An artificial intelligence-based crop growth monitoring system

Abstract

The application relates to the technical field of artificial intelligence, and provides a crop growth monitoring system based on artificial intelligence, which comprises the following steps: establishing a unique identifier for each plot, recording and maintaining plot boundary, crop type, sowing date and historical yield information, and organizing adjacent plots into a continuous plot set; based on a pre-trained growth identification basic model, performing growth grade identification on collected plot images, and performing lightweight parameter correction under the constraint of local samples to obtain the growth index of each plot; splicing the growth index of each plot into a growth trajectory in chronological order, calculating a risk index representing the severity of potential problems and an uncertainty index representing the degree of insufficient monitoring; according to the risk index and the uncertainty index, combining the operation time window, the traffic condition and the operation capacity, performing priority sorting and generating a patrol route queue; and outputting a scheduling suggestion for a target plot according to the route queue and the growth trajectory.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically an artificial intelligence-based crop growth monitoring system. Background Technology

[0002] Crop growth monitoring refers to the assessment of crop growth status and trends using remote sensing technology or traditional methods. The purpose is to provide information for field management and to make early yield estimates. It can provide a macroscopic understanding of crop growth and analyze differences in crop growth by comparing data from different time periods.

[0003] The Chinese patent application (CN202210201040.7) discloses a method, system, equipment, and medium for monitoring crop growth. The configuration method includes: acquiring the normalized vegetation index (NVI) of at least two sub-regions of crops in the current region; determining the crop growth status based on the acquisition time of the NVI, where the growth status includes at least the sowing period, growing period, and maturity period; comparing the NVI of each sub-region with at least one reference range included in the crop's growth status, and determining and counting the number of corresponding sub-regions within each reference range; and providing tiered early warnings based on the number of counted sub-regions. The method compares multiple NVIs of crops in the current region with preset reference ranges, and monitors crop growth based on the number of corresponding sub-regions within each reference range. It visualizes alarm information, making it easier for users to monitor crop growth status.

[0004] In the field of artificial intelligence technology, although there are technical solutions that compare multiple normalized vegetation indices of crops in the current region with a preset reference range and monitor the growth of crops based on the number of corresponding sub-regions in the reference range, in the current scenario of small and scattered specialty economic crops, many small farmers in various regions choose to grow specialty crops such as grapes, blueberries, coffee, and tea. Due to the scattered plots and small area, and the fact that specialty crops are relatively sensitive to growth and quality, most small farmers generally lack funds and technology, making it difficult for them to use complex equipment and platforms. There is a lack of a solution that can obtain growth monitoring results with a simple terminal.

[0005] To address the aforementioned issues, this invention proposes an artificial intelligence-based crop growth monitoring system. This system establishes a unique identifier for each plot of land and manages its boundaries and crop information. It utilizes a pre-trained model to identify and correct growth indices, constructs growth trajectories over time to extract risk and uncertainty indicators, and optimizes inspection routes based on operational conditions to output harvesting schedule and human-machine scheduling suggestions. Summary of the Invention

[0006] In view of the existing problems mentioned above, an artificial intelligence-based crop growth monitoring system is proposed.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is: an artificial intelligence-based crop growth monitoring system, comprising:

[0008] The land parcel information management module establishes a unique identifier for each land parcel, records and maintains information on land parcel boundaries, crop types, sowing dates and historical yields, and organizes adjacent land parcels into a continuous set of land parcels;

[0009] The growth recognition and correction module, based on a pre-trained growth recognition model, identifies the growth level of the collected plot images and performs lightweight parameter correction on the basic model under the constraint of local samples to obtain the growth index of each plot.

[0010] A growth trajectory module is constructed, which splices the growth indices of each plot in chronological order to form a growth trajectory, calculates risk indicators that characterize the severity of potential problems, and uncertainty indicators that characterize the degree of inadequate monitoring.

[0011] The inspection route decision module prioritizes the set of plots and generates an inspection route queue based on risk indicators and uncertainty indicators, combined with the operation time window, traffic conditions and operation capabilities.

[0012] The management suggestion output module outputs suggestions on harvesting rhythm, field labor and agricultural machinery scheduling for target plots based on route queues and growth trajectories.

[0013] As a preferred implementation, the specific steps of the land parcel information management module are as follows: First, import the base map data of the contracted land into the land parcel information management module. On the visual plotting interface, select and draw polygonal land parcels along the boundary points. Automatically snap to the existing boundary lines through topological correction, and call the functions of the spatial database to calculate the area and overlap. Generate a unique land parcel ID for each land parcel according to the rules of administrative region code, year, and serial number. Write the ID along with the geometric boundary into the land parcel basic table, and record the land parcel production attribute information in the attribute table.

