Intelligent system for vertical field of agriculture based on multi-agent cooperation and tool enhancement

Abstract

The application relates to the field of agricultural intelligent technology, in particular to an agricultural vertical field intelligent system based on multi-agent cooperation and tool enhancement, which comprises a state construction module, a competition falsification module, an active probe module and a path synthesis module. The state construction module is used for generating plot state information; the competition falsification module is used for generating a candidate abnormal reason set, key missing evidence and an abnormal reason ranking result; the active probe module is used for supplementarily collecting key missing evidence corresponding data and updating the plot state information; and the path synthesis module is used for generating a farm treatment path. Through the application, dynamic screening, supplementary evidence and treatment decision of agricultural abnormal reasons can be realized, and the consistency of the results and actual production conditions is improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent technology, and more specifically to an intelligent system for agricultural vertical fields based on multi-agent collaboration and tool enhancement. Background Technology

[0002] As agricultural production gradually moves towards digitalization and intelligence, intelligent systems targeting vertical agricultural sectors are beginning to undertake tasks such as pest and disease identification, anomaly analysis, and agricultural decision support. These technologies are of great significance for improving the timeliness of agricultural production management, reducing reliance on experience, and promoting the standardization of agricultural technical services, especially in areas such as crop anomaly identification, field management assistance, and the generation of treatment plans.

[0003] In existing technologies, many agricultural intelligent systems are mainly built on general question-answering models or single recognition models. While they can perform preliminary processing of input images or text, they still have significant limitations when faced with complex crop anomalies, incomplete field information, and frequent changes in environmental factors. On the one hand, existing systems often lack the ability to model a unified state around crop identity, growth stage, symptom evolution, and the relationship between agricultural events, resulting in scattered input data and difficulty in forming a stable basis for judgment. On the other hand, existing systems typically lack a categorized competitive analysis mechanism for diseases, pests, pesticide damage, nutrient imbalances, and abiotic stresses, which can easily lead to confusion about the causes of anomalies. Furthermore, when evidence is insufficient, existing systems usually cannot proactively request high-value supplementary data collection, nor can they integrate crop safety, meteorological conditions, historical agricultural conflicts, and local implementation conditions into the decision-making process in subsequent treatment stages, resulting in output results that are out of touch with actual production conditions. Summary of the Invention

[0004] This invention provides an intelligent system for the vertical agricultural domain based on multi-agent collaboration and tool enhancement, which is used to at least solve the problem of how to accurately screen the causes of agricultural anomalies and generate executable agricultural treatment paths under the condition of incomplete evidence of agricultural anomalies.

[0005] This invention provides an intelligent system for agricultural vertical fields based on multi-agent cooperation and tool enhancement. The system includes: The status construction module is used to receive agricultural multimodal data and generate plot status information including crop identity, growth stage, symptom association information and agricultural event association information; The competitive falsification module is used to call multiple anomaly analysis agents to perform case matching, mechanism constraints and conflict analysis based on the crop-specific knowledge base and plot status information corresponding to the crop identity, and generate a set of candidate anomaly causes, key missing evidence and anomaly cause ranking results. The active probe module is used to call the data acquisition tool to supplement the data corresponding to the key missing evidence when the score difference between the top two candidate anomalies in the anomaly cause ranking results is lower than a preset threshold, update the land parcel status information and generate disposal constraint information. The path synthesis module is used to generate agricultural disposal paths based on the candidate anomaly cause set and disposal constraint information, combined with crop safety constraints, meteorological window constraints, historical agricultural conflict constraints, and local executability constraints.

[0006] In one possible implementation, agricultural multimodal data includes crop information, symptom information, environmental information, and agricultural event information; the state construction module is used to determine crop identity, growth stage, symptom association information, and agricultural event association information based on the agricultural multimodal data.

[0007] In one possible implementation, the state construction module is used to generate symptom association information based on the location of symptom occurrence, symptom morphology, symptom distribution, and symptom occurrence time; the state construction module is also used to generate agricultural event association information based on the time sequence of pesticide application events, fertilization events, and irrigation events.

[0008] In one possible implementation, multiple anomaly analysis agents include a disease analysis agent, an insect pest analysis agent, a nutrient imbalance analysis agent, a pesticide damage analysis agent, and an abiotic stress analysis agent; a competitive falsification module is used to generate a set of candidate anomaly causes and key missing evidence based on the analysis results of each anomaly analysis agent.

[0009] In one possible implementation, the competitive falsification module is used to determine the score of each candidate abnormal cause based on the case matching results, mechanism constraint results, temporal consistency between the symptom occurrence time and the agricultural event time, consistency of agricultural event type, and conflict analysis results, and to generate an abnormal cause ranking result based on the scores of each candidate abnormal cause.

[0010] In one possible implementation, the active probe module is used to call the data acquisition tool to supplement the data corresponding to the key missing evidence when the score difference between the top two candidate anomaly causes in the anomaly cause ranking results is lower than a first threshold and the number of key missing evidences is greater than a second threshold, and then update the land parcel status information based on the supplemented data.

[0011] In one possible implementation, the active probe module is used to regenerate a set of candidate anomaly causes based on the updated land parcel status information, and generate disposal constraint information based on the updated set of candidate anomaly causes; the disposal constraint information includes a set of allowed actions and a set of prohibited actions, and the set of prohibited actions is the union of the prohibited actions corresponding to each candidate anomaly cause in the set of candidate anomaly causes.

[0012] In one possible implementation, the path synthesis module is used to determine the meteorological window that meets the meteorological window constraints based on the set of candidate abnormal causes, the reproductive stage, agricultural event association information and environmental information, and to recall the action template from the historical case database.

[0013] In one possible implementation, the path synthesis module is used to adjust the disposal action template according to the disposal constraint information. The adjustment includes node insertion, node deletion, node postponement, and node rearrangement to generate an agricultural disposal path that meets crop safety constraints, weather window constraints, historical agricultural conflict constraints, and local executability constraints.

[0014] In one possible implementation, agricultural action paths include immediate action, observation action, delayed action, and prohibited action; local enforceability constraints are used to characterize the availability of agricultural inputs, operational equipment conditions, and planting scale.

[0015] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By using a state construction module to uniformly receive, parse, and organize agricultural multimodal data, it achieves an integrated expression of crop identity, growth stage, symptom association information, and agricultural event association information, avoiding the problems of scattered input information and unstable judgment basis in existing technologies.

[0016] By calling multiple anomaly analysis agents through a competitive falsification module and combining them with a crop-specific knowledge base corresponding to the crop identity to perform case matching, mechanism constraints, and conflict analysis, hierarchical screening and ranking of candidate anomaly causes are achieved, reducing the problems of anomaly cause confusion and single-path misjudgment in existing technologies.

[0017] By using the active probe module to call data acquisition tools to supplement key missing evidence when the distinction between the first two candidate anomalies is insufficient, dynamic supplementation of evidence and status updates are achieved in scenarios with insufficient evidence, avoiding the problem of existing technologies directly outputting conclusions when information is incomplete.

