecture, including a device-side execution platform, an observation operator construction module, an edge computing module, a cloud-cloud joint inversion module, an observation quality assessment module, and a control and triggering module.

[0091] The observation operator construction module is connected to the end-side execution platform to perform time synchronization, spatial registration, and phenological stage marking on multimodal observation data. Based on phenological stages, canopy structure parameters, and modal time delay parameters, it constructs phenological conditional time delay observation operators and maps the time delay compensated multimodal observation features into a disease risk intensity field.

[0092] The edge computing module connects to the end-side execution platform and undertakes the most time-sensitive computational tasks: Based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field issued by the cloud-based joint inversion module, it performs forward prediction on the convection-diffusion-response model to obtain the disease risk intensity field prediction result. The uncertainty of the disease risk intensity field prediction result is quantified, and the total prediction uncertainty is decomposed into random uncertainty and cognitive uncertainty. Based on the cognitive uncertainty, resampling candidate regions are determined. Furthermore, a scoring function is constructed based on the disease risk intensity field prediction result, cognitive uncertainty, observation quality factor, and flight sampling cost. Using the decrease in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the candidate trajectory as the information gain index, constrained optimal experimental design path planning is performed for the resampling candidate regions to obtain the resampling trajectory. The above calculations are completed locally at the edge, avoiding the communication latency caused by all computations being performed in the cloud and ensuring near real-time response capability for resampling scheduling.

[0093] The cloud-based joint inversion module is connected to the edge computing module and deployed on ground stations or high-performance servers to undertake computationally intensive, high-precision tasks: establishing a convection-diffusion-response model of the disease risk intensity field evolving over time; performing joint inversion and assimilation of the spatial heterogeneous resistance parameter field and the time-varying infection source intensity field based on the adjoint state method to obtain updated spatial heterogeneous resistance parameter field and updated time-varying infection source intensity field; and sending the updated spatial heterogeneous resistance parameter field and updated time-varying infection source intensity field to the edge computing module, forming a collaborative relationship between high-precision modeling and assimilation in the cloud and real-time prediction and decision-making at the edge, achieving an optimal balance between computational accuracy and response time.

[0094] The observation quality assessment module is connected to the edge computing module and the control and triggering module respectively. It is used to calculate the observation quality factor based on the flight altitude deviation, ground speed deviation, attitude angular velocity, viewpoint deviation and positioning uncertainty. The observation quality factor is used to characterize whether the current observation is sufficient to form an effective update constraint on the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field, and provides the observation quality factor to the complex sampling path planning and sampling trigger control.

[0095] The edge computing module is also used to include the information gain of the corresponding candidate region in the path planning objective function when the observation quality factor is greater than the preset quality gate threshold.

[0096] The control and triggering module is connected to the edge computing module and the edge execution platform, respectively. It is used to convert the complex sampling trajectory into MAVLink flight control commands and send them to the flight control of the edge execution platform to control the edge execution platform to perform complex sampling. When the UAV reaches the target sampling position and meets the preset sampling trigger conditions, it outputs a trigger signal through GPIO to control the sensors in the edge execution platform to perform complex sampling. At the same time, it synchronously records the sampling timestamp, attitude parameters and position parameters corresponding to the complex sampling.

[0097] Among them, the cloud-based joint inversion module is also used to receive new observation data obtained from resampling, incrementally update the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field, and feed back the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field to the edge computing module to form a closed-loop iteration of forward prediction, uncertainty quantification, resampling candidate region determination, path planning, resampling execution and incremental update.

[0098] During implementation, the physical parameters of the unmanned aerial vehicle include: Airframe configuration: 220mm wheelbase carbon fiber integrated frame.

[0099] Power characteristics: single-aircraft takeoff mass 750g, thrust-to-weight ratio ≥2.2.

[0100] Flight constraints: Maximum cruising speed 3.5 m / s, maximum flight time 18 min.

[0101] Navigation accuracy: Integrated RTK-GNSS, horizontal positioning error ≤0.05m, vertical accuracy ≤0.1m.

[0102] Terrain following: LiDAR / ToF module to meet the requirements of stable altitude control for low-altitude operations.

[0103] Computing power base: Edge module computing power ≥20TOPS, equipped with 8GB LPDDR5 memory, supports TensorRT acceleration, and model inference latency ≤30ms.

[0104] Sensing specifications: Infrared thermal imager NETD < 50mK; Multispectral camera shutter synchronization error ≤ 1ms for each channel.

[0105] Flight control center: based on STM32H7, main frequency 480MHz, control frequency 400Hz.

