Crop growth vigor sensing and control system based on multi-mode large model

Abstract

The invention discloses a crop growth perception and control system based on a multi-modal large model, and relates to the technical field of artificial intelligence, and the system comprises a data obtaining unit which is used for obtaining a crop canopy image covering a farmland space unit, soil sensor data, environment sensor data and farming operation information in a plurality of collection periods, determining a crop growth basic value and constructing a crop growth graph signal sequence; the growth prediction unit is used for constructing a crop growth Koopman evolution diagram signal processing model on the basis of the farmland space topological diagram and the crop growth diagram signal sequence; and the control execution unit is used for controlling irrigation equipment and fertilization equipment to execute corresponding operation according to the crop growth vigor control strategy. The method can effectively solve the problems of insufficient multi-modal data fusion, lack of spatial correlation modeling, unstable growth prediction, dispersed control strategies and the like in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a crop growth perception and control system based on a multi-modal large model. BACKGROUND

[0002] In the current field of agricultural production, research on crop growth monitoring, precision irrigation and scientific fertilization has been carried out for many years, and a series of public technical routes have been formed. In common growth monitoring technologies, plant phenotype recognition methods widely rely on image processing technology, which judges leaf color, canopy coverage and field growth density through single-frame image color threshold, vegetation index or region segmentation. This method can work in different growth stages of various crops, but its disadvantage is that it usually cannot effectively utilize the correlation between different data sources. For example, images can only reflect the ground canopy shape, and it is difficult to reflect the soil moisture changes in the underground root area; soil data and environmental data are often analyzed separately from image data, resulting in that the growth evaluation results are easily affected by local condition fluctuations and lack comprehensive understanding of the overall characteristics of the farmland space.

[0003] On the other hand, in the scenario of a large spatial distribution of farmland, the public technology usually uses regional mean, interpolation grid or regular partition to describe the growth state of different regions. Although these methods are relatively simple in calculation, they cannot fully model the spatial correlation formed by obvious spatial structure characteristics in the farmland, such as fixed irrigation pipelines, drainage ditches or terrain undulations. Since the topological features of the farmland space are not utilized, the adjacency relationship between different regions cannot be reflected in the growth analysis, so the spatial continuity and multi-directional change trend cannot be reflected in the growth prediction.

[0004] In terms of growth prediction methods, existing technologies mostly use time series fitting methods, such as fitting the historical growth index change sequence into a simple trend line model, or using traditional statistical models to predict the growth period. However, this method often relies on a large number of historical trends and cannot handle spatial changes. In the scenario of a large spatial scale of farmland, the growth evolution of different regions of the farmland is affected by different soil conditions and microclimate, and its change is not linear, nor is each region completely independent. The existing prediction methods often cannot fully capture the overall influence of the current spatial state on the next time spatial state, resulting in limited prediction accuracy. In addition, these methods often have the problem of accumulated prediction deviation when facing sudden weather changes, irrigation operations or agricultural interventions, which affects the stability of automatic control. SUMMARY

[0005] The application aims to provide a crop growth perception and control system based on a multi-modal large model, which can effectively solve the problems of insufficient multi-modal data fusion, lack of spatial correlation modeling, unstable growth prediction and scattered control strategies in the prior art, thereby significantly improving the reliability and overall intelligent level of crop growth monitoring, prediction and control.

[0006] To solve the above technical problems, the application provides a crop growth perception and control system based on a multi-modal large model, which comprises: A data acquisition unit is configured to acquire crop canopy images, soil sensor data, environmental sensor data and farming operation information covering the spatial units of farmland in multiple acquisition periods, and to call a multi-modal large model to analyze agronomic procedure text data to generate a growth level threshold table and a control rule table, and to determine crop growth basic values and build a crop growth map signal sequence based on the image processing results, soil sensor data and environmental sensor data of each spatial unit of farmland. A growth prediction unit is configured to build a crop growth Koopman evolution map signal processing model based on the spatial topology map of farmland and the crop growth map signal sequence, and to obtain a crop growth map signal prediction sequence for multiple prediction periods. A control execution unit is configured to generate a crop growth control strategy based on the crop growth map signal prediction sequence and the growth level threshold table and the control rule table, and to control irrigation equipment and fertilization equipment to perform corresponding operations according to the crop growth control strategy.

[0007] Further, the data acquisition unit divides the target farmland into a grid composed of multiple spatial units of farmland, and each spatial unit of farmland corresponds to a fixed geographic location identifier; in each acquisition period, the crop canopy images covering each spatial unit of farmland are acquired by an image acquisition device; the soil sensor data includes soil moisture data and soil nutrient data; the environmental sensor data includes air temperature, air humidity and light intensity data.

[0008] Furthermore, the data acquisition unit calls a multimodal large model, taking agronomic procedure text data as input, and the multimodal large model generates a growth level threshold table and a control rule table. Image processing operations are performed on the crop canopy image of each farmland spatial unit. These operations include scaling the image to a uniform size, extracting color components, obtaining plant regions through threshold segmentation, and counting the number of pixels in each plant region. Based on this, the average green component value and canopy coverage ratio of the plant regions are calculated, and these values ​​are compared with the corresponding entries in the growth level threshold table to determine the leaf color level and canopy coverage level. Finally, the leaf color level, canopy coverage level, soil moisture data, and soil nutrient data are matched with the records in the growth level threshold table to obtain the basic crop growth values ​​for the corresponding farmland spatial unit.

[0009] Furthermore, the data acquisition unit determines the adjacency relationship based on the relative position of farmland spatial units in the farmland and the connection relationship of irrigation and drainage facilities. When two farmland spatial units share a boundary in the horizontal or vertical direction, or when irrigation pipelines or drainage facilities pass through two farmland spatial units, the adjacency relationship between the two farmland spatial units is recorded as having an adjacency relationship. A farmland spatial topology map is constructed using all farmland spatial units as nodes and pairs of farmland spatial units with adjacency relationships as lines. Within the same acquisition cycle, the basic values ​​of crop growth of all farmland spatial units are matched one-to-one with the nodes in the farmland spatial topology map according to a predetermined order to form a crop growth map signal for the corresponding acquisition cycle. The process of constructing crop growth map signals is repeated in multiple consecutive acquisition cycles to form a sequence of crop growth map signals arranged in chronological order.

[0010] Furthermore, the growth prediction unit constructs a connection matrix based on the farmland spatial topology map. When any two farmland spatial units are connected by a line in the farmland spatial topology map, the first value is recorded in the corresponding row and column position of the connection matrix; when no connection exists, the second value is recorded in the corresponding row and column position. For each farmland spatial unit, all values ​​in the rows related to the corresponding farmland spatial unit are summed in the connection matrix to obtain the number of connections. The number of connections for each farmland spatial unit is then filled into the diagonal position of the diagonal matrix to form a degree matrix. The degree matrix and the connection matrix are then subtracted according to their corresponding row and column positions to obtain the graph Laplacian matrix. A numerical calculation program is called to perform eigenvalue decomposition on the graph Laplacian matrix to obtain a set of eigenvectors and eigenvalues ​​corresponding to each eigenvector. The eigenvectors are then sorted according to the size of the eigenvalues ​​to form an ordered set of eigenvectors.

[0011] Furthermore, for each acquisition cycle, the crop growth prediction unit arranges the basic crop growth values ​​of all farmland spatial units within that acquisition cycle according to the row and column index order of the farmland spatial units to form a basic crop growth vector. The basic crop growth vector is then multiplied element-wise with each feature vector in the ordered feature vector set, and the results are summed to obtain the spectral coefficients on the corresponding feature vectors. The ordered feature vector set is then divided into low-frequency, mid-frequency, and high-frequency feature vector subsets. For each acquisition cycle, the spectral coefficients corresponding to these subsets are averaged to obtain the low-frequency average spectral coefficient, mid-frequency average spectral coefficient, and high-frequency average spectral coefficient. Each average value is then multiplied by the positional value of the corresponding feature vector corresponding to each farmland spatial unit and accumulated to obtain the low-frequency, mid-frequency, and high-frequency growth components for each farmland spatial unit.