[0018] By combining the candidate anomaly cause set with the disposal constraint information through the path synthesis module, and simultaneously incorporating crop safety constraints, meteorological window constraints, historical agricultural conflict constraints, and local executability constraints, agricultural disposal paths oriented to real production conditions are generated. This enables the output results to not only explain anomalies but also guide subsequent processing.

[0019] Through the synergistic cooperation of the above-mentioned technical means, a complete closed loop from state modeling, anomaly analysis, dynamic supplementary verification to path output is achieved. Attached Figure Description

[0020] Figure 1 This is a block diagram of the system modules of the present invention; Figure 2 This is a schematic diagram showing the correspondence between symptoms and agricultural events in an embodiment of the present invention; Figure 3 This is a comparison chart of the comprehensive scores of candidate anomaly causes in an embodiment of the present invention; Figure 4 This is a graph showing the changes in scores for abnormal causes before and after supplementary sampling in an embodiment of the present invention; Figure 5 This is a graph showing the changes in key indicators after treatment in an embodiment of the present invention. Detailed Implementation

[0021] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0022] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0023] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0024] Multi-agent collaboration and tool enhancement is an intelligent processing mechanism for complex tasks. Multi-agent collaboration emphasizes breaking down the same task into multiple intelligent processing agents with clearly defined responsibilities and different processing focuses. Each agent undertakes functions such as information parsing, knowledge matching, anomaly identification, evidence supplementation, constraint verification, and path generation, and then forms a unified output through interactive collaboration. Tool enhancement emphasizes introducing external data acquisition tools, knowledge bases, historical case databases, environmental information interfaces, and agricultural record tools into the intelligent processing process. This allows the intelligent system to go beyond static input content and combine real-time data, structured knowledge, and scenario constraints to complete more realistic analysis and decision-making. For the agricultural vertical field, multi-agent collaboration and tool enhancement is not simply adding processing steps, but rather providing a division-of-labor, interactive, and closed-loop foundation for agricultural anomaly identification, evidence supplementation, and decision-making. Based on this, this invention proposes an intelligent system for the agricultural vertical field based on multi-agent collaboration and tool enhancement. Through processes such as state construction, competitive falsification, active probes, and path synthesis, it achieves hierarchical analysis and response path generation for agricultural anomalies.

[0025] like Figure 1 As shown, an intelligent system for agricultural vertical fields based on multi-agent cooperation and tool enhancement is disclosed. The system includes: The status construction module is used to receive agricultural multimodal data and generate plot status information including crop identity, growth stage, symptom association information and agricultural event association information; In this embodiment, the state construction module receives agricultural multimodal data and converts it into plot state information that can be directly accessed in subsequent anomaly analysis stages. Specifically, the state construction module first performs time alignment, field normalization, and semantic completion on the raw data from different data sources, and then extracts crop identity, growth stage, symptom association information, and agricultural event association information. After extraction, the state construction module summarizes the above content according to a unified field structure to form plot state information for individual plots. This plot state information retains both the current crop state of the plot and the correlation between symptom changes and agricultural operations, facilitating subsequent stages of case matching, anomaly cause screening, and treatment path generation around the same treatment object.

[0026] Agricultural multimodal data includes crop information, symptom information, environmental information, and agricultural event information; the state construction module is used to determine crop identity, growth stage, symptom association information, and agricultural event association information based on agricultural multimodal data.

[0027] In one embodiment, agricultural multimodal data includes crop information, symptom information, environmental information, and agricultural information. This embodiment adds a specific composition of agricultural multimodal data and a process for determining plot status information compared to the foregoing. This limitation is introduced because receiving only images or only text can easily lead to unstable crop type judgment, significant deviations in growth stage judgment, or missing agricultural background information, thus affecting subsequent anomaly analysis results. Therefore, after receiving agricultural multimodal data, the status construction module first establishes a plot-level data receiving unit to merge images, text, environmental records, and agricultural records received from the same plot within a preset time window. Crop information is used to characterize the planted object and its basic attributes, and may include crop name, variety name, sowing time, transplanting time, planting time, and planting area. Symptom information is used to characterize the external manifestations of plant abnormalities in the plot, and may include lesion images, leaf color change descriptions, wilting descriptions, leaf drop descriptions, fruit surface abnormality descriptions, and the extent of abnormality. Environmental information is used to characterize the external conditions when an abnormality occurs, and may include temperature, humidity, rainfall, light, wind speed, soil moisture, and greenhouse ventilation status.

[0028] Agricultural information is used to characterize recent human interventions and can include records of pesticide application, fertilization, irrigation, pruning, and mulching. When processing this data, the status construction module first aggregates it by plot number or geographical location, then sorts it by collection time, and performs completion checks on missing fields. The completion check follows these principles: prioritizing structured fields directly input by the user, followed by image recognition and text parsing results; if conflicts exist for the same field, the more recent and stable data takes precedence.

[0029] After data collection is complete, the state construction module first determines the crop identity. Crop identity can be determined by first reading the crop name from the crop information; if the crop name is missing, it can be identified based on plant morphology, leaf outline, and fruit appearance features in the symptom images, and verified against common crop types in the planting area. Next, the state construction module determines the growth stage. The growth stage can be estimated based on sowing, transplanting, or planting time, or corrected by combining information such as leaf age, plant height, flower buds, and fruit size in the images. After determining the crop identity and growth stage, the state construction module writes the crop identity, growth stage, subsequently generated symptom association information, and agricultural event association information into the same plot's state information, and adds a collection time and data source marker to each item to facilitate subsequent assessments of data timeliness and reliability.

[0030] The status construction module is used to generate symptom association information based on the location, morphology, distribution, and time of symptom occurrence; the status construction module is also used to generate agricultural event association information based on the time sequence of pesticide application events, fertilizer application events, and irrigation events.

[0031] In one embodiment, the state construction module generates symptom association information based on the location, morphology, distribution, and timing of symptom occurrence, and generates agricultural event association information based on the chronological order of pesticide application, fertilization, and irrigation events. This embodiment adds specific methods for generating symptom association information and agricultural event association information compared to the foregoing. The reason for this limitation is that if the plot state information only stores isolated symptom fields and isolated agricultural records, subsequent stages can only read static descriptions separately, making it difficult to determine where the anomaly occurred, how it spread, and whether it has a temporal relationship with recent agricultural operations. Therefore, when generating symptom association information, the state construction module does not simply record a single symptom label, but establishes associations around the location, morphology, distribution, and timing of symptom occurrence. The location of symptom occurrence indicates which part of the leaf, stem, fruit, flower, or root the anomaly first appears in.