[0014] The system pre-determines adjacent plots by setting conditions. When the shared boundary length of two plots is greater than the pre-defined ratio of their perimeter, and the crop type and sowing date are the same, it automatically generates an adjacency matrix using spatial overlay analysis. Based on this matrix, it performs connected component clustering to organize several adjacent plots into a continuous set of plots, assigns a set ID, and establishes a mapping table. When it is necessary to adjust the plot boundaries and attributes in the future, it generates a new version record while keeping the original plot ID unchanged. It incrementally recalculates the adjacency relationship and set division for the affected area and records the effective time interval.

[0015] As a preferred embodiment, the specific steps of the growth recognition and correction module are as follows:

[0016] Retrieve orthophotos of the corresponding time from the image service based on the unique plot ID, crop the plot boundary to obtain the plot canopy image, and preprocess it. Input the processed image into the pre-trained offline growth recognition basic model, maintain a small sample annotation library locally, and only unfreeze the end classification layer of the basic model to update the parameters when performing model correction. Keep the rest of the backbone network frozen and perform rapid calibration on local samples through gradient descent to obtain a corrected model consistent with the scale of the local area. Use the corrected model to re-infer all plot images, calculate the continuous growth index according to the output probability weight, and write it back to the plot information database.

[0017] The pre-trained offline growth recognition model consists of: a convolutional neural network backbone layer for extracting image features, a feature concatenation layer for fusing auxiliary features, and a multi-task output layer that outputs growth level probabilities and continuous growth scores respectively. During training, satellite imagery, UAV imagery, and close-up ground images are collected from multiple regions and years. Each image is cropped into a canopy sub-image according to plot boundaries and mapped one-to-one with the corresponding plot ID, crop type, growth stage, and measured growth indicators and growth score labels, constructing a training sample set that combines image and auxiliary feature mapping with growth labels. In the forward recognition process, the plot canopy image is processed by the convolutional backbone layer to extract multi-scale feature maps, which are then concatenated with the encoded auxiliary features and input into the classification head and regression head. The classification head outputs the probability distribution of each growth level, and the regression head outputs the continuous growth prediction value. Simultaneously, the one-hot vector of the actual labeled level and the measured growth score are used as supervision signals. The loss function is a weighted sum of classification loss and regression loss, with the following formula:

[0018] ,

[0019] in Represents the classification loss weights. This represents the regression loss weight, and i represents the index of the growth level. Let represent the probability of the i-th growth level. This represents the true label for the i-th growth level. This represents the predicted growth score output by the model's regression head. This represents the true continuous growth score;

[0020] A continuous growth index is calculated by weighting the output probabilities. The lightly modified growth recognition model is then used for batch inference for each plot of land based on the shooting date. The model outputs a probability vector of discrete growth levels for the input image. First, the probability vector is calibrated with confidence, and each level is mapped to a preset numerical score. The initial continuous growth score is obtained by summing the probabilities.

[0021] ,

[0022] Where i represents the index of the growth level, and N represents the total number of growth levels. Let represent the probability of the i-th growth level. This represents the numerical score corresponding to the i-th level;

[0023] Weighted temporal smoothing is used to improve stability and temporal continuity, fusing data from several time points. The smoothed continuous growth index is obtained. .

[0024] As a preferred embodiment, the specific steps for constructing the growth trajectory module are as follows:

[0025] Based on the plot ID, all continuous growth index records for the plot are read from the database in ascending order of date. For cases with uneven shooting intervals or missing data for individual dates, the blanks are filled in using a time-series backfilling strategy based on adjacent dates. The filled growth index sequence is then spliced ​​together over time to generate the plot growth trajectory, which is stored in the form of a time-series object.

[0026] After constructing the trajectory, multidimensional analysis is performed to extract risk indicators reflecting the severity of the problem. The first and second derivatives of the growth index curve are calculated to determine the recent rate and acceleration of growth decline, used to identify whether the growth abnormality is worsening. Combined with the crop growth period standard curve, alignment and comparison are performed at the same age to calculate the deviation of the current trajectory from the standard curve. Based on the magnitude and duration of the deviation, a persistently low index is calculated. The formula for the persistently low index is:

[0027] ,

[0028] in This indicates a risk indicator for trajectory deviation, where T represents the total number of time points and t represents the time index. This represents the continuous growth index corresponding to time t. This represents the standard growth index corresponding to time t. This represents the upper limit of the growth index scale.

[0029] The sliding window method is used to calculate the curve volatility. By utilizing the neighborhood correlation between plots, the common part of the decline in growth of adjacent plots during the same period is extracted to identify regional risks. These regional risks are then used as weighting factors to correct the risk scores of individual plots.