[0012] Furthermore, for each farmland spatial unit, the crop growth prediction unit combines the basic crop growth value, low-frequency growth component, medium-frequency growth component and high-frequency growth component in a fixed order to form a crop growth extended feature vector in a certain collection period. The crop growth extended feature vectors of all farmland spatial units in the same collection period are arranged in row and column index order to form a crop growth extended state vector. A crop growth extended state sequence is formed in multiple collection periods.

[0013] Furthermore, the growth prediction unit selects crop growth expansion state vectors from two adjacent acquisition periods in the crop growth expansion state sequence according to time order. The earlier crop growth expansion state vector is used as the current state vector, and the later crop growth expansion state vector is used as the next state vector. All current state vectors are arranged in columns according to the acquisition time order to form the current state matrix, and all next state vectors are arranged in columns according to the corresponding acquisition time order to form the next state matrix. The minimum error solution function in the numerical calculation program is called, and the linear transformation matrix from the current state matrix to the next state matrix is ​​obtained by taking the current state matrix and the next state matrix as input. The linear transformation matrix is ​​used as the crop growth Koopman evolution operator.

[0014] The crop growth sensing and control system based on a multimodal large model of the present invention has the following beneficial effects: This invention, by jointly processing crop growth map signal sequences, multimodal farmland data, and farmland spatial topology maps, simultaneously considers spatial structure, temporal evolution, and multi-source information representation, significantly improving the overall accuracy and stability of crop growth perception and control. By parsing agronomic procedure text data using a multimodal large-scale model, the traditional experience-based growth judgment method can be transformed into a structured growth level threshold table and control rule table, making the growth assessment process consistent and interpretable, and avoiding the subjective differences commonly found in manual judgment.

[0015] This invention utilizes a farmland spatial topology map to construct a graph Laplacian matrix and generates multi-scale spatial feature vectors through eigenvalue decomposition. This effectively expresses the spatial continuity, local fluctuations, and multi-scale structure of crop growth, enabling subsequent prediction stages to naturally distinguish between global trends and local anomalies. By constructing an extended state vector of crop growth containing basic and multi-scale components, this invention leverages the linear properties of the crop growth Koopman evolution operator in time-series prediction to obtain numerically stable prediction results with overall evolutionary consistency throughout the continuous prediction process. Compared to traditional methods based on single-point or independent region prediction, this invention significantly reduces the accumulation of prediction errors.

[0016] Furthermore, this invention generates crop growth control strategies by utilizing crop growth map signal prediction sequences and control rule tables. This eliminates reliance on single sensor signals or fixed thresholds in the control process, instead relying on a comprehensive judgment based on spatial state, temporal trends, and agronomic standards. This results in more precise and coordinated execution of irrigation and fertilization equipment, reducing resource waste and improving crop growth quality. The invention can also optimize the control strategy by incorporating equipment capacity constraints, effectively avoiding instability caused by too many devices operating simultaneously. In summary, this invention provides comprehensive technological improvements in multimodal data fusion, spatial structure utilization, temporal prediction stability, and control strategy generation. It is applicable to various farmland scenarios and has significant practical value and promotional significance. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the spatial distribution pattern of the eigenvectors of the graph Laplacian matrix provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the principle of farmland spatial unit division and graph Laplace matrix construction provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of crop canopy image processing and basic growth value extraction provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of multi-scale growth potential component decomposition provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The crop growth perception and control system based on a multimodal large model includes a data acquisition unit, a growth prediction unit, and a control execution unit. The data acquisition unit is used to acquire image data, soil sensor data, environmental sensor data, and agricultural operation information covering farmland spatial units in multiple acquisition cycles. It also calls the multimodal large model to parse the agronomic procedure text data to generate a growth level threshold table and a control rule table. Based on the image processing results, soil sensor data, and environmental sensor data of each farmland spatial unit, it determines the basic values ​​of crop growth and constructs a crop growth map signal sequence.

[0020] In one implementation, the data acquisition unit first spatially divides the target farmland during system initialization, dividing it into a regular grid composed of multiple farmland spatial units. Each farmland spatial unit corresponds to a fixed geographical location identifier. This geographical location identifier can be a combination of row and column numbers; for example, the farmland spatial unit in the first row and first column is identified as "1-1," the farmland spatial unit in the first row and second column is identified as "1-2," and so on. This regular division method ensures that each data point is accurately associated with a unique farmland spatial unit during subsequent data collection and crop growth map signal sequence construction, avoiding identification errors caused by ambiguous geographical locations.

[0021] In terms of data acquisition, the data acquisition unit acquires crop canopy images, soil sensor data, environmental sensor data, and agricultural operation information covering all farmland spatial units in each acquisition cycle. Crop canopy images can be acquired using fixed camera equipment installed above the farmland. For example, multiple cameras can be deployed on poles on both sides or in the center of the farmland. The camera's shooting height is set, for example, 3 to 5 meters above the ground, ensuring that a single frame image covers a predetermined number of farmland spatial units. The camera's resolution is set, for example, to 1920 pixels by 1080 pixels or 3840 pixels by 2160 pixels, to ensure that each farmland spatial unit corresponds to at least several dozen pixels in the image, thus obtaining stable statistical results in subsequent image processing operations. Alternatively, in another implementation, a drone equipped with cameras can be used to fly at low altitude along a preset route, capturing images of the farmland at a fixed flight altitude and overlap rate in each acquisition cycle. The acquired image data is then uploaded to the data acquisition unit for unified processing. Using fixed camera equipment simplifies the system structure, and the data collection cycle can be set to once a day or multiple times a day; using drones is suitable for large areas of farmland and can improve coverage efficiency.

[0022] refer to Figure 3 , Figure 3 The principle and process of crop canopy image processing and basic growth value extraction are shown. Figure 3 The original canopy image is shown on the left. Figure 3 The pixel classification results are shown in the middle. Figure 3 The statistical parameter extraction process is shown on the right and bottom. Figure 3 On the left, raw crop canopy images are shown, acquired using fixed cameras or drones. These images contain multiple irregularly distributed areas of plant leaves, each displaying different shades of gray to represent leaf color variations. The background is bare soil. Each raw canopy image corresponds to the geographical extent of a single farmland spatial unit, and the image size has been uniformly scaled, for example, to 256 pixels × 256 pixels, to ensure the comparability of subsequent statistical results. Figure 3The pixel classification results are shown in the center. By thresholding the green component value of each pixel in the original canopy image, the pixels in the image are divided into plant region pixels and bare soil region pixels. When the green component value of a pixel falls within the preset threshold range, the pixel is marked as a plant region pixel and displayed in dark gray in the pixel classification result image; when the green component value of a pixel is significantly lower than the preset threshold range, the pixel is marked as a bare soil region pixel and displayed in white in the pixel classification result image. The pixel classification result image clearly shows the spatial distribution of plant regions and background regions, with each small square representing a pixel unit. This segmentation method based on color component thresholding can achieve accurate identification of plant regions without using complex deep learning models.

[0023] Figure 3 The bottom left shows the statistical parameter extraction process. First, the number of pixels in the plant region, Np, is counted, which is the total number of pixels marked as plant regions in the pixel classification result image; at the same time, the total number of pixels in the image, Ntotal, is recorded. By calculating the ratio of the number of pixels in the plant region to the total number of pixels, the canopy cover ratio C is obtained. This ratio directly reflects the degree of ground cover of the crop in this farmland spatial unit. In addition, the green component values ​​of all plant region pixels are accumulated and divided by the number of plant region pixels Np to obtain the average green component G. This value reflects the overall depth of leaf color and is closely related to the nitrogen content and growth status of the leaves. Figure 3 The bottom right shows the process for determining the baseline crop growth values. Leaf color level is determined using a mapping function based on the average green component G, and canopy cover level is determined using a mapping function based on the canopy cover ratio C. The leaf color level and canopy cover level are then comprehensively matched with soil moisture and nutrient data obtained from soil sensors. By querying the growth level threshold table generated by the multimodal large model, the baseline crop growth values ​​for this farmland spatial unit are finally determined. These baseline crop growth values ​​comprehensively reflect the crop growth status of this farmland spatial unit, providing basic data for the subsequent construction of crop growth map signals.

