Planting management method and system for smart agriculture

Abstract

The invention relates to the technical field of intelligent agriculture, in particular to a planting management method and system for intelligent agriculture, and the method comprises the steps: S1, obtaining multi-modal data through configuring an independent sensing module and a local control execution end for each micro-block; s2, recognizing a current crop variety in real time according to the multi-modal data by adopting a low-power-consumption edge computing device in combination with image recognition; when the system is used, each micro-block operates independently, maintenance of each plant can be refined, multi-source sensing data can be fused, the problems of data lag and weight deviation among different sensors are solved, full-automatic feedback regulation and control are achieved, the labor burden of workers is reduced, low-power-consumption and high-real-time recognition is facilitated, multi-source data such as images and fusion of temperature and humidity are achieved, and the system is suitable for popularization and application. According to the method, the identification accuracy is improved, the problem of crop confusion which is like a long variety can be solved conveniently, accurate management in multi-variety mixed cropping planting can be achieved conveniently, and the accuracy of fine management and control of different demands of multiple different crops in a small range can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart agriculture, and in particular to a planting management method and system for smart agriculture. Background Art

[0002] Smart agricultural planting management is a new agricultural management model that uses modern information technology, the Internet of Things (IoT), big data, artificial intelligence (AI) and other advanced technologies to comprehensively monitor and fine-tune the agricultural production process.

[0003] The patent publication number is CN118466647A, which states in its specification that "the present invention discloses a planting management method and system for smart agriculture, which relates to the field of planting management technology. The present invention includes greenhouse information monitoring, temperature prediction analysis, equipment temperature testing, equipment demand analysis and result display. By analyzing the heat transfer rate corresponding to each monitoring point in each sub-area, the predicted temperature corresponding to each monitoring point in each sub-area is predicted, and the maximum temperature operating limit corresponding to the hot fan is analyzed, and then compared with the predicted temperature corresponding to each monitoring point in each sub-area, the equipment demand corresponding to each target sub-area is obtained, thereby realizing the reasonable installation of equipment in the greenhouse and ensuring the crop While these technologies achieve precision agricultural management through intelligent temperature monitoring and equipment control, primarily involving zone division and calculation of heat transfer rates, temperature prediction at each monitoring point based on environmental data, evaluation of fan performance, selection of optimal upper operating temperature limits, temperature comparison, calculation of equipment placement and demand, and visualization of fan configuration plans, they struggle to achieve precise management when dealing with mixed cropping. Traditional smart agriculture platforms are more suitable for field crops or single-crop cultivation scenarios. However, in today's urban and facility agriculture, a growing number of farms are adopting mixed cropping (e.g., strawberry + vanilla, tomato + cucumber, etc.). However, existing systems typically manage by "zone" or "single crop," making it difficult to precisely manage the varying needs (temperature, light, water, fertilizer, etc.) of multiple crops within a small area.

[0004] To sum up, developing a planting management method and system for smart agriculture is still a key issue that needs to be urgently addressed in the field of smart agricultural technology. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem in the prior art that although the above-mentioned technology realizes precision agricultural management through intelligent temperature monitoring and equipment control, it mainly involves dividing the area and calculating the heat transfer rate, predicting the temperature of each monitoring point based on environmental data, evaluating the performance of the hot fan, screening the upper limit of the optimal operating temperature, comparing temperature data, calculating the equipment placement direction and demand, and visually displaying the hot fan configuration plan, the above-mentioned technology is difficult to achieve precise management when facing multi-variety mixed planting. Traditional smart agricultural platforms are more suitable for field crops or single crop planting scenarios, and in today's urban agriculture and facility agriculture, more and more plantations are showing a trend of multi-variety mixed planting (such as strawberry + vanilla, tomato + cucumber, etc.). However, the existing system is usually managed by "region" or "single crop", and it is difficult to finely control the different needs (temperature, light, water and fertilizer, etc.) of multiple different crops in a small range. The present invention provides a planting management method and system for smart agriculture.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention provides a planting management method for smart agriculture, comprising: S1, acquiring multimodal data by configuring an independent perception module and a local control execution terminal for each micro-block;

[0008] S2. Using a low-power edge computing device combined with image recognition to identify the current crop type in real time based on the multimodal data;

[0009] S3. The local terminal adjusts the microenvironment based on the identified crop type by retrieving its phased control parameters;

[0010] S4. Based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed to dynamically determine the crop stage.