[0030] While calculating risk indicators, uncertainty indicators representing the degree of insufficient monitoring are also calculated. The model uncertainty is calculated based on the probability output of the growth recognition module. The proportion and distribution of data from different image sources in the trajectory are statistically analyzed. The length of missing data segments in the trajectory is calculated, and based on the impact of missing data on trend interpretation, it is mapped to a monitoring gap indicator. The difference before and after smoothing in the trajectory is used as a curve stability indicator. The formula for curve stability uncertainty is:

[0031] ,

[0032] in Indicators representing the uncertainty of curve stability This represents the initial growth index at time t. This represents the smoothed growth exponent at time t;

[0033] Risk indicators and uncertainty indicators are written together into the growth trajectory data structure, and a unified plot trajectory health report is generated.

[0034] As a preferred embodiment, the specific steps of the inspection route decision module are as follows:

[0035] Based on consecutive land parcel sets, the comprehensive risk index of each set is... With comprehensive uncertainty indicators Normalization is performed, and an inspection necessity score is calculated according to preset weights. Additionally, based on crop-related information, plots near key agricultural milestones are given extra weight, forming an inspection priority list for the current date. The formula for calculating the inspection necessity score is as follows:

[0036] ,

[0037] Where j represents the index of the set of land parcels, Indicates risk weight. Indicates the uncertainty weight. Indicates the weight of key agricultural events. This represents the risk index of the j-th consecutive set of land parcels. This represents the uncertainty index of the j-th consecutive set of land parcels. The weighting factor represents the neighboring key agricultural nodes of the j-th contiguous set of land parcels;

[0038] The geometric center points of each continuous set of land parcels are read from the land parcel information management module. Combined with the pre-imported farm roads, farm roads and obstacle layers, a road network map is constructed with the work points as nodes and the actual travel time as edge weights. Then, based on the available inspection force for the day and the maximum working time and maximum travel distance of each type of work unit, corresponding work capacity constraints are established for each type of work unit. In the set of land parcels with the highest inspection necessity, a heuristic path generation algorithm is used to connect several sets of land parcels into an initial inspection route according to the road network distance, under the condition that the single shift work time does not exceed the limit. The determined inspection route is output as an inspection route queue.

[0039] As a preferred implementation, the specific steps of the management suggestion output module are as follows:

[0040] For each set of consecutive plots, based on the deviation of the growth trajectory from the standard curve of the growth period, the maturity score and the expected time window to reach the suitable harvest period are calculated. Plots with a maturity score higher than the preset threshold and that have entered the suitable harvest period are marked as priority harvest plots, and plots with growth weaker than the average growth level but not yet at the harvest period are marked as key tracking plots. The corresponding inspection routes and expected arrival times are recorded.

[0041] The system iterates through the inspection route queue daily, estimating the time window for agricultural operations that can be performed concurrently for each route. Combining this with the company's labor and machinery resource ledger, it reverse-engineers the manpower and machinery requirements by day and by route. For each route, it calculates the working hours and machine hours required to complete the recommended harvesting tasks. The results are matched with the available resources for the day, providing the number of team members, recommended machine types and quantities, and suggested shift arrangements for each route. When resources are insufficient to cover all priority harvesting tasks, the module marks some plots as delayed batches based on maturity scores, risk indicators, and the weight of nearby key agricultural nodes, and provides a delay period and potential yield risk warnings in the management suggestions.

[0042] Beneficial effects

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] 1. By establishing a long-term, stable, and unique identifier for each plot of land and automatically organizing a set of consecutive plots based on topological relationships, and by adopting versioned management when adjusting plot boundaries, the system achieves unified spatial and temporal association of growth trajectory, yield data, and inspection records, thus solving the problems of chaotic plot numbering and difficulty in tracing across years in existing technologies.

[0045] 2. By combining comprehensive risk indicators and uncertainty indicators into an inspection necessity score, and combining operation time windows, road conditions and operation capabilities, a heuristic path generation algorithm is used to construct inspection routes at the level of continuous plot sets. This allows limited inspection resources to prioritize covering high-risk and unreliable plots, rather than simply patrolling by distance or in fixed shifts, thereby improving inspection efficiency and problem detection rate. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0047] Figure 1 This is a block diagram of the present invention;

[0048] Figure 2 This is a comparison diagram of the effects of the present invention and the prior art, where gray bars represent the prior art and black bars represent the present invention. Detailed Implementation

[0049] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0050] Example 1:

[0051] To achieve the above objectives, please refer to Figure 1 This invention provides an artificial intelligence-based crop growth monitoring system, which includes the following modules:

[0052] The land parcel information management module establishes a unique identifier for each land parcel, records and maintains information on land parcel boundaries, crop types, sowing dates and historical yields, and organizes adjacent land parcels into a continuous set of land parcels;

[0053] The growth recognition and correction module, based on a pre-trained growth recognition model, identifies the growth level of the collected plot images and performs lightweight parameter correction on the basic model under the constraint of local samples to obtain the growth index of each plot.