[0024] Soil sensor data includes soil moisture and soil nutrient data. Soil moisture sensors can be buried near the center of each farmland spatial unit, at a depth of 10 to 30 centimeters below the soil surface, to reflect the water content of the main root distribution area. Soil nutrient sensors can be deployed in rows or columns, for example, one group of soil nutrient sensors can be shared by every four farmland spatial units, and the collected soil nutrient data can be distributed to the corresponding farmland spatial units through interpolation. Environmental sensor data includes air temperature, air humidity, and light intensity data, which can be obtained through environmental monitoring points deployed at the edges or in the fields. The deployment density of environmental monitoring points can be configured according to the farmland area and topographical complexity. In small to medium-sized farmlands with minimal topographical changes and relatively stable wind direction, only one environmental monitoring point is sufficient. Agricultural operation information can be recorded through a management system connected to agricultural machinery terminals. For example, at the beginning and end of operations such as fertilization, spraying, tilling and sowing, the operator selects the operation type, operation area and operation time on the agricultural machinery terminal. The data acquisition unit reads the agricultural operation information from the management system periodically and stores the agricultural operation information in association with the corresponding farmland space unit.

[0025] In terms of text information processing, the data acquisition unit calls the multimodal big data model to parse the agronomic procedure text data to generate a growth level threshold table and a control rule table. The multimodal big data model can be deployed on a local server or a remote server. In one implementation, the data acquisition unit sends the agronomic procedure text data to the multimodal big data model in raw text form. The multimodal big data model segments and divides the text into sentences, distinguishing between statements describing crop growth assessment conditions and statements describing control measures. For statements describing growth assessment conditions, such as "when the leaves are dark green and the soil area in the field is relatively small, it is judged as vigorous growth," the multimodal big data model, based on its internally trained language mapping capabilities, maps language descriptions such as "dark green" and "relatively small soil area" to operable numerical ranges. For example, it uses conditions such as the average green component of leaves falling within a certain numerical range and the canopy coverage ratio being higher than a certain percentage as conditions for vigorous growth.

[0026] For statements describing control measures, such as "increase irrigation when growth is weak and soil moisture is low," the multimodal large model extracts the conditional relationships between "weak growth," "low soil moisture," and "increase irrigation" to generate structured control rule entries. The generated growth level threshold table records the correspondence between different growth levels and leaf color, canopy cover, soil moisture, and soil nutrients. The generated control rule table records the combination of growth level and soil condition conditions, along with the corresponding control actions. In this way, agronomic experience that was originally in textual form can be transformed into structured data that can be directly used for automated processing, making subsequent growth assessment and control strategy generation independent of manual interpretation of textual descriptions.

[0027] In terms of image processing, the data acquisition unit performs image processing operations on the crop canopy image corresponding to each farmland spatial unit to obtain the image processing results. Taking a single frame image acquired by a fixed camera as an example, the data acquisition unit first establishes the correspondence between image coordinates and farmland spatial units based on the camera's installation position, shooting angle, and the layout of the farmland spatial units. The entire image is then divided into multiple regions, each corresponding to a farmland spatial unit. When the camera's field of view contains multiple farmland spatial units, pre-calibration allows for precise mapping of pixel rows and columns in the image to specific farmland spatial units, thus dividing the entire image into several sub-images. Each sub-image is the crop canopy image corresponding to a farmland spatial unit. For each crop canopy image, the data acquisition unit scales the image to a uniform size, for example, to 256 pixels by 256 pixels. The advantage of a uniform size is that when comparing statistical results between different acquisition periods, each image has the same number of pixels, allowing direct comparison of the number of pixels in the plant area and the average color value, avoiding statistical bias caused by differences in image size.

[0028] refer to Figure 2 , Figure 2 This illustrates the principle of dividing farmland into spatial units and constructing the Graph Laplace matrix. Figure 2 The left side shows the grid division structure of farmland spatial units. Figure 2 The right side shows how the connection matrix A and the degree matrix D are constructed. Figure 2On the left, the target farmland is divided into a regular grid consisting of multiple farmland spatial units, comprising 36 units in 6 rows and 6 columns. Each farmland spatial unit corresponds to a unique node number, assigned sequentially from top to bottom and left to right. The farmland spatial unit in the 1st row and 1st column is numbered 1, the one in the 1st row and 2nd column is numbered 2, and so on until the one in the 6th row and 6th column is numbered 36. The dashed lines in the diagram indicate the adjacency connections between nodes. Taking node 15 in the 3rd row and 3rd column as an example, this node has adjacent lines connecting it to node 9 above, node 21 below, node 14 to the left, and node 16 to the right, indicating that these farmland spatial units are physically adjacent. Figure 2 The upper right side shows how the connection matrix A is constructed. The number of rows and columns in connection matrix A is equal to the number of farmland spatial units; the diagram exemplifies a 6×6 matrix structure. When any two farmland spatial units are adjacent, a value of 1 is recorded in the corresponding row and column position of connection matrix A; when two farmland spatial units are not adjacent, a value of 0 is recorded in the corresponding row and column position. The diagonal positions of the matrix correspond to the same farmland spatial unit, therefore all diagonal elements are 0.

[0029] In this way, the connectivity matrix A numerically and completely describes the adjacency relationships between the various farmland spatial units in the farmland spatial topology map. Figure 2 The lower right section illustrates the construction of the degree matrix D. The degree matrix D is a diagonal matrix, where the values ​​at the diagonal positions represent the number of connections between corresponding farmland spatial units. Specifically, by summing each row of the connectivity matrix A, the number of connections between the corresponding farmland spatial unit and other farmland spatial units in that row is obtained, and this number is then filled into the corresponding diagonal position of the degree matrix D. For example, farmland spatial units located at the edge of a field typically have 2 or 3 adjacent units, corresponding to a diagonal value of 2 or 3 in the degree matrix; farmland spatial units located inside a field typically have 4 adjacent units, corresponding to a diagonal value of 4 in the degree matrix. All off-diagonal elements of the degree matrix D are 0. The formula for calculating the graph Laplacian matrix L is: L = DA. The graph Laplacian matrix L is obtained by performing a difference operation between the degree matrix D and the connectivity matrix A at corresponding positions. The graph Laplacian matrix L contains both the connection strength information of each farmland spatial unit and the adjacency information, providing a mathematical foundation for subsequent graph signal processing and eigenvalue decomposition.

[0030] During color processing, the data acquisition unit converts each crop canopy image into a color representation containing red, green, and blue components. In one implementation, the red, green, and blue component values ​​of each pixel can be directly read from the original color image. To distinguish between plant areas and background areas, the data acquisition unit initially selects several representative images. Within these images, operators circle typical plant and non-plant areas, calculating the distribution of green component values ​​for pixels in plant areas and non-plant areas, thus determining a preset threshold range that effectively distinguishes the two types of pixels. During operation, the data acquisition unit traverses each image pixel-by-pixel. Pixels with green component values ​​falling within the preset threshold range are marked as plant area pixels. Pixels with green component values ​​significantly below this range are marked as bare soil area pixels. Pixels with green component values ​​significantly above this range and accompanied by abnormal blue or red component values ​​can be identified as bright reflective areas or non-target areas. For example, when using plastic film for covering, the reflective characteristics of the plastic film differ significantly from those of plant leaves, and this can be identified through the combination of green component values ​​with other components.