[0032] Symptom morphology indicates the manifestation of the abnormality, which may include yellowing, brown spots, rot, curling, wilting, perforation, dry tips, and surface attachments. Symptom distribution indicates the coverage pattern of the abnormality in individual plants and plots, which may include sporadic distribution, patchy distribution, edge-concentrated distribution, and localized clustering. Symptom occurrence time indicates the time when the user first discovered the abnormality or the earliest time in the image where the abnormality can be confirmed. When processing the above content, the state construction module first arranges multiple symptom records in chronological order, and then determines whether two adjacent symptom records belong to the same abnormal evolution process. The judgment criteria may include whether the occurrence location is continuous, whether the symptom morphology has a progressive relationship, and whether the collection time interval is within a preset continuous observation period. If it is determined to be the same abnormal evolution process, the preceding and following symptom records are established as related items, forming symptom association information. The generation method of agricultural event association information is similar.

[0033] The state construction module first sorts the pesticide application, fertilization, and irrigation events by time, then extracts the execution time, target, scope, and dosage or intensity of each event. Subsequently, the module generates agricultural event association information based on the chronological order of the events, representing the chain of agricultural operations before and after the anomaly occurred on the same plot. If multiple types of agricultural operations exist simultaneously within the same time window, a sequential relationship is established based on their execution time; if the execution times are the same, the original order is saved according to the entry time, with a concurrency flag added. After generating the symptom association information and agricultural event association information, the state construction module writes these two types of association information, along with crop identity and growth stage, into the plot state information. This way, when subsequent stages read the plot state information, they can not only obtain the static state of the current plot but also the temporal correlation between the symptom evolution process and the agricultural operation process, providing a consistent data foundation for subsequent anomaly cause analysis and response decisions.

[0034] The competitive falsification module is used to call multiple anomaly analysis agents to perform case matching, mechanism constraints and conflict analysis based on the crop-specific knowledge base and plot status information corresponding to the crop identity, and generate a set of candidate anomaly causes, key missing evidence and anomaly cause ranking results. In this embodiment, the competitive falsification module receives the plot status information output by the state construction module and generates a candidate set of anomaly causes, key missing evidence, and anomaly cause ranking results based on the plot status information and the crop-specific knowledge base corresponding to the crop identity. Specifically, the competitive falsification module first selects the corresponding crop-specific knowledge base according to the crop identity, and then calls multiple anomaly analysis agents to analyze the plot status information respectively, obtaining multiple anomaly analysis results. Subsequently, the competitive falsification module merges the multiple anomaly analysis results, filtering out anomaly causes that are obviously inconsistent with the current crop identity, growth stage, and agricultural event association information, and retaining anomaly causes that meet the basic matching conditions as a candidate set of anomaly causes. The competitive falsification module continues to perform evidence integrity checks on the candidate set of anomaly causes, extracts missing fields that affect the further differentiation of anomaly causes as key missing evidence, and generates anomaly cause ranking results based on the matching and conflict situations of each candidate anomaly cause. The anomaly cause ranking results are used to provide triggering basis for the subsequent active probe stage, and the key missing evidence is used to provide clear data requirements for the subsequent data supplementation stage.

[0035] Multiple anomaly analysis agents include disease analysis agents, pest analysis agents, nutrient imbalance analysis agents, pesticide damage analysis agents, and abiotic stress analysis agents; the competitive falsification module is used to generate a set of candidate anomaly causes and key missing evidence based on the analysis results of each anomaly analysis agent.

[0036] In one embodiment, multiple anomaly analysis agents include disease analysis agents, pest analysis agents, nutrient imbalance analysis agents, pesticide damage analysis agents, and abiotic stress analysis agents. This embodiment adds specific components of the anomaly analysis agents and their analysis boundaries compared to the foregoing. This limitation is introduced because agricultural anomalies have significant type differences; the same symptom may simultaneously correspond to multiple sources such as diseases, pests, pesticide damage, fertilizer damage, drought stress, and waterlogging stress. If only a single analysis module is used for unified judgment, bias can easily form in the initial screening stage of anomalies, leading to unstable subsequent ranking results. Therefore, after switching to the crop-specific knowledge base, the competitive falsification module first calls different anomaly analysis agents according to the anomaly cause category.

[0037] The disease analysis AI primarily handles anomalies caused by fungi, bacteria, viruses, etc., focusing on reading the location, morphology, direction of symptom spread, humidity conditions, and historical pesticide application information. The insect pest analysis AI primarily handles anomalies caused by chewing insects, piercing-sucking insects, and vector-borne diseases, focusing on reading insect holes, leaf curling, chlorotic patches, insect images, insect population distribution, and recent pest control information. The nutrient imbalance analysis AI primarily handles anomalies caused by nutrient deficiencies, over-fertilization, and nutrient imbalances, focusing on reading leaf color changes, affected leaf positions, fertilization records, and irrigation background.

[0038] The pesticide damage analysis agent primarily handles abnormalities such as burns, curling, chlorosis, and deformities resulting from pesticide application, focusing on the application event, application time, pesticide formulation, pesticide concentration, and symptom onset time. The abiotic stress analysis agent primarily handles abnormalities such as high temperature, low temperature, waterlogging, drought, salinity, and mechanical damage, focusing on environmental information and non-pathogenic damage characteristics. After receiving the analysis results from each anomaly analysis agent, the competitive falsification module does not directly output a single conclusion. Instead, it summarizes the anomaly causes, supporting evidence, and non-supporting evidence provided by each anomaly analysis agent to form a candidate set of anomaly causes. For anomalies deemed potentially valid by multiple anomaly analysis agents, the competitive falsification module retains them as priority candidates. For anomalies valid only by a single anomaly analysis agent and clearly conflicting with crop identity or growth stage, the competitive falsification module lowers their priority or eliminates them directly.

[0039] After this processing, the competitive falsification module further checks whether the necessary evidence for each candidate anomaly is complete. For example, disease analysis typically requires close-up symptom data and environmental humidity information, pesticide damage analysis typically requires application time and pesticide name, and nutrient imbalance analysis typically requires fertilization records and leaf position of occurrence. Missing necessary evidence is recorded as critical missing evidence for subsequent active probe phase re-collection. In this way, the competitive falsification module can first perform triage at the anomaly cause type level, and then perform preliminary convergence at the evidence level, avoiding confusion between different anomalies in the same analysis path.

[0040] The competitive falsification module is used to determine the score of each candidate abnormal cause based on the case matching results, mechanism constraint results, temporal consistency between the symptom occurrence time and the agricultural event time, consistency of agricultural event type, and conflict analysis results, and to generate an abnormal cause ranking result based on the score of each candidate abnormal cause.

[0041] In one embodiment, the competitive falsification module determines the score of each candidate abnormal cause based on case matching results, mechanism constraint results, temporal consistency between symptom onset time and agricultural event time, consistency of agricultural event type, and conflict analysis results. It then generates an abnormal cause ranking result based on these scores. This embodiment adds a method for generating the candidate abnormal cause ranking result compared to the aforementioned content. This limitation is introduced because the candidate abnormal cause set only represents the range of possible abnormal causes. Without further clarifying the ranking rules, the subsequent active probe stage cannot determine which abnormal causes should be prioritized, nor can it determine whether the conditions for directly entering the path synthesis stage have been met. Therefore, after forming the candidate abnormal cause set, the competitive falsification module calculates the case matching result, mechanism constraint result, temporal consistency relationship, agricultural event type consistency relationship, and conflict analysis result for each candidate abnormal cause.