[0106] Underlying protocol: Supports the MAVLink standard instruction set (MISSION_ITEM, SET_POSITION_TARGET_LOCAL_NED, etc.).

[0107] Physical triggering link: The edge GPIO outputs a 3.3V TTL push-pull signal, which triggers the camera shutter on the rising edge. The pulse width is ≥5ms and the trigger delay is ≤1ms, ensuring image spatial alignment under high dynamic range inspection.

[0108] In this implementation, the computationally intensive propagation model building and parameter inversion assimilation tasks are deployed in the cloud through a cloud-edge-device collaborative architecture, achieving an organic integration of high-precision assimilation computation and real-time sampling scheduling. The prediction, decision-making and execution tasks with high real-time requirements are deployed at the edge and on-device execution platforms, forming a sustainable autonomous closed loop. This balances high-precision modeling capabilities with low-latency execution capabilities, enhancing the overall real-time performance, executability and closed-loop update capabilities of the system, and enabling multiple rounds of iterative optimization without human intervention.

[0109] In the above embodiments, preferably, the end-side execution platform includes a flight controller, a positioning module, a lidar or ToF module, and a multimodal sensor assembly; The multimodal sensor assembly includes a visible light camera, a multispectral camera, and a thermal infrared sensor, used to acquire visible features of lesions, red edge index features, and canopy temperature difference features, respectively.

[0110] The positioning module is used to output the position parameters corresponding to the multimodal observation data; The flight controller is used to output attitude parameters corresponding to multimodal observation data; LiDAR or ToF modules are used to output flight altitude-related information; Among these features, the flight altitude-related information output by the lidar or ToF module can be used for low-altitude operational altitude stability control, ensuring a relatively stable observation distance between the multimodal observation payload and the crop canopy. This helps reduce imaging scale variations and improve the spatial consistency of multimodal data. Furthermore, the lidar or ToF module also provides ranging input for terrain-following control. In addition to flight control, the flight controller is used to control the platform to fly along the resampling trajectory according to the MAVLink flight control commands issued by the control and triggering module, enabling the resampling task to reach the designated area and waypoints according to the planned trajectory.

[0111] In this embodiment, by integrating flight control, positioning, altitude sensing, and multimodal sensor components into the end-side execution platform, the system achieves integrated observation and acquisition, flight control, position and attitude output, and altitude stability control, providing the basic hardware support for performing multimodal inspection and active resampling tasks.

[0112] In the above embodiments, preferably, the control and triggering module includes a MAVLink control submodule, a GPIO triggering submodule, and a synchronization recording submodule; The MAVLink control submodule encapsulates the complex sampled trajectory into MAVLink messages containing waypoint control information and sends them to the flight controller on the end-side execution platform, enabling the upper-layer planning results to be converted into executable flight controller commands. Specifically, the trajectory point sequence is encapsulated according to the MAVLink v2.0 protocol specification into waypoint control messages containing parameters such as target position coordinates, target velocity, and waypoint reach tolerance radius. The message type is selected based on the flight phase: SET_POSITION_TARGET_LOCAL_NED (continuous velocity / position control) or MISSION_ITEM (mission waypoint). The encapsulated message is sent to the flight controller module on the end-side execution platform via the control link. The flight controller tracks the trajectory in real time at a control frequency of 400Hz, achieving strong coupling between algorithm decision-making and the underlying flight controller execution, eliminating communication delays and command conversion losses between the algorithm layer and the flight control layer in existing technologies.

[0113] The GPIO trigger submodule determines that the sampling trigger conditions are met when the end-side execution platform enters the neighborhood of the target waypoint (position error enters the convergence radius), the horizontal ground speed is below a set threshold, the attitude angular velocity is below a stability threshold, and the observation quality factor is not below a preset quality threshold. It then immediately outputs a 3.3V TTL push-pull trigger pulse via GPIO, with the rising edge triggering synchronous imaging by the multi-modal sensor. The dual determination of ground speed and attitude in the sampling trigger constraints ensures that the UAV is in a relatively stable motion state at the sampling moment, reducing the impact of motion blur on image quality and the registration error between the image and position coordinates. Furthermore, by introducing an observation quality factor threshold determination, factors such as flight altitude deviation, viewing angle deviation, and positioning uncertainty can be comprehensively constrained, avoiding invalid sampling when position and attitude conditions are met but observation availability is insufficient.