[0031] After the plant region pixels are labeled, the data acquisition unit calculates the ratio between the number of pixels in the plant region and the total number of pixels in the entire image to obtain the canopy cover ratio. The canopy cover ratio directly reflects the degree of ground cover of the crop in that farmland spatial unit; a higher canopy cover ratio generally indicates a larger number of plants and denser growth. Because the canopy cover ratio is calculated based on the number of pixels, it is less susceptible to local anomalies compared to indicators that rely solely on single-point measurements. For example, changes caused by a single hole or localized lodging are smoothed out by the overall pixel statistics, thus improving the stability of growth assessment to some extent. Simultaneously, the data acquisition unit accumulates the green component values ​​of the plant region pixels and divides them by the number of pixels in the plant region to obtain the average green component value. The average green component value reflects the depth of leaf color. When the nitrogen content of the leaves is low or early senescence occurs, the leaf color usually gradually lightens from dark green, and the average green component value changes accordingly. By combining the canopy cover ratio and the average green component value, a more comprehensive assessment of growth can be made.

[0032] Soil moisture data from soil sensors can be directly read using capacity-type or time-domain reflectometry sensors. The data acquisition unit can obtain soil moisture ranges for different crop growth stages from the description of suitable soil moisture ranges in the agronomic regulations and the growth level threshold table generated by the multimodal large model. For example, soil moisture is required to be within a certain range during the seedling stage, and within a slightly higher range during the heading stage. For each farmland spatial unit within each acquisition cycle, the data acquisition unit compares the soil moisture data with the corresponding soil moisture range and categorizes the soil moisture data as suitable, low, or high. Soil nutrient data can be obtained using conductivity sensors or ion-selective electrodes. Similarly, the data acquisition unit categorizes the soil nutrient data as suitable or insufficient based on the soil nutrient ranges recorded in the growth level threshold table. Air temperature, air humidity, and light intensity data from environmental sensors can be averaged according to the acquisition cycle or have representative values ​​calculated within a preset time window, thus being stored in association with the corresponding acquisition cycle. Although environmental sensor data does not directly generate growth levels in this embodiment, it can serve as supplementary information for reference when generating growth level threshold tables and control rule tables.

[0033] When determining the baseline values ​​for crop growth, the data acquisition unit matches the leaf color level, canopy cover level, soil moisture data, and soil nutrient data corresponding to each farmland spatial unit with the records in the growth level threshold table. The growth level threshold table can include multiple growth levels, such as weak, average, good, and vigorous. Each growth level corresponds to a set of condition ranges; for example, a certain growth level requires a medium or dark leaf color level, a certain canopy cover level, suitable soil moisture, and soil nutrient levels not lower than suitable. For a given farmland spatial unit, the data acquisition unit compares its leaf color level and canopy cover level with each level condition in the growth level threshold table. When a combination of conditions for a certain growth level is found to be compatible with the leaf color level and canopy cover level of the farmland spatial unit, the soil moisture data and soil nutrient data are further determined to fall within the soil interval corresponding to that growth level. If all conditions match, the baseline crop growth value corresponding to that growth level is assigned to the farmland spatial unit. The basic values ​​of crop growth can be represented in integer form, for example, weak, average, good, and vigorous can be represented as 1, 2, 3, and 4, respectively. In this way, leaf color, canopy cover, and soil conditions, which were originally composed of multiple dimensions, are reduced to a single value, which not only preserves multidimensional information but also facilitates subsequent graph signal processing and crop growth Koopman evolution graph signal processing based on farmland spatial topology maps.

[0034] When constructing a crop growth map signal sequence from basic crop growth data, the data acquisition unit arranges the basic crop growth data of all farmland spatial units in a predetermined order within each acquisition cycle, corresponding one-to-one with nodes in the farmland spatial topology map. The predetermined order can be row-major, first arranging by row number from smallest to largest, then by column number within the same row from smallest to largest. This fixed order ensures that the value at a certain position always corresponds to the same farmland spatial unit across different acquisition cycles, facilitating comparison and analysis over time. The set of basic crop growth data generated within each acquisition cycle constitutes the crop growth map signal for that cycle. As the acquisition cycle progresses, the data acquisition unit continuously adds new crop growth map signals to storage, forming a crop growth map signal sequence arranged chronologically. This crop growth map signal sequence reflects both the spatial distribution of farmland and records the temporal evolution process, and can be directly used as input for the crop growth prediction unit to perform Koopman evolution map signal processing.

[0035] In another optional implementation, the data acquisition unit can employ a segmentation method based on a combination of brightness and hue during image processing. Specifically, the crop canopy image can be converted into a color representation composed of brightness, hue, and saturation, and pixels with lower brightness and hues falling within the green range are selected as plant region pixels. Compared to using only the green component threshold, this method can better exclude non-plant areas and reduce false positives when there are bright sky areas or reflective areas. In this implementation, the average green component value can be replaced by an average hue or a combination of average brightness and saturation, with the hue and saturation range conditions recorded accordingly in the growth level threshold table. This implementation and the aforementioned implementations can be selected or combined based on the actual sensor performance and crop variety characteristics.

[0036] The crop growth prediction unit is used to construct a crop growth Koopman evolution map signal processing model based on farmland spatial topology map and crop growth map signal sequence, and obtain crop growth map signal prediction sequence for multiple prediction periods.

[0037] In one implementation, after receiving the farmland spatial topology map and the crop growth map signal sequence arranged in chronological order provided by the data acquisition unit, the growth prediction unit first constructs a basic matrix structure for graph signal processing based on the farmland spatial topology map, then constructs a crop growth Koopman evolution graph signal processing model on this basis, and finally outputs a crop growth map signal prediction sequence covering multiple prediction periods.

[0038] In its implementation, the growth prediction unit assigns a unique row and column index number to each farmland spatial unit in the farmland spatial topology map. For example, it can assign number 1 to the farmland spatial unit in the first row and first column, number 2 to the farmland spatial unit in the first row and second column, and so on, until all farmland spatial units are numbered. Then, the growth prediction unit constructs a connection matrix based on the pairs of farmland spatial units with adjacent relationships recorded in the farmland spatial topology map. Specifically, the number of rows and columns of the connection matrix is ​​equal to the number of farmland spatial units. In the connection matrix, when any two farmland spatial units are connected by a line in the farmland spatial topology map, a first value is recorded in the corresponding row and column position of the connection matrix, for example, recorded as 1. When two farmland spatial units are not connected by a line in the farmland spatial topology map, a second value is recorded in the corresponding row and column position, for example, recorded as 0. In this way, each row reflects the connection between the corresponding farmland spatial unit and other farmland spatial units, and each column corresponds one-to-one with it. The connection matrix as a whole describes the discrete structure of spatial adjacency relationships in farmland.

[0039] After obtaining the connectivity matrix, the growth prediction unit sums all the values ​​in the corresponding row of each farmland spatial unit to obtain the number of connections for that farmland spatial unit. For example, if a farmland spatial unit is connected to four farmland spatial units (up, down, left, and right), its number of connections is 4. The growth prediction unit then fills the connection counts of all farmland spatial units into the diagonal positions of a diagonal matrix according to their row and column indices, forming a degree matrix. Each diagonal position of the degree matrix corresponds to a farmland spatial unit, recording the connection strength between that farmland spatial unit and its surrounding farmland spatial units, representing the number of "neighbors" of that farmland spatial unit in the entire farmland topology map. Finally, the growth prediction unit performs a difference operation between the degree matrix and the connectivity matrix according to their corresponding row and column positions to obtain the graph Laplacian matrix. The Graph Laplace matrix reflects the relative relationships between farmland spatial units on a farmland spatial topology map. When there is a significant difference in the basic crop growth values ​​between two adjacent farmland spatial units, the correlation quantity constructed through the Graph Laplace matrix will become larger, thus naturally tending to balance the growth difference between neighboring farmland spatial units in subsequent processing. In this way, the Graph Laplace matrix transforms the intuitive understanding that "the growth of adjacent farmland spatial units should have a certain continuity" into a structure that can be directly used in numerical calculations.