[0042] Case matching results characterize the similarity between the current plot status information and similar abnormal scenarios in historical cases. Specifically, this can be obtained through multi-field comparison based on crop identity, growth stage, symptom location, symptom morphology, symptom distribution, environmental information, and agricultural information. Mechanistic constraint results characterize whether candidate anomaly causes conform to the current crop stage and anomaly evolution patterns. For example, some diseases typically spread rapidly under high humidity conditions, some pesticide damage usually appears shortly after application, and some nutrient imbalances usually manifest first in older or newer leaves. The temporal consistency between symptom onset time and agricultural event time characterizes whether there is a reasonable chronological relationship between the anomaly occurrence time and events such as pesticide application, fertilization, and irrigation. This consistency can be determined based on the interval between the first symptom appearance time and the execution time of the relevant agricultural event; when the interval is within a preset valid range, it can be considered temporally consistent.

[0043] The preset effective range can be set based on the common occurrence cycle of the corresponding anomaly. For example, pesticide damage usually appears within a short period after application, while fertilizer damage and waterlogging damage may occur over a longer period after fertilization or irrigation. The consistency of agricultural event types characterizes whether there is a category correspondence between the current candidate anomaly and related agricultural event types. For example, pesticide damage is more strongly associated with pesticide application events, nutrient imbalance is more strongly associated with fertilizer application events, and waterlogging is more strongly associated with irrigation or rainfall events. Conflict analysis results characterize whether there are significant contradictions between the current candidate anomaly and existing evidence. For example, if a candidate anomaly requires high humidity but environmental information shows prolonged dryness, a conflict exists; or if a candidate anomaly usually occurs first at the leaf margin but current symptoms are concentrated in the roots, a conflict exists. After completing the above analysis, the competitive falsification module generates a comprehensive score for each candidate anomaly. A higher score indicates greater consistency between the candidate anomaly and the current plot status information, fewer conflicts, and more sufficient evidence support. After calculating the scores for all candidate anomalies, the competitive falsification module generates a ranking of the anomalies based on their scores. If two adjacent candidate anomaly causes have similar scores, they are retained as key targets for differentiation in the subsequent active probe stage. If the score of the first candidate anomaly cause is significantly higher than that of the subsequent candidate anomaly causes, and the number of key missing evidence is small, the subsequent process can directly use this ranking result as the priority judgment basis. Through the above processing, the candidate anomaly cause set, key missing evidence, and anomaly cause ranking results can form a closed data loop that is interconnected.

[0044] The active probe module is used to call the data acquisition tool to supplement the data corresponding to the key missing evidence when the score difference between the top two candidate anomalies in the anomaly cause ranking results is lower than a preset threshold, update the land parcel status information and generate disposal constraint information. In one embodiment, the active probe module receives the anomaly cause ranking results and key missing evidence, and determines whether a supplementary data collection phase is needed. Specifically, the active probe module first reads the two candidate anomaly causes with the highest ranking and their corresponding scores. If the difference between the two scores is lower than a preset threshold, it is determined that the current anomaly cause lacks sufficient discriminative power, and new evidence data needs to be collected. Subsequently, the active probe module determines the supplementary collection target, supplementary collection order, and supplementary collection tool based on the key missing evidence, calls the data acquisition tool to obtain the new data, and writes the new data into the plot status information. After writing is completed, the active probe module performs consistency and integrity checks on the updated plot status information, then generates disposal constraint information and sends the disposal constraint information to the path synthesis module to constrain the generation range of subsequent agricultural disposal paths.

[0045] The active probe module is used to call the data acquisition tool to supplement the data corresponding to the key missing evidence when the score difference between the top two candidate anomalies in the anomaly cause ranking results is lower than the first threshold and the number of key missing evidences is greater than the second threshold, and to update the land parcel status information based on the supplemented data.

[0046] In one embodiment, the active probe module is used to, when the score difference between the top two candidate anomaly causes in the anomaly cause ranking results is lower than a first threshold and the number of key missing evidences is greater than a second threshold, invoke a data acquisition tool to supplement the data corresponding to the key missing evidence, and update the land parcel status information based on the supplemented data. This implementation adds active probe triggering conditions, a method for determining supplementary data collection targets, and status update rules compared to the foregoing. The reason for introducing this limitation is that using only the score difference between the top two candidate anomaly causes as the triggering basis may cause unnecessary duplicate data collection when there is very little key missing evidence; using only the number of key missing evidences as the triggering basis may introduce redundant steps when the candidate anomaly causes are already relatively clear. Therefore, the active probe module uses both the first and second thresholds as dual triggering conditions.

[0047] The first threshold characterizes the minimum distinguishability between the top two candidate anomalies. A score difference below the first threshold indicates that existing evidence is insufficient to reliably distinguish the two candidate anomalies. The second threshold characterizes the minimum required number of critical missing evidence. A number of critical missing evidence exceeding the second threshold indicates a significant gap that could affect the anomaly diagnosis. The first threshold can be set based on the average score difference between the top and second-ranked anomalies in historical cases under stable conditions, or it can be set separately for different crop categories and anomaly category categories. The second threshold can be set based on the impact of missing fields in the current plot status information on anomaly differentiation; for example, missing application time, missing close-up images of symptoms, or missing soil moisture information can all be counted as critical missing evidence. After meeting the dual triggering conditions, the active probe module first classifies the critical missing evidence. Classification methods can include image-based evidence, environmental evidence, agricultural evidence, and supplementary text evidence. Image-based evidence typically includes close-up views of lesions, images of the underside of leaves, images of the base of stems, images of the surface of fruits, and images of the root zone; environmental evidence typically includes temperature, humidity, rainfall, soil moisture, and ventilation conditions inside the greenhouse; agricultural evidence typically includes the time of pesticide application, the name of the pesticide, the time of fertilization, the type of fertilizer, and the time of irrigation; supplementary textual evidence typically includes the time when the anomaly was first discovered, the extent of the anomaly, and whether the anomaly is concentrated in a certain area.