[0011] Furthermore, in step S1, by configuring an independent perception module and a local control execution terminal for each micro-block, a method for acquiring multimodal data is as follows:

[0012] The micro-block is equipped with an independent sensing module to obtain environmental temperature and humidity, light, and soil parameters. The environmental temperature and humidity are collected by an infrared thermal array, and the light is measured by a multi-channel spectroscopic reflectometer. The soil parameters represent volumetric moisture content, electrical conductivity, pH, and nutrient concentration. An all-in-one soil probe is used to continuously monitor soil parameters. A weighted time synchronization fusion model is introduced to fuse the environmental temperature and humidity, light, and soil parameters to obtain multimodal data x i (t), expression formula: where x c (t) is the multimodal data of micro-block c at time t, M is the total number of perceptual channels, qj (t) is the dynamic weight coefficient of the jth mode, is the data of the jth mode after the lag time, α cj is the time delay of the jth mode on the micro-block c relative to the unified time t. The local control execution end includes but is not limited to sprinkler irrigation, fill light, and fan. The controller of the local control execution end generates control instructions through a nonlinear state-behavior mapping function. The control instructions control the sprinkler irrigation, fill light, and fan to operate respectively.

[0013] Furthermore, in step S2, a method for real-time identification of the current crop type based on the multimodal data using a low-power edge computing device combined with image recognition is as follows:

[0014] The low-power edge computing device obtains image frames from the camera at every moment, maps them into high-dimensional image feature vectors through the visual encoder, and at the same time obtains the multimodal data to construct a joint semantic embedding space: Q m (t), in the joint semantic embedding space, crop species recognition is regarded as a multi-category image-environment joint classification task. The image recognition calculates the posterior probability of crop species through a multimodal Bayesian inference framework, which is expressed as: Where W(w k |Q m (t)) represents the given feature Q m (t) belongs to the wth k Class probability, Q m (t) Joint semantic embedding space, β k represents the mean vector of category k, ∑ k represents the covariance matrix of category k, represents the inverse of the covariance matrix of category k, β u represents the mean vector of category u, represents the covariance matrix of category u, It means to sum up the Gaussian likelihood values of all categories, K represents the number of all categories, and the recognition result is: where Q m (t) Joint semantic embedding space, represents the crop prediction category corresponding to the micro-block c at time t, w c Indicates that category c is an element in the set E, W(w k |Q m (t)) represents the given joint semantic embedding space Q at time t m (t) Micro-block c belongs to category w k The probability of E represents the set of all possible categories, which is the set of all crop types and growth stages.

[0015] Furthermore, in step S3, the local end uses the identified crop type and retrieves its phased control parameters to adjust the microenvironment in the following manner:

[0016] The crop species can be divided into multiple key growth stages. The multiple key growth stages include but are not limited to germination, jointing, heading, and filling. The hidden variables of the growth stages are defined and time series prediction is performed using a hidden Markov model (HMM). The expression formula is: Where O(p i (t)|a 1:t ) is the total observation sequence a from time 1 to t given at time t 1:t Under the premise that the crop state is p i (t), A is the normalization factor, O(a t |p i (t)) indicates that if the crop is in state p i (t) Then we observe the current observation a at time t t the possibility of Indicates that a weighted sum is to be made for all possible previous states, O(p i (t)|p i (t-1)) indicates that the crop changes from state p i (t-1) transfer to the current state p i The probability of (t), O(p i (t-1)|a 1:t-1 ) indicates that the crop is in state p at the previous time t-1 given the previous t-1 observation sequence i The probability of (t-1), a t is the observation data from images, environment and sensors, a 1:t is the sequence of all observations from time 1 to time t, p i (t) represents the state of the i-th possible crop growth stage at time t, and outputs the current inferred crop type stage: in is the predicted crop growth stage of the i-th micro-block c at time t, It means that among all possible crop growth stages p, the posterior probability O(p|a 1:t ) is the largest stage, O(p|a 1:t ) is a given observation sequence a1, a2, ..., a at time t t The probability that the crop is in growth stage p under the premise of 1:t is the sequence of all observations from time 1 to time t.

[0017] Furthermore, in step S3, the local end uses the identified crop type and retrieves its phased control parameters to adjust the microenvironment in the following manner:

[0018] The crop type stage corresponds to a set of most suitable environmental parameter intervals, and a mapping table is stored in the form of a database. After looking up the table, the corresponding control target vector is dynamically loaded. where g i (t) represents the environmental control target parameter vector for the i-th micro-block c at time t, It represents the optimal environmental temperature and humidity value that needs to be maintained in micro-block c for the current crop type and stage. Indicates the target light intensity required for the current crop growth stage. represents the soil parameter control item, Indicates fan control items, [·] ⊥ is the transpose symbol, Represents vector g i (t) is a crop class identified by and growth stage The control parameter space of the decision is extracted, is the predicted crop growth stage of the i-th micro-block c at time t, Represents the crop prediction category corresponding to the micro-block c at time t, and then the multimodal data x of the actual environmental state is i (t) and the control target vector g i (t) to compare the deviation and generate the control input sequence in the future Δt period, expressed as: in is the optimization operation, x i (χ) is the system state vector of the i-th micro-block c at time χ, I i (χ) is the expected environmental state corresponding to the current time χ, represents the weighted square difference between the current state and the target state, d i (χ) is the control amount applied to the micro-block c at time χ, is the weighted square of the control input, It is the overall optimization of all states and control inputs for the next P steps from the current time t, (st)x i (χ+1)=f{x i (χ),d i (χ)} describes how to i (χ) and control input d i(χ) Deducing the next state, the local end actuator adjusts the local control execution end in real time after receiving it, executes the work of sprinkler irrigation, fill light, and fan, and periodically updates the state feedback to trigger incremental re-optimization.