[0054] A growth trajectory module is constructed, which splices the growth indices of each plot in chronological order to form a growth trajectory, calculates risk indicators that characterize the severity of potential problems, and uncertainty indicators that characterize the degree of inadequate monitoring.

[0055] The inspection route decision module prioritizes the set of plots and generates an inspection route queue based on risk indicators and uncertainty indicators, combined with the operation time window, traffic conditions and operation capabilities.

[0056] The management suggestion output module outputs suggestions on harvesting rhythm, field labor and agricultural machinery scheduling for target plots based on route queues and growth trajectories.

[0057] Example 2:

[0058] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the land parcel information management module:

[0059] First, import the vector data and remote sensing orthophotos of the contracted land into the land information management module. On the visual plotting interface, select and draw polygonal plots along the boundaries of field ridges, farm roads, and irrigation ditches. The plots are automatically snapped to the existing boundaries through topological correction, and the area and overlap are calculated by calling the functions of the spatial database. A unique plot ID is generated for each plot according to the rules of administrative region code, year, and serial number. The ID, along with the geometric boundary, is written into the plot basic table, and the crop type, variety, sowing date, crop rotation method, soil type, and annual yield are recorded in the attribute table.

[0060] To reflect the spatial relationship between plots, the module presets the conditions for determining adjacent plots. When the proportion of the shared boundary length to the perimeter of two plots is greater than a preset value, and the crop type and sowing date are the same, an adjacency matrix is ​​automatically generated using spatial overlay analysis. Based on this, connected component clustering is performed to organize several adjacent plots into a continuous plot set, assigning a set ID and establishing a mapping table. When it is necessary to adjust the plot boundaries and attributes in the future, a new version record is generated while keeping the original plot ID unchanged. The adjacency relationship and set division of the affected area are incrementally recalculated, and the effective time interval is recorded.

[0061] Unlike existing technologies that manage land parcel information only according to a fixed grid, this invention links adjacent land parcels into a set, thereby enabling traceable management of the land parcel evolution process and providing a stable, consistent, and detailed spatial index for crop growth trajectory construction and inspection route planning down to the individual land parcel level.

[0062] Example 3:

[0063] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the growth identification and correction module:

[0064] The orthophoto of the corresponding time is retrieved from the image service according to the unique plot ID. The plot canopy image is cropped according to the plot boundary and preprocessed such as resolution unification and brightness normalization. The image service consists of multi-source image data pre-accessed by the system. The processed image is input into the pre-trained offline growth recognition basic model. The growth recognition basic model is trained based on historical samples of multiple regions and crops and can output the initial growth level and probability vector of each plot. In order to adapt to the differences in varieties and management modes in this region, a small sample annotation library is maintained locally. This library includes typical growth samples and their true levels or key growth indicators verified by agricultural technicians. When performing model correction, only the end classification layer of the basic model is unfrozen and a small number of iterative parameter updates are performed. The rest of the backbone network is kept frozen. Gradient descent is used to quickly calibrate on local samples to obtain a corrected model consistent with the scale of this region. The corrected model is used to re-infer all plot images. The continuous growth index is calculated by weighting the output probability and written back to the plot information database.

[0065] Specifically, the pre-trained offline growth recognition model comprises: a convolutional neural network backbone layer for extracting image features; a feature stitching layer for fusing auxiliary features such as crop type, growth stage, and shooting season; and a multi-task output layer that outputs growth level probabilities and continuous growth scores, respectively. During training, satellite images, UAV images, and close-up ground images are first collected from multiple regions and years. Each image is cropped into a canopy sub-image according to the plot boundaries and compared with the corresponding plot ID, crop type, growth stage, and measured growth rate. The growth potential index and growth score labels are mapped one-to-one, constructing a training sample set that combines image and auxiliary features with the growth potential label mapping. During the forward recognition process, the canopy image of the plot is processed through a convolutional backbone layer to extract multi-scale feature maps, which are then concatenated with coded auxiliary features such as crop type and growth stage. These are input into the classification head and regression head. The classification head outputs the probability distribution of each growth potential level, and the regression head outputs continuous growth potential prediction values. Simultaneously, the one-hot vector of the actual labeled level and the measured growth potential score are used as supervision signals. The loss function is a weighted sum of classification loss and regression loss, with the following formula:

[0066] ,

[0067] in Represents the classification loss weights. This represents the regression loss weight, and i represents the index of the growth level. Let represent the probability of the i-th growth level. This represents the true label for the i-th growth level. This represents the predicted growth score output by the model's regression head. This represents the true continuous growth score;

[0068] Specifically, the modified model is used to re-infer all plot images, and a continuous growth index is calculated by weighting the output probabilities. The lightly modified growth recognition model is then used to batch infer for each plot according to the shooting date. The model outputs a probability vector of discrete growth levels for the input image. First, the probability vector is calibrated with confidence, and each level is mapped to a preset numerical score. The initial continuous growth score is obtained by summing the probabilities.