[0048] After classification, the active probe module determines the order of data collection based on the impact of key missing evidence on distinguishing candidate anomaly causes. Key missing evidence with a higher impact is collected first; for example, when pesticide damage and disease are difficult to distinguish, application time and pesticide name take precedence over general environmental descriptions; when nutrient imbalance and abiotic stress are difficult to distinguish, fertilization records and irrigation records take precedence over general textual supplementation. Subsequently, the active probe module calls the appropriate data acquisition tools according to the evidence category. Image-based evidence can use photography tools or image upload tools; environmental evidence can use meteorological reading tools or sensor reading tools; and agricultural evidence can use agricultural record query tools or manual data entry tools. After data collection is completed, the active probe module writes the new data into the plot status information according to field categories, retaining the collection time, data source, and corresponding key missing evidence markers. If new data conflicts with existing fields, data with more recent collection time, more direct source, and more relevant to the current candidate anomaly cause analysis is retained, while the conflicting fields are recorded in the status update record for subsequent verification. Through the above processing, the active probe module can form a complete closed loop in three aspects: supplementary sampling triggering, supplementary sampling target selection, and status update, avoiding indiscriminate supplementary sampling and supplementary sampling without basis.

[0049] The active probe module is used to regenerate a set of candidate anomaly causes based on the updated land parcel status information, and generate disposal constraint information based on the updated set of candidate anomaly causes. The disposal constraint information includes a set of allowed actions and a set of prohibited actions. The set of prohibited actions is the union of the prohibited actions corresponding to each candidate anomaly cause in the set of candidate anomaly causes.

[0050] In one embodiment, the active probe module is used to regenerate a set of candidate anomaly causes based on the updated land parcel status information, and generate disposal constraint information based on the updated set of candidate anomaly causes. The disposal constraint information includes a set of allowed actions and a set of prohibited actions, where the set of prohibited actions is the union of the prohibited actions corresponding to each candidate anomaly cause in the candidate anomaly cause set. This embodiment adds a re-analysis process after supplementary data collection and a process for generating disposal constraint information compared to the foregoing. The reason for introducing this constraint is that the purpose of supplementary data collection is not only to update the land parcel status information, but more importantly, to use the new data to narrow down the range of anomaly causes and to provide an executable boundary before the anomaly cause is completely and uniquely determined, so as to avoid directly generating disposal actions that conflict with existing evidence in the subsequent path synthesis stage. Based on this, after the land parcel status information is updated, the active probe module does not directly end the processing, but instead inputs the updated land parcel status information again into the competitive falsification process.

[0051] During reanalysis, the active probe module can either call the same anomaly analysis agent as the previous round, or only call the anomaly analysis agents corresponding to the top two candidate anomaly causes, to reduce unnecessary redundant calculations. The reanalysis results are used to regenerate the candidate anomaly cause set. If the updated new data can clearly distinguish the two main candidate anomaly causes, the regenerated candidate anomaly cause set will usually be reduced to a smaller number; if the new data is still insufficient to make a unique judgment, it can at least narrow down the range of some invalid anomaly causes. Subsequently, the active probe module generates disposal constraint information based on the updated candidate anomaly cause set. The set of allowed actions represents the range of actions that can enter the path planning stage under the current evidence conditions, while the set of prohibited actions represents the range of actions that cannot be executed under the current evidence conditions. The method for determining the set of allowed actions may include: extracting actions related to the current candidate anomaly cause and not causing significant secondary risks from the historical case database or the preset disposal rule database, such as observation, isolation, local re-shooting, suspending additional input, recording anomaly expansion, etc. The method for determining the set of prohibited actions may include: extracting the prohibited actions corresponding to each candidate anomaly cause, and then taking the union of the prohibited actions. For example, when the set of candidate abnormal causes includes both pesticide damage and nutrient imbalance, it may be necessary to prohibit the continued application of the same type of pesticide; when the set of candidate abnormal causes includes both disease and waterlogging, it may be necessary to prohibit the continued high-frequency irrigation.

[0052] The reason for using the union approach is that, when the cause of an anomaly is not uniquely determined, if an action corresponding to a candidate anomaly cause may lead to significant risk, that action should not be included in the subsequent path synthesis results. To avoid an overly broad set of prohibited actions, the active probe module can also add effective conditions to prohibited actions, such as limiting their effectiveness to the current reproductive stage, current environmental conditions, or current agricultural context. After generating the set of allowed and prohibited actions, the active probe module writes both into the disposal constraint information and sends the disposal constraint information to the path synthesis module. In this way, when generating subsequent agricultural disposal paths, the path synthesis module can only select actions from the set of allowed actions and must not generate actions from the set of prohibited actions. Through the above processing, the updated plot status information, the regenerated set of candidate anomaly causes, and the disposal constraint information can form a sequential processing chain, enabling the active probe stage to not only perform supplementary data collection but also risk boundary convergence.

[0053] The path synthesis module is used to generate agricultural disposal paths based on the candidate anomaly cause set and disposal constraint information, combined with crop safety constraints, meteorological window constraints, historical agricultural conflict constraints, and local executability constraints.

[0054] In this embodiment, the path synthesis module receives a set of candidate anomaly causes and disposal constraint information, and generates an agricultural disposal path based on crop safety constraints, meteorological window constraints, historical agricultural conflict constraints, and local executability constraints. Specifically, the path synthesis module first reads the anomaly cause type, growth stage, agricultural event association information, and environmental information from the candidate anomaly cause set to determine the time range and action range within which the current plot can enter the disposal stage. Then, it recalls the disposal action template closest to the current plot's state from the historical case database. Subsequently, the path synthesis module filters, supplements, and rearranges the disposal action templates according to the disposal constraint information to obtain an agricultural disposal path that meets the current constraints. The generated agricultural disposal path continues to be output to the execution stage to guide plot management, anomaly review, and risk avoidance.

[0055] The path synthesis module is used to determine the meteorological window that meets the meteorological window constraints based on the candidate abnormal cause set, reproductive stage, agricultural event association information and environmental information, and to recall the disposal action template from the historical case database.

[0056] In one embodiment, the path synthesis module is used to determine a meteorological window that meets the meteorological window constraints based on the candidate anomaly cause set, growth stage, agricultural event association information, and environmental information, and to recall the treatment action template from the historical case database. This implementation adds a meteorological window determination process and a treatment action template recall process compared to the foregoing. The reason for introducing this constraint is that the timing of treatment for the same anomaly cause differs depending on the growth stage, weather conditions, and agricultural background. If an executable meteorological window is not determined first, the subsequently output agricultural treatment path may conflict with the field conditions; for example, scheduling spraying actions during a period of continuous rainfall, or scheduling foliar treatment actions prone to pesticide damage during a period of high temperature. Therefore, after receiving the candidate anomaly cause set, the path synthesis module first reads the corresponding meteorological requirement rules according to the category of the candidate anomaly cause.