[0019] Furthermore, in step S4, based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed. The method for dynamically determining the crop stage is as follows:

[0020] The multimodal data includes but is not limited to a continuously acquired image modality sequence, an environmental modality vector sequence, and crop types. The multimodal data are aligned and constructed into a fusion tensor: in is the temporal feature input obtained by fusion of the image and sensor of the i-th micro-block c at time t, is a real matrix space indicating that is a size of ε×(Z q′ +Z e′ ), ε represents the time length, Z q′ is the image feature dimension, Z e′ is the environmental feature dimension, Concat[·] means concatenating two feature matrices into a whole temporal input in the feature dimension, CNN(q c ′(·)) for the continuous images q of micro-block c c ′(·) is input to apply convolutional neural network (CNN), is the continuous environmental data of the i-th micro-block c. The image feature extraction of the image modality sequence adopts a multi-scale convolutional feature encoder. Combined with the growth model, a neural differential system is introduced to capture the nonlinear time-varying feature evolution process. The hidden state at the current moment is obtained by ODE-Solver, and a local environment graph is constructed at the same time. Each node of the local environment graph is a local micro-block. The hidden state of the local micro-block is updated by graph convolution, and the expression formula is: in is the hidden state of the i′th node in the l+1th layer at time t, φ(·) is the GELU activation function, It is a summation operation for each node j′ in the set of all neighbor nodes N(i′) belonging to node j′. Indicates that at time t, the importance of node j' to node i' satisfies H l is the learnable linear transformation matrix of the lth layer, Represent the feature representation of the j′th adjacent node in the lth layer at time t, and build a "crop growth stage automatic identification algorithm system" driven by time series.

[0021] Furthermore, in step S4, based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed. The method for dynamically determining the crop stage is as follows:

[0022] Based on the hidden state, a multi-class classifier is used to output the current stage of the crop. Then, the stage of the crop is further determined, which is expressed as:

[0023] in Represents the current temporal fusion feature of the known micro-block c and crop type prediction results Under the premise of the crop's growth stage is the probability of the k′th stage in the m′th class, and softmax(·) maps the output vector to a probability distribution. It is the feature vector of micro-block c at time t that integrates image, temperature, humidity, lighting and other data. is the crop type prediction result, Targeted at specific crop types The designed linear classifier weight matrix, For crop types The bias vector defined is is the result of the growth phase determination of the i-th micro-block c at time t, It means finding the stage number m′ that makes the subsequent probability value the largest. Indicates that the current fusion feature of the given block Under the premise that the probability value of the crop belongs to the k'th stage under the m'th category, is the probability of the k′th stage in the m′th category. The “crop growth stage automatic identification algorithm system” introduces a confidence dynamic smoothing mechanism to perform time series dynamic correction and adaptive confidence filtering on the stage of the crop.

[0024] On the other hand, the present invention also provides a smart agriculture planting management system, comprising:

[0025] The distributed micro-block management module acquires multimodal data by configuring an independent perception module and a local control execution terminal for each micro-block;

[0026] The automatic crop identification module uses low-power edge computing devices combined with image recognition to identify the current crop type in real time based on the multimodal data;

[0027] The strategy dynamic matching module, based on the identified crop type, retrieves its stage-specific control parameters to adjust the microenvironment and dynamically determines the crop stage;

[0028] The automatic identification module, based on the multimodal data, combines the crop types and growth models to build a time series-driven "crop growth stage automatic identification algorithm system" to dynamically determine the crop stage.

[0029] The operation process of the distributed micro-block management module includes:

[0030] The micro-block is equipped with an independent sensing module to obtain environmental temperature and humidity, light, and soil parameters. The environmental temperature and humidity are collected by an infrared thermal array, and the light is measured by a multi-channel spectroscopic reflectometer. The soil parameters represent volumetric moisture content, electrical conductivity, pH, and nutrient concentration. An all-in-one soil probe is used to continuously monitor soil parameters. A weighted time synchronization fusion model is introduced to fuse the environmental temperature and humidity, light, and soil parameters to obtain multimodal data x i (t), expression formula: where x c (t) is the multimodal data of micro-block c at time t, M is the total number of perceptual channels, q j (t) is the dynamic weight coefficient of the jth mode, is the data of the jth mode after the lag time, ɑ cj is the time delay of the jth mode on the micro-block c relative to the unified time t. The local control execution end includes but is not limited to sprinkler irrigation, fill light, and fan. The controller of the local control execution end generates control instructions through a nonlinear state-behavior mapping function. The control instructions control the sprinkler irrigation, fill light, and fan to operate respectively;

[0031] The operation process of the automatic crop identification module includes:

[0032] The low-power edge computing device obtains image frames from the camera at every moment, maps them into high-dimensional image feature vectors through the visual encoder, and at the same time obtains the multimodal data to construct a joint semantic embedding space: Q m (t), in the joint semantic embedding space, crop species recognition is regarded as a multi-category image-environment joint classification task. The image recognition calculates the posterior probability of crop species through a multimodal Bayesian inference framework, which is expressed as: Where W(w k |Q m (t)) represents the given feature Q m (t) belongs to the wth k Class probability, Q m (t) Joint semantic embedding space, β k represents the mean vector of category k, ∑ k represents the covariance matrix of category k, represents the inverse of the covariance matrix of category k, β urepresents the mean vector of category u, represents the covariance matrix of category u, It means to sum up the Gaussian likelihood values of all categories, K represents the number of all categories, and the recognition result is: where Q m (t) Joint semantic embedding space, represents the crop prediction category corresponding to the micro-block c at time t, w c Indicates that category c is an element in the set E, W(w k |Q m (t)) represents the given joint semantic embedding space Q at time t m (t) Micro-block c belongs to category w k The probability of E represents the set of all possible categories, which is the set of all crop types and growth stages.

[0033] The operation process of the strategy dynamic matching module includes:

[0034] The crop species can be divided into multiple key growth stages. The multiple key growth stages include but are not limited to germination, jointing, heading, and filling. The hidden variables of the growth stages are defined and time series prediction is performed using a hidden Markov model (HMM). The expression formula is: Where O(p i (t)|a 1:t ) is the total observation sequence a from time 1 to t given at time t 1:t Under the premise that the crop state is p i (t), A is the normalization factor, O(a t |p i (t)) indicates that if the crop is in state p i (t) Then we observe the current observation a at time t t the possibility of Indicates that a weighted sum is to be made for all possible previous states, O(p i (t)|p i (t-1)) indicates that the crop changes from state p i (t-1) transfer to the current state p i The probability of (t), O(p i (t-1)|a 1:t-1 ) indicates that the crop is in state p at the previous time t-1 given the previous t-1 observation sequence i The probability of (t-1), a t is the observation data from images, environment and sensors, a 1:t is the sequence of all observations from time 1 to time t, p i(t) represents the state of the i-th possible crop growth stage at time t, and outputs the current inferred crop type stage: in is the predicted crop growth stage of the i-th micro-block c at time t, It means that among all possible crop growth stages p, the posterior probability O(p|a 1:t ) is the largest stage, O(p|a 1:t ) is a given observation sequence a1, a2, ..., a at time t t The probability that the crop is in growth stage p under the premise of 1:t is the sequence of all observations from time 1 to time t.

[0035] The crop type stage corresponds to a set of most suitable environmental parameter intervals, and a mapping table is stored in the form of a database. After looking up the table, the corresponding control target vector is dynamically loaded. where g i (t) represents the environmental control target parameter vector for the i-th micro-block c at time t, It represents the optimal environmental temperature and humidity value that needs to be maintained in micro-block c for the current crop type and stage. Indicates the target light intensity required for the current crop growth stage. represents the soil parameter control item, Indicates fan control items, [·] ⊥ is the transpose symbol, Represents vector g i (t) is a crop class identified by and growth stage The control parameter space of the decision is extracted, is the predicted crop growth stage of the i-th micro-block c at time t, Represents the crop prediction category corresponding to the micro-block c at time t, and then the multimodal data x of the actual environmental state is i (t) and the control target vector g i (t) to compare the deviation and generate the control input sequence in the future Δt period, expressed as: in is the optimization operation, x i (χ) is the system state vector of the i-th micro-block c at time χ, I i (χ) is the expected environmental state corresponding to the current time χ, represents the weighted square difference between the current state and the target state, d i (χ) is the control amount applied to the micro-block c at time χ, is the weighted square of the control input, It is the overall optimization of all states and control inputs for the next P steps from the current time t, (st)x i (χ+1)=f{x i (χ),d i (χ)} describes how to i (χ) and control input d i (x) Deducing the next state, the local end actuator adjusts the local control execution end in real time after receiving it, executes the sprinkler irrigation, fill light, and fan operations, and periodically updates the state feedback to trigger incremental re-optimization;

[0036] The operation process of the automatic identification module includes:

[0037] The multimodal data includes but is not limited to a continuously acquired image modality sequence, an environmental modality vector sequence, and crop types. The multimodal data are aligned and constructed into a fusion tensor: in is the temporal feature input obtained by fusion of the image and sensor of the i-th micro-block c at time t, is a real matrix space indicating that is a size of ε×(Z q′ +Z e′ ), ε represents the time length, Z q′ is the image feature dimension, Z e′ is the environmental feature dimension, Concat[·] means concatenating two feature matrices into a whole temporal input in the feature dimension, CNN(q c ′(·)) for the continuous images q of micro-block c c ′(·) is input to apply convolutional neural network (CNN), is the continuous environmental data of the i-th micro-block c. The image feature extraction of the image modality sequence adopts a multi-scale convolutional feature encoder. Combined with the growth model, a neural differential system is introduced to capture the nonlinear time-varying feature evolution process. The hidden state at the current moment is obtained by ODE-Solver, and a local environment graph is constructed at the same time. Each node of the local environment graph is a local micro-block. The hidden state of the local micro-block is updated by graph convolution, and the expression formula is: in is the hidden state of the i′th node in the l+1th layer at time t, φ(·) is the GELU activation function, It is a summation operation for each node j′ in the set of all neighbor nodes N(i′) belonging to node j′. Indicates that at time t, the importance of node j' to node i' satisfies H l is the learnable linear transformation matrix of the lth layer, Represent the feature representation of the j′th adjacent node in the lth layer at time t, and build a "crop growth stage automatic identification algorithm system" driven by time series.

[0038] Based on the hidden state, a multi-class classifier is used to output the current stage of the crop. Then, the stage of the crop is further determined, which is expressed as:

[0039] in Represents the current temporal fusion feature of the known micro-block c and crop type prediction results Under the premise of the crop's growth stage is the probability of the k′th stage in the m′th class, and softmax(·) maps the output vector to a probability distribution. It is the feature vector of micro-block c at time t that integrates image, temperature, humidity, lighting and other data. is the crop type prediction result, Targeted at specific crop types The designed linear classifier weight matrix, For crop types The bias vector defined is is the result of the growth phase determination of the i-th micro-block c at time t, It means finding the stage number m′ that makes the subsequent probability value the largest. Indicates that the current fusion feature of the given block Under the premise that the probability value of the crop belongs to the k'th stage under the m'th category, is the probability of the k′th stage in the m′th category. The “crop growth stage automatic identification algorithm system” introduces a confidence dynamic smoothing mechanism to perform time series dynamic correction and adaptive confidence filtering on the stage of the crop.

[0040] Beneficial effects

[0041] Compared with the known public technology, the technical solution provided by the present invention has the following advantages:

[0042] Beneficial effects:

[0043] When the present invention is in use, each micro-block operates independently, can be refined to the maintenance of each plant, integrate multi-source perception data, solve the problems of data lag and weight deviation between different sensors, and automatically feedback control to reduce the burden of manual labor, which is conducive to low-power and high-real-time recognition. By imaging and integrating multi-source data such as temperature and humidity, the recognition accuracy is improved, which is convenient for solving the confusion problem of "similar" crops. Through time series judgment and personalized control, it is conducive to accurately matching the current growth needs of crops, avoiding waste of resources, and achieving the purpose of modeling different crops and different stages. It no longer relies on manual judgment, is conducive to automatic adaptation to the optimal environment, and facilitates the realization of precise management in multi-variety mixed planting, which is conducive to improving the accuracy of refined management and control of the different needs of multiple different crops in a small range.

[0044] When in use, the present invention significantly improves the accuracy of growth stage identification based on joint modeling of time series and space, and introduces a confidence filtering mechanism to effectively avoid stage identification misjudgment caused by single data fluctuations, avoids inaccurate matching of growth stage identification and management rhythm, and improves the accuracy of growth stage judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a planting management method for smart agriculture of the present invention;

[0046] Figure 2 This is a system diagram of a smart agriculture planting management system according to the present invention. DETAILED DESCRIPTION

[0047] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0048] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0049] The present invention is described in further detail below with reference to the accompanying drawings:

[0050] Example 1:

[0051] like Figure 1 As shown, the present invention provides a planting management method for smart agriculture, comprising:

[0052] S1, acquire multimodal data by configuring an independent perception module and a local control execution terminal for each micro-block;

[0053] Furthermore, in step S1, by configuring an independent perception module and a local control execution terminal for each micro-block, a method for acquiring multimodal data is as follows:

[0054] The micro-block is equipped with an independent sensing module to obtain environmental temperature and humidity, light, and soil parameters. The environmental temperature and humidity are collected by an infrared thermal array, and the light is measured by a multi-channel spectroscopic reflectometer. The soil parameters represent volumetric moisture content, electrical conductivity, pH, and nutrient concentration. An all-in-one soil probe is used to continuously monitor soil parameters. A weighted time synchronization fusion model is introduced to fuse the environmental temperature and humidity, light, and soil parameters to obtain multimodal data x i (t), expression formula: where x c (t) is the multimodal data of micro-block c at time t, M is the total number of perceptual channels, q j (t) is the dynamic weight coefficient of the jth mode, is the data of the jth mode after the lag time, α cj is the time delay of the jth mode on the micro-block c relative to the unified time t. The local control execution end includes but is not limited to sprinkler irrigation, fill light, and fan. The controller of the local control execution end generates control instructions through a nonlinear state-behavior mapping function. The control instructions control the sprinkler irrigation, fill light, and fan to operate respectively;

[0055] Specifically, at Smart Agriculture Demonstration Base A, a micro-block unit is arranged every 10 square meters, each equipped with a sensor probe and sprinkler control equipment. For example, if the sensor system in micro-block No. 12 detects insufficient light and low soil moisture, the method automatically issues control commands to "turn on the fill light" and "start the small drip irrigation equipment" after calculation, without the need for human intervention. This method can also dynamically adjust the amount of water and light supplied according to the different growth stages of crops, facilitating precise perception and control based on "local conditions." Each micro-block operates independently, enabling detailed maintenance of each plant, integrating multi-source perception data to address data lag and weight bias between different sensors, and providing fully automatic feedback control, reducing the burden of manual labor.