[0114] The synchronous recording submodule is used to synchronously record the sampling timestamp, attitude parameters, and position parameters at the same time as the GPIO output trigger pulse. It also associates and stores the sampling timestamp, attitude parameters, and position parameters with the new observation data obtained by resampling, so that the subsequent observation operator can accurately project the sensor pixels to the physical space coordinates, calculate the observation residuals, and input them into the cloud parameters for inversion and assimilation.

[0115] In this modular structure, the control and triggering modules are no longer just simple data forwarding units, but simultaneously serve as the planning execution interface, sampling triggering interface, and spatiotemporal information binding interface. The MAVLink control submodule ensures that the planned trajectory is accurately transmitted, the GPIO triggering submodule ensures that the sampling action is triggered under appropriate flight conditions and that the observation quality during sampling triggering meets preset requirements, and the synchronization recording submodule ensures a one-to-one correspondence between the sampled data and the spatiotemporal parameters. Thus, planning results, flight control execution, sensor sampling, and data recording are organized into a unified execution unit, reducing timing mismatches between different functional modules, providing high-quality spatiotemporally aligned observation input for cloud parameter inversion and assimilation, and ensuring the assimilation accuracy and stability of the system in multiple rounds of closed-loop iteration.

[0116] In this implementation, by subdividing the control and triggering module into a MAVLink control submodule, a GPIO triggering submodule, and a synchronization recording submodule, the division of labor among instruction issuance, sampling triggering, and spatiotemporal recording during the resampling execution process is clearly defined, which helps to improve the clarity of the system execution chain, timing consistency, and data registration stability. Furthermore, by introducing an observation quality factor threshold condition in the GPIO triggering submodule, the effectiveness of resampling execution and the quality of observation input are further improved.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation, characterized in that, include: Acquire multimodal observation data collected by UAV during the inspection of the target farmland area, as well as the position parameters, attitude parameters and sampling timestamps corresponding to the multimodal observation data, and acquire the canopy structure parameters corresponding to the multimodal observation data; The multimodal observation data are time-synchronized, spatially registered, and phenological stage marked. An observation operator is constructed based on the modal time delay parameters constrained by the phenological stage marking results and the canopy structure parameters. Based on the multimodal observation data, a disease risk intensity field for the target farmland area is constructed after mapping by the observation operator, and a convection-diffusion-response model for the evolution of the disease risk intensity field over time is established. The convection-diffusion-response model includes a spatial heterogeneous resistance parameter field characterizing the spatial heterogeneous resistance of crops and a time-varying infection source intensity field characterizing the spatiotemporal changes of pathogen pressure. The spatial heterogeneity resistance parameter field and the time-varying infection source intensity field are jointly inverted and assimilated. In the joint inversion and assimilation, the propagation blocking boundary constraint is combined to limit the cross-regional expansion of the time-varying infection source intensity field, so as to obtain the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field. Based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field, the convection-diffusion-response model is used for forward prediction to obtain the disease risk intensity field prediction results; The uncertainty of the disease risk intensity field prediction results is quantified based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field, and the total prediction uncertainty is decomposed into accidental uncertainty and cognitive uncertainty. A cognitive uncertainty trigger threshold is set, and regions with cognitive uncertainty higher than the cognitive uncertainty trigger threshold are identified as resampling candidate regions. A scoring function is constructed based on the disease risk intensity field prediction results, the cognitive uncertainty, the expected observation quality, and the flight sampling cost. The expected observation quality is used to characterize whether the candidate observations are sufficient to form effective update constraints on the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. The decrease in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the candidate trajectory is used as the information gain index. Constrained optimal experimental design path planning is performed on the resampling candidate regions to obtain the resampling trajectory. The complex sampling trajectory is converted into flight control commands and sent to the UAV flight controller to control the UAV to perform complex sampling. When the UAV reaches the target sampling position and meets the preset sampling trigger conditions, a trigger signal is output to control the sensor to perform complex sampling. At the same time, the sampling timestamp, attitude parameters and position parameters corresponding to the complex sampling are recorded simultaneously. The new observation data obtained from resampling is fed back to the joint inversion assimilation process to incrementally update the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. Based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field, the disease risk intensity field is re-predicted to enter the next round of forward prediction, uncertainty quantification, resampling candidate region determination, path planning and resampling execution closed-loop iteration.

2. The active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation according to claim 1, characterized in that, The observation operator for constructing the disease risk intensity field is a phenological conditional time-delay observation operator. The phenological conditional time-delay observation operator configures weight parameters and response time-delay parameters for lesion proportion characteristics, normalized red edge index characteristics, and normalized canopy temperature difference characteristics for different phenological stages. The weight parameters and response time-delay parameters are jointly constrained by the phenological stage labeling results and at least one canopy structure parameter among canopy leaf area index, canopy closure degree, ground surface exposed ratio, and canopy shading degree, and map each modal feature after time-delay compensation to a unified disease risk intensity value.