[0057] Meteorological demand rules are used to define the appropriate action to be taken under specific temperature, humidity, rainfall, wind speed, and light conditions for a given type of anomaly. Environmental information provides the weather conditions for the current plot during the current time period and subsequent preset time periods, including temperature, humidity, rainfall probability, rainfall amount, wind speed, sunshine duration, and soil moisture. Growth stage determines whether a particular crop is suitable for a specific agricultural action; for example, the range of actionable actions typically differs between the seedling, flowering, fruiting, and maturity stages. Agricultural event association information determines whether recent operations such as pesticide application, fertilization, and irrigation have occurred, in order to determine whether new actions need to avoid the impact cycle of existing agricultural activities. When determining the meteorological window, the path synthesis module first generates candidate weather intervals arranged chronologically according to the environmental information, then compares each candidate interval with the meteorological demand rules to filter out candidate time periods that meet the basic execution conditions. If multiple candidate time periods exist, a secondary filtering is performed by combining the growth stage and agricultural event association information. For example, when a candidate anomaly requires foliar spraying, the path synthesis module will first exclude periods of continuous rainfall, then exclude periods that are sensitive to high temperatures and periods that have recently completed similar spraying, thus retaining meteorological windows that meet the meteorological window constraints.

[0058] Once the meteorological window is determined, the path synthesis module retrieves action templates from the historical case database based on the candidate anomaly cause set, growth stage, meteorological window, and agricultural event association information. The action templates stored in the historical case database are not single actions, but rather action sequences recorded around specific anomaly causes, which may include observation, isolation, cessation of additional treatment, local treatment, whole-field treatment, review, and escalation treatment. When retrieving action templates, the path synthesis module first filters historical cases by crop category, then matches them by anomaly cause type, growth stage, and agricultural background, and finally retains one or more action templates that are closest to the current plot conditions. If multiple action templates meet the conditions, the template that is more consistent with the current environmental information and agricultural event association information is selected first. Through this process, the path synthesis stage can first determine the executable time, and then select action templates that match the current plot status, providing stable input for subsequent template adjustments.

[0059] The path synthesis module is used to adjust the action template based on the disposal constraint information. The adjustment includes node insertion, node deletion, node postponement, and node rearrangement to generate agricultural disposal paths that meet crop safety constraints, weather window constraints, historical agricultural conflict constraints, and local executability constraints.

[0060] In one embodiment, the path synthesis module adjusts the action template based on disposal constraint information. This adjustment includes node insertion, deletion, postponement, and rearrangement to generate an agricultural disposal path that satisfies crop safety constraints, weather window constraints, historical agricultural conflict constraints, and local executability constraints. This implementation adds a method for adjusting the action template and rules for generating agricultural disposal paths compared to the foregoing. This limitation is introduced because while the action templates in the historical case database provide a reference framework, different plots differ in growth stage, weather window, recent agricultural background, and execution resources, making it impossible to directly copy historical templates. Therefore, after recalling the action template, the path synthesis module first checks each action node in the template based on the disposal constraint information. The disposal constraint information limits which actions can and cannot be performed on the current plot.

[0061] The path synthesis module first traverses all action nodes in the action template. If an action node is a prohibited action, it is marked as a deletion object. If an action node is not a prohibited action, but its execution time falls outside the non-executable weather window, it is marked as a postponed object. If an action node lacks execution prerequisites, such as the template requiring the completion of review image acquisition first, but the current template does not include this action, the review image acquisition is used as an insertion object. If the order of multiple action nodes conflicts with the current agricultural event's associated information, such as the template arranging pesticide application before review, but the current plot needs to be reviewed first before deciding whether to apply pesticide, the corresponding node is marked as a rearrangement object. When inserting execution nodes, the path synthesis module can insert actions such as on-site review, additional sampling, suspension of input, local isolation, and manual confirmation. When deleting execution nodes, it can delete actions that are prohibited from execution, incompatible with the current crop stage, or have too high a risk under the current weather conditions. When delaying execution nodes, it can move actions that need to wait for the weather to improve or for the impact of existing agricultural activities to subside to subsequent meteorological windows. When rearranging execution nodes, it can rearrange actions according to risk control priority, execution dependencies, and on-site operability sequence.

[0062] Crop safety constraints limit whether certain actions are permitted at the current growth stage, such as actions to avoid damaging flowers during flowering and actions to avoid affecting fruit quality during fruiting. Historical agricultural conflict constraints limit whether there are time, pesticide, or operational conflicts between the path to be generated and recently executed pesticide, fertilization, or irrigation actions. Local executability constraints limit whether the current plot has the resource conditions to perform a certain type of action, such as whether the corresponding inputs are available locally, whether the corresponding operating equipment is available on site, and whether the current planting scale supports high-frequency manual inspections. After completing the adjustment of all nodes, the path synthesis module connects the retained action nodes according to the new execution order to generate agricultural disposal paths. Through the above processing, agricultural disposal paths can retain the disposal experience of historical templates while adapting to the actual conditions of the current plot, avoiding execution deviations caused by directly copying historical solutions.

[0063] Agricultural action pathways include immediate action, observation action, delayed action, and prohibited action; local enforceability constraints are used to characterize the availability of agricultural inputs, operational equipment conditions, and planting scale.

[0064] In one embodiment, the agricultural action path includes immediate action, observation action, delayed action, and prohibited action; local executability constraints are used to characterize the availability of agricultural inputs, operational equipment conditions, and planting scale. This implementation adds an output structure for the agricultural action path and a specific meaning for the local executability constraints compared to the foregoing. The reason for introducing this constraint is that if only a single action list is output, on-site personnel will find it difficult to distinguish which actions need to be executed immediately, which require continuous observation, which require waiting for an opportune moment, and which actions are clearly prohibited, easily affecting the execution order and risk control. Therefore, after generating the agricultural action path, the path synthesis module categorizes all actions according to their execution nature.

[0065] Immediate actions are those that can be performed at the current moment without relying on subsequent supplementary conditions, such as stopping a high-risk input, marking abnormal areas, implementing local isolation, and recording the current abnormal status. Observation actions are those that require continuous tracking or reconfirmation over a certain period, such as rechecking symptom changes after a preset interval, taking additional images of specific areas, recording the extent of the abnormality, and recording weather changes. Delayed actions are those that are not suitable for immediate execution but can be implemented after subsequent conditions are met, such as waiting for rainfall to end before spraying, waiting for soil moisture content to decrease before adjusting irrigation, and waiting for the end of a high-temperature period before performing foliar treatment. Prohibited actions are those that cannot be performed under the current set of candidate abnormal causes and the constraints of the action plan, such as continuing to add similar agents, implementing sensitive treatments under high-risk weather, and performing irreversible operations before completing a follow-up review.

[0066] When forming the above classifications, the path synthesis module first determines the execution nature of the action nodes based on their status after template adjustment, and then corrects them according to crop safety constraints, meteorological window constraints, and historical agricultural conflict constraints. Local executability constraints at this stage are mainly used to determine whether the actions have the practical conditions for implementation. Agricultural input availability characterizes whether the area where the current plot is located can obtain the inputs required for the treatment action. If not, the relevant action cannot be implemented immediately. Operating equipment conditions characterize whether the site has basic operating conditions such as spraying equipment, measuring equipment, sampling tools, and transportation vehicles. If the necessary equipment is lacking, the relevant action needs to be converted into a delayed treatment action or an observation action. Planting scale characterizes the current plot's management area and the degree of human coverage. When the planting scale is large and human resources are limited, the path synthesis module can prioritize retaining immediate treatment actions in higher-risk areas and convert some low-priority actions into observation actions or delayed treatment actions. After completing the above classifications, the path synthesis module outputs agricultural treatment paths according to execution priority, enabling on-site personnel to implement them in the order of immediate treatment, continuous observation, execution upon condition fulfillment, and prohibition of execution. Through the above processing, the output format of agricultural disposal paths can be kept consistent with on-site execution habits, while local executability constraints are truly involved in action selection and action classification, rather than merely existing as additional explanations.