[0069] ,

[0070] Where i represents the index of the growth level, and N represents the total number of growth levels. Let represent the probability of the i-th growth level. This represents the numerical score corresponding to the i-th level;

[0071] Weighted time series smoothing is employed to improve stability and temporal continuity, such as exponentially weighted moving average, which integrates data from several time points. The smoothed continuous growth index is obtained. .

[0072] Unlike existing technologies that directly use fixed models, this module achieves adaptive calibration to local environmental and varietal differences through lightweight parameter adjustments without large-scale retraining, balancing versatility and local accuracy.

[0073] Example 4:

[0074] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps for constructing the growth trajectory module:

[0075] Based on the plot ID, all continuous growth index records for that plot are read from the database in ascending order of date. For cases with uneven shooting intervals or missing data for individual dates, a time-series backfilling strategy based on adjacent dates is used to fill in the gaps. Specifically, the time-series backfilling strategy automatically generates reasonable estimates based on valid data from consecutive dates by using neighboring values ​​to ensure the continuity of the trajectory on the time axis. The supplemented growth index sequence is then spliced ​​together over time to generate the plot's growth trajectory, which is stored as a time series object. Each trajectory point contains information such as the growth index, model uncertainty, image timestamp, and image source.

[0076] After constructing the trajectory, multidimensional analysis is performed to extract risk indicators reflecting the severity of the problem. The first and second derivatives of the growth index curve are calculated to determine the recent rate and acceleration of growth decline, used to identify whether the growth abnormality is worsening. Combined with the crop growth period standard curve, alignment and comparison are performed at the same age to calculate the deviation of the current trajectory from the standard curve. Based on the magnitude and duration of the deviation, a persistently low index is calculated. The formula for the persistently low index is:

[0077] ,

[0078] in This indicates a risk indicator for trajectory deviation, where T represents the total number of time points and t represents the time index. This represents the continuous growth index corresponding to time t. This represents the standard growth index corresponding to time t. This represents the upper limit of the growth index scale.

[0079] The curve volatility is calculated using a sliding window. By utilizing the neighborhood correlation between plots, the common part of the decline in growth of adjacent plots during the same period is extracted, regional risks such as low temperature and continuous rain are identified, and the regional risk is used as a weighting factor to correct the risk score of individual plots.

[0080] While calculating risk indicators, uncertainty indicators representing the degree of insufficient monitoring are also calculated. The model uncertainty is calculated based on the probability output of the growth recognition module. The proportion and distribution of data from different image sources in the trajectory are statistically analyzed. The length of missing data segments in the trajectory is calculated, and based on the impact of missing data on trend interpretation, it is mapped to a monitoring gap indicator. The difference before and after smoothing in the trajectory is used as a curve stability indicator. The formula for curve stability uncertainty is:

[0081] ,

[0082] in Indicators representing the uncertainty of curve stability This represents the initial growth index at time t. This represents the smoothed growth exponent at time t;

[0083] Risk indicators and uncertainty indicators are written together into the growth trajectory data structure, and a unified plot trajectory health report is generated, which includes: original growth trajectory, risk level label, uncertainty level, explanation of risk source, and time coverage of trajectory.

[0084] Compared to existing technologies that only make point-like judgments on the growth status at a single moment, this invention achieves continuous characterization of the growth status of the plot through time-series splicing, trend analysis, and uncertainty assessment. It can also clearly indicate the degree of risk and the credibility of monitoring, thus providing more accurate and interpretable input for subsequent inspection route planning. This constitutes the significant technical difference of this invention in the trajectory construction stage.

[0085] Example 5:

[0086] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the inspection route decision module:

[0087] Based on consecutive land parcel sets, the comprehensive risk index of each set is... With comprehensive uncertainty indicators Normalization is performed, and an inspection necessity score is calculated according to preset weights. Additionally, based on supplementary information such as the crop's current growth stage and planned harvest time, plots near key agricultural milestones (e.g., sowing, harvesting) are given extra weight, forming an inspection priority list for the current date. The formula for calculating the inspection necessity score is:

[0088] ,

[0089] Where j represents the index of the set of land parcels, Indicates risk weight. Indicates the uncertainty weight. Indicates the weight of key agricultural events. This represents the risk index of the j-th consecutive set of land parcels. This represents the uncertainty index of the j-th consecutive set of land parcels. The weighting factor represents the neighboring key agricultural nodes of the j-th contiguous set of land parcels;