3. The active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation according to claim 1, characterized in that, The joint inversion assimilation includes: A target functional is constructed that includes observation residuals, prior constraints of spatial heterogeneity resistance parameter fields, and sparse constraints of time-varying infection source intensity fields. The prior constraints of spatial heterogeneity resistance parameter fields are constrained by root zone soil moisture gradient, canopy leaf area index, and propagation blocking boundary. The sparse constraints of time-varying infection source intensity fields are used to limit the unconstrained expansion of infection sources in non-locally enhanced regions. Based on the adjoint state method, an adjoint equation corresponding to the convection-diffusion-response model is constructed, and the adjoint variables are solved in reverse time to calculate the gradient of the target functional with respect to the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field, respectively. The spatial heterogeneity resistance parameter field and the time-varying infection source intensity field are iteratively updated based on the gradient, and the cross-regional expansion of the time-varying infection source intensity field is restricted at the propagation blocking boundaries formed by plot boundaries, field ridges, ditches, or bare soil isolation zones.

4. The active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation according to claim 3, characterized in that, The time-varying infection source intensity field is expressed by expanding the time basis function as follows: in, For the preset time basis function, Let be the spatial coefficient field to be estimated.

5. The active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation according to claim 1, characterized in that, The decomposition of total forecast uncertainty into accidental uncertainty and cognitive uncertainty includes: By jointly perturbing and sampling the posterior distribution of the spatially heterogeneous resistance parameter field and the observation noise, multiple sets of disease risk intensity field prediction results are obtained, and the total prediction uncertainty is determined accordingly. By fixing the updated spatial heterogeneity resistance parameter field, and only perturbing the observation noise and environmental random terms, random uncertainty is obtained; Cognitive uncertainty is determined based on the difference between the total prediction uncertainty and the random uncertainty, and regions with negative differences are treated as zero. The cognitive uncertainty represents the knowledge blind spots caused by insufficient estimation of the host resistance spatial heterogeneity parameter field and the time-varying intensity field of the infection source.

6. The active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation according to claim 1, characterized in that, The process of setting a threshold for cognitive uncertainty and constructing a scoring function includes: Regions where cognitive uncertainty exceeds the cognitive uncertainty trigger threshold are included in the resampling candidate set; A candidate region scoring function is constructed based on the risk prediction value corresponding to the disease risk intensity field prediction result, the cognitive uncertainty, the observation quality factor, and the comprehensive cost of flight and sampling. The observation quality factor is determined by flight altitude deviation, ground speed deviation, attitude angular velocity, viewpoint deviation, and positioning uncertainty, and is used to characterize whether the current observation is sufficient to form an effective update constraint on the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. The information gain of the corresponding candidate region will be included in the path planning objective function only when the observation quality factor is greater than the preset quality gating threshold. The candidate region scoring function is used as the target driving force for the constrained optimal experimental design path planning; The process of performing constrained optimal experimental design path planning on the resampling candidate region to obtain the resampling trajectory includes: The reduction in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field is used as the optimization objective, and a path planning objective function is constructed by combining the observation quality factor, flight energy consumption cost and sensor activation duration cost. Under constraints of remaining battery power, flight altitude, flight speed, turning maneuver, and sampling trigger, a trajectory search is performed on the resampling candidate region to obtain the resampling trajectory.

7. The active resampling method for information detection based on intelligent sensing and source-destination decoupling assimilation according to claim 1, characterized in that, A trigger signal is output via GPIO. The trigger signal is output when the following conditions are met: the distance between the UAV and the target sampling position is not greater than a preset neighborhood radius, the ground speed is not greater than a preset ground speed threshold, the attitude angular velocity is not greater than a preset stability threshold, and the observation quality factor is not lower than a preset quality threshold. The observation quality evaluation model used in the sampling triggering stage is consistent with the observation quality evaluation model used for information gain gating in the path planning stage.