[0067] In one specific embodiment, a watermelon planting plot in a facility was selected as the implementation object. The plot area was 2.4 acres, with 1320 plants planted, and the crop was in the early fruit-setting stage. At 08:30 on July 8, the on-site management personnel found that the upper leaves were curled, had irregular brown patches, and showed local chlorosis. At the same time, 18 images, 3 text descriptions, 1 set of meteorological records for the past 7 days, and 5 days of agricultural records were submitted. In the state construction stage, the image acquisition time, text entry time, and sensor time were first aligned to form unified plot state information. After analysis, the plot state information was determined as follows: the crop is watermelon, the growth stage is early fruit-setting stage, the abnormalities are mainly concentrated on the upper new leaves and areas with high spray coverage, the symptoms were first discovered on the morning of July 8, the average temperature for the past 3 days was 33.6 degrees Celsius, the average nighttime humidity was 81%, a fungicide and foliar nutrient solution was sprayed at 18:30 on July 7, and drip irrigation was carried out in the early morning of July 8. In this way, the subsequent competitive falsification stage can directly analyze the status information of the same plot without having to read the original data separately again.

[0068] In one embodiment, the competitive falsification phase invokes disease analysis agents, pest analysis agents, nutrient imbalance analysis agents, pesticide damage analysis agents, and abiotic stress analysis agents to analyze the status information of the same plot in parallel. The disease analysis agent, after reading close-up images of leaves, environmental humidity, and the direction of lesion expansion, gives a disease candidate score of 74 points; the pesticide damage analysis agent, combining spraying time, mixing records, symptom concentration areas, and young leaf damage characteristics, gives a pesticide damage candidate score of 78 points; the abiotic stress analysis agent, combining high temperature, drip irrigation, and localized leaf curling phenomena, gives an abiotic stress candidate score of 52 points; the nutrient imbalance analysis agent, combining fertilization records and chlorosis manifestations, gives a nutrient imbalance candidate score of 47 points; the pest analysis agent, due to the absence of obvious insect holes and insect bodies, only gives 18 points. The competitive falsification module further performs conflict analysis. While candidate diseases exhibited high humidity conditions, the lack of images of the undersides of leaves prevented confirmation of the presence of mold or spores. Candidate pesticide damage, though closely related to spraying time, lacked information on pesticide concentration and spray coverage uniformity. Candidate abiotic stress matched high-temperature conditions, but symptoms were concentrated in the sprayed area, not entirely consistent with simple high-temperature burns. This resulted in a set of candidate abnormal causes, in the following order: pesticide damage, disease, abiotic stress, nutrient imbalance, and insect infestation. Three key missing pieces of evidence were extracted: images of the undersides of leaves, records of mixed pesticide concentrations, and a map of the abnormal area distribution. Figure 2 This is a diagram showing the correspondence between symptoms and the timing of agricultural events.

[0069] like Figure 2 As shown, from July 8th to July 13th, the symptom coverage rate first increased and then decreased. Mixed spraying occurred on July 8th, leaf underside re-photographing was completed on July 9th, leaf washing with clean water was performed on July 10th, and localized follow-up inspections were conducted on July 11th. The figure also shows the symptom change nodes, including mild leaf curling, increased patches, worsening at the upper part of the leaf, controlled spread, edge retraction, and color recovery. This figure reflects a clear temporal correspondence between symptom changes and agricultural events. Figure 3 This is a comparison chart of the comprehensive scores of candidate anomaly causes. (Example) Figure 3 As shown, the score for pesticide damage was 78 points, and the score for disease damage was 74 points. The difference between the two scores was only 4 points, which is lower than the preset threshold of 8 points, indicating that the existing evidence is not yet sufficient to make a stable distinction. The scores for abiotic stress, nutritional imbalance and insect pests were significantly lower, and they can be temporarily not used as the primary judgment objects.

[0070] The active probe phase employs a dual triggering condition. The first threshold is set at 8 points, based on the average score difference between the top and second-ranked candidates in a stable judgment state among historical samples of similar watermelon leaf anomalies. The second threshold is set at 2 items, based on the experience-based assessment that when more than two key missing pieces of evidence are present, misjudgment is likely. In the current case, the score difference between the top two candidate anomalies is 4 points, and the number of key missing pieces of evidence is 3, thus triggering a supplementary sampling. The active probe module first uses an image acquisition tool to capture 6 images of the underside of the leaves, then uses an agricultural record query tool to extract the spray tank preparation record, confirming that the mixed pesticide on July 7th contained mancozeb and amino acid foliar nutrients, with a concentration 22% higher than the recommended concentration. Simultaneously, a regional distribution acquisition tool generates an anomaly distribution map, revealing that the anomalies are mainly concentrated in the overlapping area of ​​the spray nozzles and two rows on the south side of the greenhouse. After the supplementary sampling, the plot status information is updated as follows: no obvious mold layer on the underside of the leaves, the pesticide mixture concentration is too high, and the anomaly area is highly consistent with the spray overlap area. The competitive falsification module re-analyzed the plot status information based on the updated information. The score for pesticide damage increased to 91 points, the score for disease decreased to 39 points, the score for abiotic stress decreased to 35 points, the score for nutrient imbalance decreased to 24 points, the score for pests decreased to 9 points, the difference between the first and second place scores widened to 52 points, and the number of key missing evidence items decreased to 0. Figure 4 This is a graph showing the changes in scores for abnormal causes before and after supplementary sampling.

[0071] like Figure 4 As shown, before re-collection, the scores for pesticide damage and disease were similar; after re-collection, the score for pesticide damage increased significantly, while the scores for disease and other candidate abnormal causes decreased significantly, indicating that the re-collection data directly changed the ranking of abnormal causes. Based on this, the active probe module generates treatment constraint information. The set of permissible actions includes washing leaves with clean water, partial shading, 72-hour follow-up inspection, suspending new foliar spraying, and recording the affected area; the set of prohibited actions includes continuing to apply the same type of foliar pesticide, spraying again during the midday high-temperature period, and applying high-concentration foliar nutrients before completing the follow-up inspection. The set of prohibited actions is generated using a union method; if an action corresponding to a candidate abnormal cause poses a significant risk, that action is included in the prohibited range. In this way, the path synthesis stage obtains clear executable boundaries while the abnormal causes have converged.