[0056] S2. Using a low-power edge computing device combined with image recognition to identify the current crop type in real time based on the multimodal data;

[0057] Furthermore, in step S2, a method for real-time identification of the current crop type based on the multimodal data using a low-power edge computing device combined with image recognition is as follows:

[0058] The low-power edge computing device obtains image frames from the camera at every moment, maps them into high-dimensional image feature vectors through the visual encoder, and at the same time obtains the multimodal data to construct a joint semantic embedding space: Q m (t), in the joint semantic embedding space, crop species recognition is regarded as a multi-category image-environment joint classification task. The image recognition calculates the posterior probability of crop species through a multimodal Bayesian inference framework, which is expressed as: Where W(w k |Q m (t)) represents the given feature Q m (t) belongs to the wth k Class probability, Q m (t) Joint semantic embedding space, β k represents the mean vector of category k, ∑ k represents the covariance matrix of category k, represents the inverse of the covariance matrix of category k, β u represents the mean vector of category u, represents the covariance matrix of category u, It means to sum up the Gaussian likelihood values of all categories, K represents the number of all categories, and the recognition result is: where Q m (t) Joint semantic embedding space, represents the crop prediction category corresponding to the micro-block c at time t, w c Indicates that category c is an element in the set E, W(w k |Q m(t)) represents the given joint semantic embedding space Q at time t m (t) Micro-block c belongs to category w k The probability of E represents the set of all possible categories, which is the set of all crop types and growth stages;

[0059] Specifically, tomatoes, chili peppers, and cucumbers were grown simultaneously in the smart greenhouse of Smart Agriculture Demonstration Base A. Multiple low-power edge nodes were deployed across different micro-blocks. Cameras captured images of plant leaves, stems, and fruits, combining these with local temperature and humidity data. This method identified that the current image features and environmental parameters of micro-block No. 3 more closely matched the joint embedding distribution of the "tomato" category, thus determining that the crop planted in this block was tomato. Requiring no cloud computing or manual labeling, the entire process is completed at the edge, facilitating low-power, high-real-time recognition. By integrating images with multi-source data such as temperature and humidity, recognition accuracy is improved, making it easier to resolve the confusion between similar-looking crops.

[0060] S3. The local terminal adjusts the microenvironment based on the identified crop type by retrieving its phased control parameters;

[0061] Furthermore, in step S3, the local end uses the identified crop type and retrieves its phased control parameters to adjust the microenvironment in the following manner:

[0062] The crop species can be divided into multiple key growth stages. The multiple key growth stages include but are not limited to germination, jointing, heading, and filling. The hidden variables of the growth stages are defined and time series prediction is performed using a hidden Markov model (HMM). The expression formula is: Where O(p i (t)|a 1:t ) is the total observation sequence a from time 1 to t given at time t 1:t Under the premise that the crop state is p i (t), A is the normalization factor, O(a t |p i (t)) indicates that if the crop is in state p i (t) Then we observe the current observation a at time t t the possibility of Indicates that a weighted sum is to be made for all possible previous states, O(p i (t)|p i (t-1)) indicates that the crop changes from state p i (t-1) transfer to the current state p i The probability of (t), O(p i (t-1)|a 1:t-1) indicates that the crop is in state p at the previous time t-1 given the previous t-1 observation sequence i The probability of (t-1), a t is the observation data from images, environment and sensors, a 1:t is the sequence of all observations from time 1 to time t, p i (t) represents the state of the i-th possible crop growth stage at time t, and outputs the current inferred crop type stage: in is the predicted crop growth stage of the i-th micro-block c at time t, It means that among all possible crop growth stages p, the posterior probability O(p|a 1:t ) is the largest stage, O(p|a 1:t ) is a given observation sequence a1, a2, ..., a at time t t The probability that the crop is in growth stage p under the premise of 1:t is the sequence of all observations from time 1 to time t.