8. An active resampling system for information detection based on intelligent sensing and source-destination decoupling assimilation, characterized in that, It includes an edge execution platform, an observation operator construction module, an edge computing module, a cloud-based joint inversion module, an observation quality assessment module, and a control and triggering module; The end-side execution platform is used to perform inspection flights in the target farmland area and collect multimodal observation data; The observation operator construction module is connected to the end-side execution platform and is used to perform time synchronization, spatial registration and phenological stage marking on the multimodal observation data. Based on the phenological stage, canopy structure parameters and modal time delay parameters, a phenological conditional time delay observation operator is constructed, and the multimodal observation features after time delay compensation are mapped into a disease risk intensity field. The edge computing module is connected to the observation operator construction module and is used to perform forward prediction on the convection-diffusion-response model based on the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field issued by the cloud joint inversion module, to obtain the disease risk intensity field prediction result, to quantify the uncertainty of the disease risk intensity field prediction result, to decompose the total prediction uncertainty into accidental uncertainty and cognitive uncertainty, and to determine the resampling candidate region based on the cognitive uncertainty; The cloud-based joint inversion module is connected to the edge computing module and is used to establish a convection-diffusion-response model of the disease risk intensity field evolving over time. It performs joint inversion and assimilation on the spatial heterogeneous resistance parameter field and the time-varying infection source intensity field to obtain the updated spatial heterogeneous resistance parameter field and the updated time-varying infection source intensity field. The updated spatial heterogeneous resistance parameter field and the updated time-varying infection source intensity field are then sent to the edge computing module. The observation quality assessment module is connected to the edge computing module and the control and triggering module. It is used to calculate the observation quality factor based on the flight altitude deviation, ground speed deviation, attitude angular velocity, viewpoint deviation and positioning uncertainty, and to provide the observation quality factor to the complex sampling path planning and sampling triggering control. The observation quality factor is used to characterize whether the current observation is sufficient to form an effective update constraint on the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field. The edge computing module is also used to construct a scoring function based on the disease risk intensity field prediction results, the cognitive uncertainty, the observation quality factor and the flight sampling cost, and to use the decrease in the joint posterior variance of the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field after executing the candidate trajectory as the information gain index, and to perform constrained optimal experimental design path planning for the resampling candidate region to obtain the resampling trajectory. The edge computing module is also used to include the information gain of the corresponding candidate region in the path planning objective function when the observation quality factor is greater than the preset quality gate threshold. The control and triggering module is connected to the edge computing module and the edge execution platform respectively. It is used to convert the complex sampling trajectory into flight control commands and send them to the flight control of the edge execution platform to control the edge execution platform to perform complex sampling. When the UAV reaches the target sampling position and meets the preset sampling triggering conditions, it outputs a trigger signal to control the sensors in the edge execution platform to perform complex sampling. At the same time, it synchronously records the sampling timestamp, attitude parameters and position parameters corresponding to the complex sampling. The cloud-based joint inversion module is also used to receive new observation data obtained from resampling, incrementally update the spatial heterogeneity resistance parameter field and the time-varying infection source intensity field, and feed back the updated spatial heterogeneity resistance parameter field and the updated time-varying infection source intensity field to the edge computing module, so as to form a closed-loop iteration of forward prediction, uncertainty quantification, resampling candidate region determination, path planning, resampling execution and incremental update.

9. The information detection active resampling system based on intelligent sensing and source-destination decoupling assimilation according to claim 8, characterized in that, The end-side execution platform includes a flight controller, a positioning module, a lidar or ToF module, and a multimodal sensor assembly; The multimodal sensor assembly includes a visible light camera, a multispectral camera, and a thermal infrared sensor; The positioning module is used to output position parameters corresponding to the multimodal observation data; The flight controller is used to output attitude parameters corresponding to the multimodal observation data; The lidar or ToF module is used to output flight altitude-related information and provide ranging input for terrain-following control; The flight controller is also used to control the end-side execution platform to fly along the complex sampling trajectory according to the MAVLink flight control commands issued by the control and triggering module.

10. The information detection active resampling system based on intelligent sensing and source-destination decoupling assimilation according to claim 8, characterized in that, The control and triggering module includes a MAVLink control submodule, a GPIO triggering submodule, and a synchronization recording submodule; The MAVLink control submodule is used to encapsulate the complex sampled trajectory into an MAVLink message containing waypoint control information and send it to the flight control of the end-side execution platform; The GPIO trigger submodule is used to output a trigger pulse when the end-side execution platform enters the neighborhood of the target waypoint, the ground speed is not greater than a preset ground speed threshold, the attitude angular velocity is not greater than a preset stability threshold, and the observation quality factor is not lower than a preset quality threshold, so as to trigger the sensors in the end-side execution platform to perform synchronous sampling. The synchronous recording submodule is used to synchronously record the sampling timestamp, attitude parameters, and position parameters when the GPIO is triggered, and to associate and store the sampling timestamp, attitude parameters, and position parameters with the new observation data obtained by resampling.

Images