[0040] The crop growth prediction unit calls the eigenvalue decomposition function in the numerical calculation program to perform eigenvalue decomposition on the graph Laplacian matrix, obtaining a set of eigenvectors and corresponding eigenvalues ​​for each eigenvector. Eigenvectors can be viewed as a series of spatial distribution patterns formed on the farmland spatial topology map. The value at each position in each eigenvector represents the degree of participation of different farmland spatial units within that spatial distribution pattern. The magnitude of the eigenvalue reflects the rate of change of the spatial distribution pattern. Generally, eigenvectors with smaller eigenvalues ​​show more gradual changes between neighboring farmland spatial units, more closely resembling a large-scale, slowly changing growth distribution; eigenvectors with larger eigenvalues ​​show more dramatic changes between adjacent farmland spatial units, more closely resembling local, detailed differences in growth. The growth prediction unit sorts all eigenvalues ​​in ascending order and rearranges the eigenvectors according to the sorting result, forming an ordered set of eigenvectors. This ordered set of eigenvectors allows for convenient subsequent decomposition of information in the crop growth map signal into three scales: low-frequency, mid-frequency, and high-frequency, facilitating the analysis of overall growth trends and local anomalies.

[0041] refer to Figure 1 , Figure 1 The distribution patterns of three typical eigenvectors on the farmland spatial topology map are shown to illustrate the differences in the spatial scale of change represented by eigenvectors of different frequencies. Figure 1 The diagram contains three independent spatial distribution grids from left to right, corresponding to low-frequency feature vectors, mid-frequency feature vectors, and high-frequency feature vectors, respectively. Figure 1 The left side shows the spatial distribution pattern of feature vector 1, whose corresponding eigenvalue λ1 is 0.15, belonging to the low-frequency feature vector category. This part consists of a regular grid of 64 farmland spatial units arranged in 8 rows and 8 columns. Each square in the grid represents a farmland spatial unit, and the gray level filling the square indicates the magnitude of the feature vector at the corresponding spatial location. A specific numerical value is also marked at the center of each square. The values ​​are represented to two decimal places, ranging from -1.00 to 1.00. From the gray level distribution of this grid, it can be observed that the entire 8×8 area exhibits a large-scale smooth gradient characteristic, with relatively small numerical changes between adjacent farmland spatial units, showing a slow transition from the upper left to the lower right corner. Specifically, the values ​​in the upper left corner of the grid are mostly positive and darker, while the values ​​in the lower right corner gradually transition to negative values ​​and lighter gray levels, with the middle area being the transition region. This distribution pattern reflects the large-scale, low-frequency spatial variation pattern represented by the low-frequency feature vector, that is, it shows an overall trend change over a large spatial range, which is suitable for expressing the macroscopic distribution of the overall growth of farmland.

[0042] Figure 1The central section shows the spatial distribution pattern of feature vector 15, and the corresponding eigenvalue λ. 15 The value is 2.34, falling within the mid-frequency eigenvector category. This part also consists of a regular 8x8 grid, with the same grid structure as the left side. Within this grid, the grayscale distribution exhibits a spatial variation pattern significantly different from the left side. Specifically, some areas within the grid display positive values ​​with darker grayscale, while others display negative values ​​with lighter grayscale, forming a clear striped or blocky alternation pattern. The numerical variation amplitude between adjacent farmland spatial units is larger than that of the low-frequency eigenvector, but remains relatively smooth compared to the high-frequency eigenvector. For example, the grid may show a distribution where several columns on the left are positive and several columns on the right are negative, or the upper and lower halves show opposite signs. This distribution pattern reflects the medium-scale spatial variation law represented by the mid-frequency eigenvector, capturing the differences in growth between areas composed of several adjacent spatial units in farmland. It is suitable for expressing the growth distribution characteristics between different management zones or areas with different terrain conditions within a field.

[0043] Figure 1 The right side shows the spatial distribution pattern of feature vector 30, and the corresponding eigenvalue λ. 30 The value is 5.67, falling into the category of high-frequency feature vectors. This part also employs an 8x8 grid structure. Within this grid, the grayscale distribution exhibits a rapidly oscillating spatial variation pattern, with significant numerical differences between adjacent farmland spatial units. Positive and negative values ​​alternate with a high frequency in space. Specifically, a particular farmland spatial unit may have a large positive value, while its immediate neighbors (up, down, left, and right) may have large negative values, forming a checkerboard-like or finely striped distribution pattern. This distribution pattern reflects the local details and high-frequency spatial variation patterns characterized by high-frequency feature vectors, capable of capturing abnormal growth or localized changes within a single or a few farmland spatial units. It is suitable for expressing small-scale differences in growth caused by factors such as pests and diseases, localized waterlogging, and soil anomalies.

[0044] Figure 1 Each grid cell is labeled above its corresponding eigenvector number, eigenvalue magnitude, and frequency band type. For example, the left grid cell is labeled "Eigenvector 1 (λ1=0.15)" with the annotation "Low frequency - large scale variation," while the middle grid cell is labeled "Eigenvector 15 (λ1=0.15)." 15 =2.34)" and noted "intermediate frequency-intermediate scale variation", the grid on the right is labeled "eigenvector 30(λ)". 30=5.67)" and noted "High-frequency - local detail changes". These annotations clearly indicate the sequence number of each eigenvector, its corresponding eigenvalue, and its frequency range, facilitating the understanding of the role of different eigenvectors in multi-scale analysis of crop growth. Through Figure 1 The three typical eigenvector spatial distribution patterns shown can intuitively illustrate the role of graph Laplacian matrix eigenvalue decomposition in crop growth map signal processing. Projecting the crop growth map signal onto these eigenvectors of different frequencies decomposes the complex growth spatial distribution into a superposition of large-scale trend components, medium-scale regional components, and local detail components, thus providing a richer and more structured state description foundation for subsequent crop growth Koopman evolution map signal processing models.

[0045] When processing crop growth map signal sequences, the growth prediction unit, for each acquisition period, first arranges the basic crop growth values ​​of all farmland spatial units within that period according to the row and column index order of the farmland spatial units, forming a basic crop growth vector. For example, in a farmland containing 100 farmland spatial units, the basic crop growth vector contains 100 values, each value corresponding to the basic crop growth value of one farmland spatial unit. The growth prediction unit performs element-wise multiplication and addition of the basic crop growth vector with each feature vector in the ordered feature vector set, and sums the results to obtain the spectral coefficients on the corresponding feature vectors. Element-wise multiplication and addition means multiplying the value at each position in the basic crop growth vector with the value at the same position in the current feature vector, and then summing all the products to obtain a single real number, which is the spectral coefficient on the current feature vector for that acquisition period.

[0046] Spectral coefficients represent the "projection degree" of the basic crop growth vector onto a certain spatial distribution pattern. A larger spectral coefficient indicates a strong consistency between the current basic crop growth numerical distribution and the spatial distribution pattern represented by the corresponding feature vector. For example, if a feature vector is positive overall on the left side of the field and negative overall on the right side, this feature vector represents an "east-west growth difference pattern." If the corresponding spectral coefficient is large, it indicates a significant east-west growth difference during the current data collection period. By calculating the spectral coefficient for each feature vector, the growth prediction unit transforms the original growth distribution on the farmland spatial topology map into a combined representation on a set of spatial distribution patterns, facilitating subsequent grouping, compression, and expansion of these combinations.

[0047] After the ordered feature vector set is constructed and the spectral coefficients for each acquisition cycle are calculated, the growth prediction unit sequentially divides the ordered feature vector set into low-frequency, mid-frequency, and high-frequency feature vector subsets. In one specific implementation, the ordered feature vector set can be divided into three segments based on the number of feature values. For example, when the ordered feature vector set contains 60 feature vectors, the first 20 feature vectors can be divided into the low-frequency feature vector subset, the middle 20 feature vectors into the mid-frequency feature vector subset, and the last 20 feature vectors into the high-frequency feature vector subset. The feature vectors in the low-frequency feature vector subset correspond to smaller feature values, suitable for expressing large-scale, smooth growth trends; the feature vectors in the high-frequency feature vector subset correspond to larger feature values, more suitable for expressing fine-grained fluctuations in local areas, such as abnormal growth near a certain field ridge; the mid-frequency feature vector subset lies in between, taking into account medium-scale changes. This grouping method provides a clear hierarchy for subsequent multi-scale analysis without changing the basic feature structure.