[0063] Furthermore, in step S3, the local end uses the identified crop type and retrieves its phased control parameters to adjust the microenvironment in the following manner:

[0064] The crop type stage corresponds to a set of most suitable environmental parameter intervals, and a mapping table is stored in the form of a database. After looking up the table, the corresponding control target vector is dynamically loaded. where g i (t) represents the environmental control target parameter vector for the i-th micro-block c at time t, It represents the optimal environmental temperature and humidity value that needs to be maintained in micro-block c for the current crop type and stage. Indicates the target light intensity required for the current crop growth stage. represents the soil parameter control item, Indicates fan control items, [·] ⊥ is the transpose symbol, Represents vector g i (t) is a crop class identified by and growth stage The control parameter space of the decision is extracted, is the predicted crop growth stage of the i-th micro-block c at time t, Represents the crop prediction category corresponding to the micro-block c at time t, and then the multimodal data x of the actual environmental state is i (t) and the control target vector g i (t) to compare the deviation and generate the control input sequence in the future Δt period, expressed as: in is the optimization operation, x i (χ) is the system state vector of the i-th micro-block c at time χ, I i (χ) is the expected environmental state corresponding to the current time χ, represents the weighted square difference between the current state and the target state, d i (χ) is the control amount applied to the micro-block c at time χ, is the weighted square of the control input, It is the overall optimization of all states and control inputs for the next P steps from the current time t, (st)x i (χ+1)=f{x i (χ),d i (χ)} describes how to i (χ) and control input d i (x) Deducing the next state, the local end actuator adjusts the local control execution end in real time after receiving it, executes the sprinkler irrigation, fill light, and fan operations, and periodically updates the state feedback to trigger incremental re-optimization;

[0065] Specifically, using wheat grown in a greenhouse in northern China as an example, this method uses image recognition and sensor data to determine that the wheat has entered the jointing stage. It then retrieves the optimal environmental parameters for "jointing stage-wheat" from a database: for example, daytime temperatures between 18 and 22°C, humidity around 55%, and slightly damp soil moisture. If the temperature in this micro-block is detected to be slightly elevated (25°C) and humidity slightly low (45%), control instructions are issued: fans are activated for cooling, sprinkler irrigation is activated for humidification, and the brightness of the fill lights is reduced to prevent excessive evaporation. Simultaneously, an optimization algorithm controls fan operation time and water volume, ensuring that each adjustment balances energy conservation and target achievement. Further fine-tuning is then carried out in the next cycle based on new data, forming a closed loop. This facilitates precise matching of crop growth needs through time-series judgment and personalized control, avoiding resource waste and achieving the goal of modeling different crops and stages separately. This eliminates reliance on manual judgment and facilitates automatic adaptation to the optimal environment.

[0066] S4. Based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed to dynamically determine the crop stage;

[0067] Furthermore, in step S4, based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed. The method for dynamically determining the crop stage is as follows:

[0068] The multimodal data includes but is not limited to a continuously acquired image modality sequence, an environmental modality vector sequence, and crop types. The multimodal data are aligned and constructed into a fusion tensor: in is the temporal feature input obtained by fusion of the image and sensor of the i-th micro-block c at time t, is a real matrix space indicating that is a size of ε×(Z q′ +Z e′ ), ε represents the time length, Z q′ is the image feature dimension, Z e′ is the environmental feature dimension, Concat[·] means concatenating two feature matrices into a whole temporal input in the feature dimension, CNN(q c ′(·)) for the continuous images q of micro-block c c ′(·) is input to apply convolutional neural network (CNN), is the continuous environmental data of the i-th micro-block c. The image feature extraction of the image modality sequence adopts a multi-scale convolutional feature encoder. Combined with the growth model, a neural differential system is introduced to capture the nonlinear time-varying feature evolution process. The hidden state at the current moment is obtained by ODE-Solver, and a local environment graph is constructed at the same time. Each node of the local environment graph is a local micro-block. The hidden state of the local micro-block is updated by graph convolution, and the expression formula is: in is the hidden state of the i′th node in the l+1th layer at time t, φ(·) is the GELU activation function, It is a summation operation for each node j′ in the set of all neighbor nodes N(i′) belonging to node j′. Indicates that at time t, the importance of node j' to node i' satisfies H l is the learnable linear transformation matrix of the lth layer, Represent the feature representation of the j′th adjacent node in the lth layer at time t, and build a "crop growth stage automatic identification algorithm system" driven by time series.

[0069] Furthermore, in step S4, based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed. The method for dynamically determining the crop stage is as follows:

[0070] Based on the hidden state, a multi-class classifier is used to output the current stage of the crop. Then, the stage of the crop is further determined, which is expressed as:

[0071] in Represents the current temporal fusion feature of the known micro-block c and crop type prediction results Under the premise of the crop's growth stage is the probability of the k′th stage in the m′th class, and softmax(·) maps the output vector to a probability distribution. It is the feature vector of micro-block c at time t that integrates image, temperature, humidity, lighting and other data. is the crop type prediction result, Targeted at specific crop types The designed linear classifier weight matrix, For crop types The bias vector defined is is the result of the growth phase determination of the i-th micro-block c at time t, It means finding the stage number m′ that makes the subsequent probability value the largest. Indicates that the current fusion feature of the given block Under the premise that the probability value of the crop belongs to the k'th stage under the m'th category, is the probability of the k′th stage in the m′th category. The “crop growth stage automatic identification algorithm system” introduces a confidence dynamic smoothing mechanism to perform time series dynamic correction and adaptive confidence filtering on the crop stage.