[0072] The path synthesis phase receives a set of candidate abnormal causes and disposal constraints, and generates an agricultural disposal path by combining the growth stage, agricultural event correlation information, and environmental information. First, a meteorological window is determined based on the weather records for the next 48 hours. On July 10th, the daytime high temperature is 31.2 degrees Celsius, the wind speed is 1.8 meters per second, and the probability of rainfall is 10%, suitable for leaf washing and partial shading; on the morning of July 11th, the temperature is 29.7 degrees Celsius, suitable for reviewing and recording recovery status. Then, three disposal action templates with conditions similar to "watermelon, early fruit setting, pesticide damage, high concentration of mixed pesticides, and damage to upper new leaves" are recalled from the historical case database. After comparison, the template closest to the current environment is selected. The path synthesis module adjusts the template according to the disposal constraints: the "immediately add disease prevention spraying" node is deleted, the "wash leaves with 200 liters of water per acre" node is inserted, the judgment on "whether to restore nutrient supplementation" is postponed to the 4th day, and the "re-shoot abnormal area" node is rearranged after "wash leaves with water". The final agricultural treatment path was as follows: On the morning of the first day, the leaves were washed with clean water and local shading was carried out; on the afternoon of the first day, the boundary of the affected area was recorded; on the second and third days, the upper new leaves were observed twice a day, morning and evening, and photos were taken; on the fourth day, the condition of the new leaves was used to determine whether to resume low-concentration foliar nutrient supplementation; on the fifth day, a second review was completed and the prohibited actions were turned off.

[0073] To verify the effectiveness, an empirical treatment method was used in another adjacent plot with similar conditions. Only a pause in spraying and manual observation were implemented; no active re-sampling or pathway synthesis was performed. After 7 days, the symptom incidence rate in this example plot decreased from 14.1% to 2.1%, the new patch rate decreased from 11.3% to 1.8%, and the leaf curl index decreased from 0.62 to 0.19. In the control plot, the symptom incidence rate decreased from 13.8% to 6.4%, the new patch rate decreased from 10.9% to 5.7%, and the leaf curl index decreased from 0.60 to 0.34. This example plot used water washing and partial shading, with an additional cost of 225 yuan per mu. The control plot, due to a second attempt at remedial spraying, incurred an additional cost of 314 yuan per mu. This example plot reduced the cost by 89 yuan per mu compared to the control plot within 7 days, with an additional 4.3 percentage point decrease in the symptom incidence rate and an additional 3.9 percentage point decrease in the new patch rate.

[0074] Figure 5 This is a graph showing the changes in key indicators after the treatment. For example... Figure 5As shown, the symptom incidence, leaf curl index, and new patch incidence all gradually decreased with treatment time, with the rate of decrease accelerating significantly after the third day. This indicates that the key missing evidence supplemented during the active probe phase effectively improved the accuracy of anomaly cause identification, and the agricultural treatment path generated during the path synthesis phase could converge the anomaly range more quickly without increasing unnecessary inputs. This verifies that this embodiment can generate executable treatment paths based on agricultural multimodal data, competitive falsification results, and active probe supplementation results, and achieve stable anomaly identification and treatment effects in actual fields.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0076] The above are merely embodiments of the present invention and are not intended to limit the 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 principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. An intelligent system for vertical agricultural sectors based on multi-agent collaboration and tool enhancement, characterized in that, The system includes: The status construction module is used to receive agricultural multimodal data and generate plot status information including crop identity, growth stage, symptom association information and agricultural event association information; The competitive falsification module is used to call multiple anomaly analysis agents to perform case matching, mechanism constraints and conflict analysis based on the crop-specific knowledge base corresponding to the crop identity and the state information of the plot, and generate a set of candidate anomaly causes, key missing evidence and anomaly cause ranking results. The active probe module is used to call the data acquisition tool to supplement the data corresponding to the key missing evidence when the score difference between the top two candidate abnormal causes in the abnormal cause ranking results is lower than a preset threshold, update the land status information and generate disposal constraint information. The path synthesis module is used to generate agricultural disposal paths based on the candidate set of abnormal causes and the disposal constraint information, combined with crop safety constraints, meteorological window constraints, historical agricultural conflict constraints and local executability constraints.

2. The system according to claim 1, characterized in that, The agricultural multimodal data includes crop information, symptom information, environmental information, and agricultural information; The state construction module is used to determine the crop identity, the reproductive stage, the symptom association information, and the agricultural event association information based on the agricultural multimodal data.

3. The system according to claim 2, characterized in that, The status construction module is used to generate the symptom association information based on the location of symptom occurrence, symptom morphology, symptom distribution, and symptom onset time; The state construction module is also used to generate the agricultural event association information according to the time sequence of the pesticide application event, fertilizer application event, and irrigation event.

4. The system according to claim 1, characterized in that, The multiple anomaly analysis agents include disease analysis agents, pest analysis agents, nutrient imbalance analysis agents, pesticide damage analysis agents, and abiotic stress analysis agents. The competitive falsification module is used to generate the candidate anomaly cause set and the key missing evidence based on the analysis results of each anomaly analysis agent.

5. The system according to claim 4, characterized in that, The competitive falsification module is used to determine the score of each candidate abnormal cause based on the case matching results, mechanism constraint results, temporal consistency between the symptom occurrence time and the agricultural event time, consistency of agricultural event type, and conflict analysis results, and to generate the abnormal cause ranking result based on the score of each candidate abnormal cause.

6. The system according to claim 5, characterized in that, The active probe module is used to call the data acquisition tool to supplement the data corresponding to the key missing evidence when the score difference between the top two candidate anomalies in the anomaly cause ranking results is lower than a first threshold and the number of key missing evidences is greater than a second threshold, and to update the land parcel status information based on the supplemented data.

7. The system according to claim 6, characterized in that, The active probe module is used to regenerate the candidate anomaly cause set based on the updated land parcel status information, and generate the disposal constraint information based on the updated candidate anomaly cause set; The handling constraint information includes a set of allowed actions and a set of prohibited actions, wherein the set of prohibited actions is the union of the prohibited actions corresponding to each candidate abnormal cause in the set of candidate abnormal causes.

8. The system according to claim 2, characterized in that, The path synthesis module is used to determine a meteorological window that meets the meteorological window constraints based on the candidate abnormal cause set, the reproductive stage, the agricultural event association information, and the environmental information, and to recall the handling action template from the historical case database.

9. The system according to claim 8, characterized in that, The path synthesis module is used to adjust the disposal action template according to the disposal constraint information. The adjustment includes node insertion, node deletion, node postponement, and node rearrangement to generate the agricultural disposal path that satisfies the crop safety constraint, the weather window constraint, the historical agricultural conflict constraint, and the local executability constraint.

10. The system according to claim 9, characterized in that, The agricultural response path includes immediate response actions, observation actions, delayed response actions, and prohibited actions; The local executability constraints are used to characterize the availability of agricultural inputs, operating equipment conditions, and planting scale.

Images