[0048] For each acquisition period, the growth prediction unit averages the spectral coefficients corresponding to the low-frequency, mid-frequency, and high-frequency feature vector subsets, respectively, to obtain the low-frequency average spectral coefficient, mid-frequency average spectral coefficient, and high-frequency average spectral coefficient. The averaging process can be viewed as integrating the contributions of multiple similar spatial distribution patterns within the same frequency band to obtain a value representing the overall contribution of that frequency band. This approach reduces the number of values ​​that need subsequent processing, compressing multiple spectral coefficients into a few representative values ​​while maintaining the overall strength relationship across different frequency bands. Based on this, the growth prediction unit traverses each feature vector in the low-frequency feature vector subset. For each farmland spatial unit, it reads the position value corresponding to that farmland spatial unit from the feature vector, multiplies this position value by the low-frequency average spectral coefficient, and accumulates all the product results in the low-frequency feature vector subset to obtain the low-frequency growth component of that farmland spatial unit in that acquisition period. The calculation process for the mid-frequency and high-frequency growth potential components is the same as that for the low-frequency growth potential components, using the mid-frequency eigenvector subset and mid-frequency average spectral coefficient, and the high-frequency eigenvector subset and high-frequency average spectral coefficient, respectively.

[0049] The low-frequency, mid-frequency, and high-frequency growth components reflect the growth characteristics of the farmland spatial unit at different spatial scales. The low-frequency growth component focuses on reflecting the overall growth over a large area, such as whether the entire plot is generally weak or strong; the high-frequency growth component focuses more on local differences, such as abnormal growth in a small area caused by disease or waterlogging. By decomposing the basic crop growth values ​​into a basic component and three frequency band components, the growth prediction unit provides a richer state description for the subsequent construction of the crop growth Koopman evolution diagram signal processing model, enabling the model to not only fit overall trend changes but also pay attention to spatial details of evolution.

[0050] After obtaining the low-frequency, mid-frequency, and high-frequency growth components, the growth prediction unit, for each farmland spatial unit, combines the basic crop growth values, low-frequency growth components, mid-frequency growth components, and high-frequency growth components of that farmland spatial unit in a fixed order during a certain acquisition period to form a crop growth extended feature vector. For example, the combination order can be agreed upon as basic crop growth values, low-frequency growth components, mid-frequency growth components, and high-frequency growth components, and this order can be maintained throughout the entire system to ensure accurate reconstruction of each component during subsequent decomposition. The growth prediction unit arranges the crop growth extended feature vectors of all farmland spatial units within the same acquisition period in the row and column index order of the farmland spatial units, and connects them into a longer vector, which is the crop growth extended state vector for that acquisition period. As multiple data collection cycles progress, the crop growth prediction unit can obtain a sequence of crop growth extension states arranged in chronological order. This sequence of crop growth extension states unifies multi-scale information in space with the evolutionary order in time, providing a complete data foundation for the subsequent construction of the crop growth Koopman evolution operator.

[0051] refer to Figure 4 , Figure 4 The principle of multi-scale component decomposition of basic crop growth values ​​is shown. Figure 4 The spatial distribution patterns of the original growth potential distribution, low-frequency growth potential component, mid-frequency growth potential component, and high-frequency growth potential component are shown from left to right. Figure 4 On the far left, the original crop growth distribution of a specific area within a given data collection period is shown. This distribution is presented in a 7×7 grid, with each grid cell corresponding to a farmland spatial unit. The grayscale value within each cell represents the baseline crop growth value for that unit; darker grayscale values ​​indicate greater growth. The original crop growth distribution exhibits a complex spatial variation pattern, encompassing both large-scale overall trend changes and detailed undulations in local areas, reflecting the actual spatial heterogeneity of crop growth in farmland. Figure 4The second column from the left shows the spatial distribution of the low-frequency growth component. This component is obtained by projecting and reconstructing the original growth distribution onto a subset of low-frequency eigenvectors; it corresponds to the eigenvector with the smallest eigenvalue in the graphical Laplacian matrix. The low-frequency growth component exhibits a smooth spatial gradient, with relatively slow numerical changes between adjacent farmland spatial units, showing an overall trend from the upper left to the lower right or from one side to the other. This large-scale smooth variation typically reflects overall differences in soil fertility, topographic slope, or irrigation conditions across the entire field, representing macroscopic factors affecting crop growth.

[0052] exist Figure 4 The second column from the right shows the spatial distribution of the mid-frequency growth potential component. This component corresponds to the eigenvectors in the graphical Laplacian matrix whose eigenvalues ​​are in the medium range; spatially, it lies between low and high frequencies. The mid-frequency growth potential component exhibits a periodic or semi-periodic spatial oscillation pattern, with some numerical fluctuations between adjacent farmland spatial units, but the amplitude of these fluctuations is relatively moderate. This medium-scale variation pattern may reflect zonal differences in field management operations, small-scale intercropping of crop varieties, or medium-scale variations in soil texture. Figure 4 On the far right, the spatial distribution of the high-frequency growth component is shown. This high-frequency growth component corresponds to the eigenvector with the largest eigenvalue in the graphical Laplacian matrix; it captures rapid changes and local details in the original growth distribution. The high-frequency growth component exhibits dramatic spatial fluctuations, with significant differences in values ​​between adjacent farmland spatial units, forming distinct patchy or checkerboard patterns. These rapid, small-scale changes typically reflect differences in growth caused by local factors such as pest and disease infestation, localized waterlogging, field ridge effects, or abnormal development of individual crops.

[0053] When constructing the Koopman evolution operator for crop growth, the growth prediction unit selects the crop growth expansion state vectors from two adjacent collection periods in chronological order within the crop growth expansion state sequence. The earlier crop growth expansion state vector is taken as the current state vector, and the later crop growth expansion state vector is taken as the next state vector. For the entire crop growth expansion state sequence, the growth prediction unit arranges all current state vectors in chronological order to form a current state matrix, and arranges all next state vectors in chronological order to form a next state matrix. This can be understood as the current state matrix recording the crop growth expansion state at the beginning of each collection period, while the next state matrix recording the crop growth expansion state in the next collection period after the end of the corresponding collection period.

[0054] The crop growth prediction unit calls the minimum error solution function in the numerical calculation program. Taking the current state matrix and the next state matrix as input, it solves for the linear transformation matrix from the current state matrix to the next state matrix, and uses this linear transformation matrix as the crop growth Koopman evolution operator. The minimum error solution function finds a linear transformation matrix such that the overall difference between each column vector in the current state matrix and the corresponding column vector in the next state matrix after transformation is minimized. The crop growth Koopman evolution operator obtained in this way can approximately describe the relationship of "how the crop growth extension state changes as a whole within a collection period". Compared to directly fitting the change relationship on the original crop growth base values, constructing a linear transformation matrix in the crop growth extension state vector space can simultaneously consider the changes of the base components and multi-scale components, making the crop growth Koopman evolution graph signal processing model more fully characterize the complex growth evolution.