[0090] The geometric center points of each continuous set of plots are read from the plot information management module. Combined with the pre-imported farm roads, farm roads and obstacle layers, a road network map is constructed with the work point as the node and the actual travel time as the edge weight. Then, based on the available inspection force for the day and parameters such as the maximum working time and maximum travel distance of each type of work unit, corresponding work capacity constraints are established for each type of work unit. In the set of plots with the highest inspection necessity, a heuristic path generation algorithm is used to connect several sets of plots into an initial inspection route according to the road network distance, under the condition that the single shift work time does not exceed the limit. Specifically, given the road network and the set of plots to be inspected, the heuristic path generation algorithm prioritizes starting from plots with high inspection necessity and spatial proximity, and iteratively constructs and optimizes the travel order according to the strategy of visiting nearby and making local adjustments, so as to obtain an approximate optimal route with a shorter journey and higher coverage value without exhausting all combinations.

[0091] The determined inspection routes are output as an inspection route queue. Each route contains the set ID of the plots arranged in the order of visit, the estimated arrival time, the estimated departure time, and the estimated driving distance, and is labeled with the corresponding work unit type.

[0092] Unlike existing technologies that generate patrol routes based solely on fixed grids, this embodiment introduces a patrol necessity degree with added risk and uncertainty, a spatial organization unit of continuous plot sets, and constraints on operational capacity and time windows during path planning. This significantly improves patrol effectiveness and scheduling efficiency with limited patrol resources, forming a technical feature that is distinctly different from traditional single shortest distance or simple rotation patrol strategies.

[0093] Example 6:

[0094] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the management suggestion output module:

[0095] For each set of consecutive plots, based on the deviation of the growth trajectory from the standard curve of the growth period, the maturity score and the expected time window to reach the suitable harvest period are calculated. Plots with high maturity scores and that have entered the suitable harvest period are marked as priority harvest plots, and plots with weak growth but not yet at the harvest period are marked as key tracking plots. The corresponding inspection routes and expected arrival times are recorded.

[0096] The inspection route queue is traversed daily, and the time window for agricultural operations that can be performed along each route is estimated. For example, if a route passes through multiple plots with high maturity scores in a day, the upper limit of the harvestable area that can be completed that day is calculated based on the area of ​​each plot, the expected work efficiency per acre, and the type of machinery available, and a list of priority harvesting plots and the corresponding batch harvesting order is generated.

[0097] Based on the enterprise's labor and agricultural machinery resource ledger, the demand for manpower and machinery is reversed by day and by route. For each route, the working hours and machine hours required to complete the recommended harvesting task are calculated. The results are matched with the available resources for the day, and the number of team members, recommended machine types and numbers, and suggested shift arrangements are given for each route. When the resources are insufficient to cover all priority harvesting tasks, the module marks some plots as delayed batches based on maturity scores, risk indicators, and the weight of nearby key agricultural nodes, and provides a delay period range and potential yield risk warning in the management suggestions.

[0098] The above results are summarized to generate structured management recommendation reports and visualization interfaces, including: harvesting schedules listed by date, labor and agricultural machinery scheduling plans divided by route and work group, and a must-see list and intensive monitoring recommendations for high-risk or high-uncertainty plots.

[0099] Unlike existing technologies that only output simple alarms, this embodiment is based on the comprehensive constraints of growth trajectory, risk and uncertainty assessment, and actual inspection routes and operational capabilities. It provides information on where to look, what to do first, who to send, and what machines to turn on, so that the monitoring results can be directly converted into an executable daily production plan, forming a significant technical feature of a closed loop from intelligent monitoring to production scheduling.

[0100] like Figure 2 This is a comparison chart of the effects of an artificial intelligence-based crop growth monitoring system. The horizontal axis lists key performance indicators, and the vertical axis represents exemplified performance scores ranging from 0 to 100. Higher values ​​indicate better performance. The aim is to intuitively demonstrate the expected improvement of the present invention in key capabilities compared to typical prior art.

[0101] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. An artificial intelligence-based crop growth monitoring system, characterized in that, include: The land parcel information management module establishes a unique identifier for each land parcel, records and maintains information on land parcel boundaries, crop types, sowing dates and historical yields, and organizes adjacent land parcels into a continuous set of land parcels; The growth recognition and correction module, based on a pre-trained growth recognition model, identifies the growth level of the collected plot images and performs lightweight parameter correction on the basic model under the constraint of local samples to obtain the growth index of each plot. A growth trajectory module is constructed, which splices the growth indices of each plot in chronological order to form a growth trajectory, calculates risk indicators that characterize the severity of potential problems, and uncertainty indicators that characterize the degree of inadequate monitoring. The inspection route decision module prioritizes the set of plots and generates an inspection route queue based on risk indicators and uncertainty indicators, combined with the operation time window, traffic conditions and operation capabilities. The management suggestion output module outputs suggestions on harvesting rhythm, field labor and agricultural machinery scheduling for target plots based on route queues and growth trajectories.