[0055] In some cases, directly using the obtained crop growth Koopman evolution operator for multi-step iterative prediction may result in the prediction results gradually increasing in value or approaching zero, leading to distortion of the crop growth map signal prediction sequence. Therefore, the growth prediction unit performs stabilization processing on the crop growth Koopman evolution operator. Specifically, the growth prediction unit calls the eigenvalue decomposition function in the numerical calculation program to perform eigenvalue decomposition on the crop growth Koopman evolution operator, obtaining a set of operator eigenvectors and corresponding operator eigenvalues. Each operator eigenvalue represents the amplification or reduction factor of the change along the direction of the corresponding operator eigenvector. When the absolute value of a certain operator eigenvalue is significantly greater than 1, the component along that direction will become increasingly larger during multiple iterations, potentially leading to an unreasonable explosive growth in the overall state vector; conversely, when the absolute value of a certain operator eigenvalue is significantly less than 1, the component along that direction will rapidly decay during multiple iterations, causing some effective change patterns to be prematurely flattened. The crop growth prediction unit pre-sets an allowable range for feature values. For example, in one implementation, this is set to an absolute value not exceeding 2 and not less than 0.1. When the absolute value of a certain operator feature value is greater than the upper limit of the allowable range, the corresponding operator feature value is replaced with the upper limit of the allowable range. When the absolute value of a certain operator feature value is less than the lower limit of the allowable range, the corresponding operator feature value is replaced with the lower limit of the allowable range. When the absolute value of a certain operator feature value is within the allowable range, the corresponding operator feature value remains unchanged. Subsequently, while keeping the operator feature vector unchanged, the crop growth prediction unit uses the adjusted operator feature values ​​and corresponding operator feature vectors to reconstruct the stabilized crop growth Koopman evolution operator using the inverse operation of feature decomposition. In this way, the relative importance of different modes in crop growth evolution can be maintained, and the numerical runaway of some modes during the prediction process can be avoided, thus improving the stability of multi-step iterative prediction.

[0056] When performing crop growth prediction, the growth prediction unit selects the crop growth map signal from the last acquisition period in the crop growth map signal sequence. Following the aforementioned methods for constructing the crop growth expansion feature vector and crop growth expansion state vector, it constructs the current crop growth expansion state vector. The growth prediction unit multiplies the current crop growth expansion state vector with the stabilized crop growth Koopman evolution operator according to matrix multiplication rules to obtain the crop growth expansion state vector for the next time step. This next time step crop growth expansion state vector is then used as the new current crop growth expansion state vector, and the same matrix multiplication operation is repeated to obtain crop growth expansion state vectors for multiple prediction periods. Compared to predicting only one future time step, this iterative approach can generate the crop growth evolution process for multiple consecutive prediction periods, which is closer to the continuous decision-making needs in actual farmland management.

[0057] For each prediction period's crop growth expansion state vector, the growth prediction unit, following the fixed order used when constructing the crop growth expansion feature vector, splits this vector into crop growth expansion feature vectors for each farmland spatial unit. For each feature vector, the growth prediction unit reads the corresponding basic crop growth values ​​and fills these values ​​into a grid consistent with the farmland spatial topology, according to the row and column index order of the farmland spatial units, forming the crop growth map signal for the corresponding prediction period. By repeating this splitting and restoration process for all prediction periods, a complete crop growth map signal prediction sequence can be obtained. The crop growth map signal prediction sequence maintains the same data structure as the original crop growth map signal sequence and can be directly shared with the control execution unit for further generation of crop growth control strategies.

[0058] The control execution unit is used to generate a crop growth control strategy based on the crop growth map signal prediction sequence, the growth level threshold table, and the control rule table, and to control the irrigation and fertilization equipment to perform corresponding operations according to the crop growth control strategy.

[0059] The control execution unit, for each farmland spatial unit in each prediction period of the crop growth map signal prediction sequence, compares the basic crop growth value of the corresponding farmland spatial unit with the various growth level intervals in the growth level threshold table to determine the growth level of the corresponding farmland spatial unit in the corresponding prediction period, and records the prediction period identifier, farmland spatial unit identifier, and growth level in the growth evaluation table. It then iterates through each record in the growth evaluation table, and for each record, sequentially reads each control rule from the control rule table, compares the growth level in the record with the preset trigger conditions in the control rule table. When the trigger condition in a certain control rule matches the growth level in the corresponding record, it reads the preset control action from that control rule, combines the preset control action with the farmland spatial unit identifier and the corresponding prediction period to form a control strategy entry, and adds the control strategy entry to the crop growth control strategy list.

[0060] The control execution unit merges the control strategy entries for the same time period, the same irrigation equipment, or the same fertilization equipment from the crop growth control strategy list. For each irrigation equipment, based on the number of farmland spatial units covered by the corresponding irrigation equipment and the required irrigation water depth for each farmland spatial unit, the unit checks the correspondence between the irrigation equipment's output and the irrigation water depth in the equipment control terminal to determine the valve opening duration. For each fertilization equipment, based on the fertilization method and fertilization level in the control rule table, the unit determines the running duration or number of runs in the correspondence between the fertilization equipment's output and the fertilization level. The equipment control terminal generates structured control instructions based on the determined valve opening duration and running duration or number of runs and issues them to the corresponding irrigation and fertilization equipment to execute irrigation and fertilization operations. After the operation is completed, the next acquisition cycle begins, repeating the process from the multimodal farmland data acquisition and crop growth map signal sequence construction steps to form a crop growth perception and control closed loop based on the crop growth Koopman evolution map signal processing model.

[0061] After the growth evaluation table is established, the control execution unit iterates through each record in the table. For a given record, the control execution unit sequentially reads each control rule from the control rule table. Each control rule in the control rule table contains at least the triggering growth level condition, optional soil condition, and a corresponding control action description. For example, one control rule might specify "execute supplemental irrigation when the growth level is weak and soil moisture is below the suitable range," while another control rule might specify "reduce fertilizer application when the growth level is vigorous and soil nutrients are high." The control execution unit compares the growth level of the record in the growth evaluation table with the triggering growth level conditions of each control rule in the control rule table. When the growth level matches, the unit then reads the soil moisture and soil nutrient data of the corresponding farmland spatial unit for that record as needed and compares them with the soil condition recorded in the control rule table. When both the growth level and soil condition meet a certain control rule, the control execution unit reads the preset control action from that control rule, such as "supplemental irrigation," "increase fertilization," or "decrease fertilization."

[0062] When generating control actions, the control execution unit not only determines the action type but also adds time and intensity information to the action. Time information can be a combination of a prediction cycle identifier and a specific execution time period, such as "morning time of the 3rd prediction cycle." Intensity information can be represented in a hierarchical manner, such as "irrigation intensity level 3" or "fertilization intensity level 2." The control execution unit combines the farmland spatial unit identifier, prediction cycle identifier, specific execution time period, action type, and intensity level into control strategy entries and adds these entries to the crop growth control strategy list. This method of "generating control strategy entries by comparing evaluation records with rules" connects growth prediction results with expert control experience through a control rule table, facilitating subsequent adjustments to the control rule table for different crops or regions without altering the implementation logic of the control execution unit.

[0063] After the list of crop growth control strategies is constructed, the control execution unit merges and arranges the strategies. Since different farmland spatial units may correspond to the same irrigation pipeline or the same fertilization equipment, simply executing each control strategy item individually could lead to frequent on / off cycles of the same irrigation or fertilization equipment within a short period, resulting in increased equipment wear and energy consumption. Therefore, the control execution unit first establishes a mapping table between irrigation and fertilization equipment and farmland spatial units based on their coverage relationships. This mapping table can be entered by maintenance personnel during the system installation and commissioning phase. For example, an irrigation device might cover farmland spatial units in row 1, column 1 to row 1, column 10, and a fertilization device might cover farmland spatial units in row 2, column 1 to row 2, column 5.

[0064] The control execution unit aggregates control strategy entries from the crop growth control strategy list for the same time period, the same irrigation equipment, or the same fertilization equipment, based on the mapping table from equipment to farmland spatial units. Taking irrigation equipment as an example, when multiple farmland spatial units are determined to require "supplementary irrigation" within the same prediction period and execution time, the control execution unit counts the number of these farmland spatial units and reads the corresponding irrigation intensity level for each unit. It then converts the irrigation intensity level into a target irrigation water depth, such as level 1 corresponding to a 10 mm water depth, level 2 to a 20 mm water depth, and level 3 to a 30 mm water depth. Finally, based on the area of ​​farmland spatial units covered by the equipment, the target irrigation water depths of each unit are merged into the target average irrigation water depth for that irrigation equipment during that time period. The significance of this is that it transforms the needs of multiple discrete farmland spatial units into a unified need at the single equipment level, reducing repetitive actions of the equipment while meeting the growth control objectives.

[0065] To convert the target average irrigation water depth and fertilization intensity level into control parameters that irrigation and fertilization equipment can directly execute, a control execution unit works in conjunction with the equipment control terminal. During the system configuration phase, the equipment control terminal pre-stores the correspondence between irrigation equipment output and valve opening duration, and the correspondence between fertilization equipment output and runtime or number of runs. These correspondences can be obtained through on-site calibration, for example, by testing the actual water volume sprayed by an irrigation device under fixed water pressure conditions at different valve opening durations, converting the water volume into the water depth corresponding to the coverage area, and recording it in a correspondence table. The control execution unit uses the target average irrigation water depth as a search condition, looking up the corresponding table in the equipment control terminal to obtain the valve opening duration closest to the target average irrigation water depth. In this way, agricultural needs can be directly converted into equipment control parameters using empirical calibration data without complex fluid simulations. The process of determining the runtime or number of runs for the fertilization equipment is similar; the control execution unit looks up the corresponding table based on the fertilization intensity level to obtain the appropriate runtime or number of runs.

[0066] After determining the valve opening duration for irrigation equipment and the running time or number of operations for fertilization equipment, the control execution unit generates structured control instructions for each irrigation and fertilization device. These structured control instructions include equipment identification, execution start time, execution end time or running time, control action type, target farmland spatial unit range, and necessary safety restrictions, such as maximum permissible water pressure and restrictions on spraying during specific time periods. The control execution unit sends the structured control instructions to the corresponding equipment control terminal via wired or wireless communication. The equipment control terminal then automatically performs operations such as valve opening and closing, pump start / stop, or fertilizer operation within the specified time period. After execution, the equipment control terminal can feed back the actual execution results to the control execution unit, such as actual running time and fault alarm information. The control execution unit can choose to record these results in a log for subsequent adjustments to the control rule table or correspondence table.

[0067] The present invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A crop growth perception and control system based on a multi-modal large model, characterized in that, The system comprises: a data acquisition unit configured to acquire crop canopy images, soil sensor data, environmental sensor data, and farming operation information covering the spatial units of the farmland in a plurality of acquisition cycles, and to call a multi-modal large model to analyze the agronomic procedure text data to generate a vigor level threshold table and a control rule table, and to determine crop vigor base values based on the image processing results, soil sensor data, and environmental sensor data of each spatial unit of the farmland and construct a crop vigor map signal sequence; a vigor prediction unit configured to construct a crop vigor Koopman evolution map signal processing model based on the farmland spatial topology map and the crop vigor map signal sequence, and obtain a crop vigor map signal prediction sequence for a plurality of prediction cycles; a control execution unit configured to generate a crop vigor control strategy based on the crop vigor map signal prediction sequence and the vigor level threshold table and the control rule table, and to control the irrigation equipment and the fertilization equipment to perform corresponding operations according to the crop vigor control strategy.

2. The system of claim 1, wherein, The data acquisition unit divides the target farmland into a grid composed of a plurality of spatial units of the farmland, and each spatial unit of the farmland corresponds to a fixed geographic location identifier; In each acquisition cycle, the crop canopy images covering each spatial unit of the farmland are acquired by an image acquisition device; The soil sensor data includes soil moisture data and soil nutrient data; the environmental sensor data includes air temperature, air humidity, and light intensity data.

3. The system of claim 2, wherein, The data acquisition unit calls the multi-modal large model to generate the vigor level threshold table and the control rule table from the agronomic procedure text data as input; performs image processing operations on the crop canopy images of each spatial unit of the farmland, which includes scaling the images to a uniform size, extracting color components, obtaining plant regions by threshold segmentation and counting the number of plant region pixels, calculating the average green component value and canopy coverage ratio of the plant region, and comparing the average green component value and canopy coverage ratio with the corresponding entries in the vigor level threshold table to determine the leaf color grade and canopy coverage grade; matches the leaf color grade, canopy coverage grade, and soil moisture data and soil nutrient data with the records in the vigor level threshold table to obtain the crop vigor base values of the corresponding spatial units of the farmland.

4. The system of claim 3, wherein, The data acquisition unit determines the adjacent relationship according to the relative position relationship of the spatial units of the farmland in the farmland and the connection relationship of the irrigation facilities and the drainage facilities, and records the adjacent relationship between two spatial units of the farmland as existing adjacent relationship when the two spatial units share a boundary in the horizontal or vertical direction, or the irrigation pipeline or drainage facility penetrates the two spatial units; constructs a farmland spatial topology map with all spatial units of the farmland as nodes and spatial unit pairs with existing adjacent relationship as lines; In the same collection cycle, the crop growth basic values of all farmland space units are corresponded to the nodes in the farmland space topology map in a predetermined order to form a crop growth map signal of the corresponding collection cycle. The crop growth map signal construction process is repeated in multiple continuous collection cycles to form a crop growth map signal sequence arranged in time sequence.

5. The system of claim 4, wherein, The growth prediction unit constructs a connection matrix based on the farmland space topology map. When there is a connection between any two farmland space units in the farmland space topology map, a first value is recorded in the corresponding row and column positions of the connection matrix, and when there is no connection, a second value is recorded in the corresponding row and column positions. For each farmland space unit, the sum of all values in the row related to the corresponding farmland space unit in the connection matrix is obtained to obtain the connection quantity, and the connection quantity of each farmland space unit is filled into the diagonal position of the diagonal matrix to form a degree matrix. The degree matrix and the connection matrix are subjected to difference value operation according to the corresponding row and column positions to obtain a graph Laplacian matrix. A numerical calculation program is called to perform eigenvalue decomposition on the graph Laplacian matrix to obtain a set of eigenvectors and eigenvalues corresponding to each eigenvector, and the eigenvectors are sorted in order according to the size of the eigenvalues to form an ordered eigenvector set.

6. The system of claim 5, wherein, The growth prediction unit arranges the crop growth basic values of all farmland space units in a corresponding collection cycle in row and column index order to form a crop growth basic vector for each collection cycle, and performs element-by-element multiplication and addition of the crop growth basic vector and each eigenvector in the ordered eigenvector set and sums the multiplication and addition results to obtain a graph spectrum coefficient on the corresponding eigenvector. In the ordered eigenvector set, low-frequency eigenvector subsets, medium-frequency eigenvector subsets, and high-frequency eigenvector subsets are sequentially divided. For each collection cycle, the graph spectrum coefficients corresponding to the low-frequency eigenvector subsets, the medium-frequency eigenvector subsets, and the high-frequency eigenvector subsets are averaged to obtain low-frequency average graph spectrum coefficients, medium-frequency average graph spectrum coefficients, and high-frequency average graph spectrum coefficients. Each average value is multiplied by the position value corresponding to each farmland space unit in the corresponding eigenvector and accumulated to obtain the low-frequency growth component, the medium-frequency growth component, and the high-frequency growth component of each farmland space unit.

7. The system of claim 6, wherein, The growth prediction unit combines the crop growth basic value, the low-frequency growth component, the medium-frequency growth component, and the high-frequency growth component in a fixed order to form a crop growth extended eigenvector for each farmland space unit in a collection cycle, and arranges the crop growth extended eigenvectors of all farmland space units in the same collection cycle in row and column index order to form a crop growth extended state vector. A crop growth extended state sequence is formed in multiple collection cycles.

8. The system of claim 7, wherein, The crop growth prediction unit selects, in time sequence, crop growth expansion state vectors of two adjacent acquisition periods in a crop growth expansion state sequence, takes the crop growth expansion state vector of an earlier time as a current state vector, and takes the crop growth expansion state vector of a later time as a next state vector; arranges all current state vectors in time sequence as columns to form a current state matrix, and arranges all next state vectors in corresponding time sequence as columns to form a next state matrix; A minimum error solving function in a numerical calculation program is called to take the current state matrix and the next state matrix as inputs, to solve a linear transformation matrix from the current state matrix to the next state matrix, and to take the linear transformation matrix as a crop growth Koopman evolution operator.

Images