2. The crop growth monitoring system based on artificial intelligence according to claim 1, characterized in that: The specific steps of the land parcel information management module are as follows: First, import the base map data of the contracted land into the land parcel information management module. On the visual plotting interface, select and draw polygonal land parcels along the boundary points. Through topological correction, automatically snap to the existing boundary lines and call the functions of the spatial database to calculate the area and overlap. Generate a unique land parcel ID for each land parcel according to the rules of administrative region code, year, and serial number. Write the ID along with the geometric boundary into the land parcel base table and record the land parcel production attribute information in the attribute table. The system pre-determines adjacent plots by setting conditions. When the shared boundary length of two plots is greater than the pre-defined ratio of their perimeter, and the crop type and sowing date are the same, it automatically generates an adjacency matrix using spatial overlay analysis. Based on this matrix, it performs connected component clustering to organize several adjacent plots into a continuous set of plots, assigns a set ID, and establishes a mapping table. When it is necessary to adjust the plot boundaries and attributes in the future, it generates a new version record while keeping the original plot ID unchanged. It incrementally recalculates the adjacency relationship and set division for the affected area and records the effective time interval.

3. The crop growth monitoring system based on artificial intelligence according to claim 1, characterized in that: The specific steps of the growth identification and correction module are as follows: Retrieve orthophotos of the corresponding time from the image service based on the unique plot ID, crop the plot boundaries to obtain the plot canopy image, and preprocess it. Input the processed image into the pre-trained offline growth recognition basic model, maintain a small sample annotation library locally, and only unfreeze the end classification layer of the basic model to update the parameters when performing model correction. Keep the rest of the backbone network frozen and perform rapid calibration on local samples through gradient descent to obtain a corrected model consistent with the scale of the local area. Use the corrected model to re-infer all plot images, calculate the continuous growth index according to the output probability weight, and write it back to the plot information database.

4. The crop growth monitoring system based on artificial intelligence according to claim 3, characterized in that: The specific steps of the growth identification and correction module also include: The pre-trained offline growth recognition model consists of: a convolutional neural network backbone layer for extracting image features, a feature concatenation layer for fusing auxiliary features, and a multi-task output layer that outputs growth level probabilities and continuous growth scores respectively. During training, satellite imagery, UAV imagery, and close-up ground images are collected from multiple regions and years. Each image is cropped into a canopy sub-image according to plot boundaries and mapped one-to-one with the corresponding plot ID, crop type, growth stage, and measured growth indicators and growth score labels, constructing a training sample set that combines image and auxiliary feature mapping with growth labels. In the forward recognition process, the plot canopy image is processed by the convolutional backbone layer to extract multi-scale feature maps, which are then concatenated with the encoded auxiliary features and input into the classification head and regression head. The classification head outputs the probability distribution of each growth level, and the regression head outputs the continuous growth prediction value. Simultaneously, the one-hot vector of the actual labeled level and the measured growth score are used as supervision signals. The loss function is a weighted sum of classification loss and regression loss, with the following formula: ， in Represents the classification loss weights. This represents the regression loss weight, and i represents the index of the growth level. Let represent the probability of the i-th growth level. This represents the true label for the i-th growth level. This represents the predicted growth score output by the model's regression head. This represents the true continuous growth score.

5. The crop growth monitoring system based on artificial intelligence according to claim 3, characterized in that: The specific steps of the growth identification and correction module also include: A continuous growth index is calculated by weighting the output probabilities. The lightly modified growth recognition model is then used for batch inference for each plot of land based on the shooting date. The model outputs a probability vector of discrete growth levels for the input image. First, the probability vector is calibrated with confidence, and each level is mapped to a preset numerical score. The initial continuous growth score is obtained by summing the probabilities. ， Where i represents the index of the growth level, and N represents the total number of growth levels. Let represent the probability of the i-th growth level. This represents the numerical score corresponding to the i-th level; Weighted temporal smoothing is used to improve stability and temporal continuity, fusing data from several time points. The smoothed continuous growth index is obtained. .

6. The crop growth monitoring system based on artificial intelligence according to claim 1, characterized in that: The specific steps for constructing the growth trajectory module are as follows: Based on the plot ID, all continuous growth index records for the plot are read from the database in ascending order of date. For cases with uneven shooting intervals or missing data for individual dates, the blanks are filled in using a time-series backfilling strategy based on adjacent dates. The filled growth index sequence is then spliced ​​together over time to generate the plot growth trajectory, which is stored in the form of a time-series object. After constructing the trajectory, multidimensional analysis is performed to extract risk indicators reflecting the severity of the problem. The first and second derivatives of the growth index curve are calculated to determine the recent rate and acceleration of growth decline, used to identify whether the growth abnormality is worsening. Combined with the crop growth period standard curve, alignment and comparison are performed at the same age to calculate the deviation of the current trajectory from the standard curve. Based on the magnitude and duration of the deviation, a persistently low index is calculated. The formula for the persistently low index is: ， in This indicates a risk indicator for trajectory deviation, where T represents the total number of time points and t represents the time index. This represents the continuous growth index corresponding to time t. This represents the standard growth index corresponding to time t. This represents the upper limit of the growth index scale.

7. The crop growth monitoring system based on artificial intelligence according to claim 6, characterized in that: The specific steps for constructing the growth trajectory module also include: The sliding window method is used to calculate the curve volatility. By utilizing the neighborhood correlation between plots, the common part of the decline in growth of adjacent plots during the same period is extracted to identify regional risks. These regional risks are then used as weighting factors to correct the risk scores of individual plots. While calculating risk indicators, uncertainty indicators representing the degree of insufficient monitoring are also calculated. The model uncertainty is calculated based on the probability output of the growth recognition module. The proportion and distribution of data from different image sources in the trajectory are statistically analyzed. The length of missing data segments in the trajectory is calculated, and based on the impact of missing data on trend interpretation, it is mapped to a monitoring gap indicator. The difference before and after smoothing in the trajectory is used as a curve stability indicator. The formula for curve stability uncertainty is: ， in Indicators representing the uncertainty of curve stability This represents the initial growth index at time t. This represents the smoothed growth exponent at time t; Risk indicators and uncertainty indicators are written together into the growth trajectory data structure, and a unified plot trajectory health report is generated.

8. The crop growth monitoring system based on artificial intelligence according to claim 1, characterized in that: The specific steps of the inspection route decision module are as follows: Based on consecutive land parcel sets, the comprehensive risk index of each set is... With comprehensive uncertainty indicators Normalization is performed, and an inspection necessity score is calculated according to preset weights. Additionally, based on crop-related information, plots near key agricultural milestones are given extra weight, forming an inspection priority list for the current date. The formula for calculating the inspection necessity score is as follows: ， Where j represents the index of the set of land parcels, Indicates risk weight. Indicates the uncertainty weight. Indicates the weight of key agricultural events. This represents the risk index of the j-th consecutive set of land parcels. This represents the uncertainty index of the j-th consecutive set of land parcels. It represents the weighting factor of the nearest key agricultural nodes of the j-th consecutive set of land parcels.

9. The crop growth monitoring system based on artificial intelligence according to claim 8, characterized in that: The specific steps of the inspection route decision module also include: The geometric center points of each continuous set of land parcels are read from the land parcel information management module. Combined with the pre-imported farm roads, farm roads and obstacle layers, a road network map is constructed with the work points as nodes and the actual travel time as edge weights. Then, based on the available inspection force for the day and the maximum working time and maximum travel distance of each type of work unit, corresponding work capacity constraints are established for each type of work unit. In the set of land parcels with the highest inspection necessity, a heuristic path generation algorithm is used to connect several sets of land parcels into an initial inspection route according to the road network distance, under the condition that the single shift work time does not exceed the limit. The determined inspection route is output as an inspection route queue.

10. The crop growth monitoring system based on artificial intelligence according to claim 1, characterized in that: The specific steps of the management suggestion output module are as follows: For each set of consecutive plots, based on the deviation of the growth trajectory from the standard curve of the growth period, the maturity score and the expected time window to reach the suitable harvest period are calculated. Plots with a maturity score higher than the preset threshold and that have entered the suitable harvest period are marked as priority harvest plots, and plots with growth weaker than the average growth level but not yet at the harvest period are marked as key tracking plots. The corresponding inspection routes and expected arrival times are recorded. The system iterates through the inspection route queue daily, estimating the time window for agricultural operations that can be performed concurrently for each route. Combining this with the company's labor and machinery resource ledger, it reverse-engineers the manpower and machinery requirements by day and by route. For each route, it calculates the working hours and machine hours required to complete the recommended harvesting tasks. The results are matched with the available resources for the day, providing the number of team members, recommended machine types and quantities, and suggested shift arrangements for each route. When resources are insufficient to cover all priority harvesting tasks, the module marks some plots as delayed batches based on maturity scores, risk indicators, and the weight of nearby key agricultural nodes, and provides a delay period and potential yield risk warnings in the management suggestions.

Images