[0072] Specifically, in daily production, inaccurate matching between crop growth stage identification and management rhythms can occur. In smart farming, many automated control and early warning strategies rely on manually set "growth stages," such as the seedling stage, vegetative growth stage, and flowering and fruiting stage. However, the actual time when crops enter a particular growth stage fluctuates, affected by weather, variety, and operating habits. As a result, existing systems rely primarily on empirical estimates or time-driven methods, which are not very accurate. However, using smart greenhouse tomato cultivation as an example, this method and system uses cameras and temperature and humidity sensors deployed in each small block to regularly collect images and environmental data. The system determines whether the "flower bud differentiation period" has begun based on changes in image color and leaf structure. Combined with the trend of ambient temperature changes, it further determines whether the appropriate growth stage has been reached. Based on joint temporal and spatial modeling, the accuracy of growth stage identification is significantly improved. The introduction of a confidence filtering mechanism effectively avoids stage identification errors caused by single-time data fluctuations, prevents inaccurate matching between growth stage identification and management rhythms, and improves the accuracy of growth stage identification.

[0073] Example 2:

[0074] like Figure 2 As shown, embodiment 2 provides a planting management system for smart agriculture, including:

[0075] The distributed micro-block management module acquires multimodal data by configuring an independent perception module and a local control execution terminal for each micro-block;

[0076] The automatic crop identification module uses low-power edge computing devices combined with image recognition to identify the current crop type in real time based on the multimodal data;

[0077] The strategy dynamic matching module, based on the identified crop type, retrieves its stage-specific control parameters to adjust the microenvironment and dynamically determines the crop stage;

[0078] The automatic identification module, based on the multimodal data, combines the crop types and growth models to build a time series-driven "crop growth stage automatic identification algorithm system" to dynamically determine the crop stage.

[0079] The operation process of the distributed micro-block management module includes:

[0080] The micro-block is equipped with an independent sensing module to obtain environmental temperature and humidity, light, and soil parameters. The environmental temperature and humidity are collected by an infrared thermal array, and the light is measured by a multi-channel spectroscopic reflectometer. The soil parameters represent volumetric moisture content, electrical conductivity, pH, and nutrient concentration. An all-in-one soil probe is used to continuously monitor soil parameters. A weighted time synchronization fusion model is introduced to fuse the environmental temperature and humidity, light, and soil parameters to obtain multimodal data x i (t), expression formula: where x c (t) is the multimodal data of micro-block c at time t, M is the total number of perceptual channels, q j (t) is the dynamic weight coefficient of the jth mode, is the data of the jth mode after the lag time, α cj is the time delay of the jth mode on the micro-block c relative to the unified time t. The local control execution end includes but is not limited to sprinkler irrigation, fill light, and fan. The controller of the local control execution end generates control instructions through a nonlinear state-behavior mapping function. The control instructions control the sprinkler irrigation, fill light, and fan to operate respectively;

[0081] The operation process of the automatic crop identification module includes:

[0082] The low-power edge computing device obtains image frames from the camera at every moment, maps them into high-dimensional image feature vectors through the visual encoder, and at the same time obtains the multimodal data to construct a joint semantic embedding space: Q m(t), in the joint semantic embedding space, crop species recognition is regarded as a multi-category image-environment joint classification task. The image recognition calculates the posterior probability of crop species through a multimodal Bayesian inference framework, which is expressed as: Where W(w k |Q m (t)) represents the given feature Q m (t) belongs to the wth k Class probability, Q m (t) Joint semantic embedding space, β k represents the mean vector of category k, ∑ k represents the covariance matrix of category k, represents the inverse of the covariance matrix of category k, β u represents the mean vector of category u, represents the covariance matrix of category u, It means to sum up the Gaussian likelihood values of all categories, K represents the number of all categories, and the recognition result is: where Q m (t) Joint semantic embedding space, represents the crop prediction category corresponding to the micro-block c at time t, w c Indicates that category c is an element in the set E, W(w k |Q m (t)) represents the given joint semantic embedding space Q at time t m (t) Micro-block c belongs to category w k The probability of E represents the set of all possible categories, which is the set of all crop types and growth stages.

[0083] The operation process of the strategy dynamic matching module includes:

[0084] The crop species can be divided into multiple key growth stages. The multiple key growth stages include but are not limited to germination, jointing, heading, and filling. The hidden variables of the growth stages are defined and time series prediction is performed using a hidden Markov model (HMM). The expression formula is: Where O(p i (t)|a 1:t ) is the total observation sequence a from time 1 to t given at time t 1:t Under the premise that the crop state is p i (t), A is the normalization factor, O(a t |p i (t)) indicates that if the crop is in state p i (t) Then we observe the current observation a at time t t the possibility of Indicates that a weighted sum is to be made for all possible previous states, O(p i (t)|p i (t-1)) indicates that the crop changes from state p i (t-1) transfer to the current state p i The probability of (t), O(p i (t-1)|a 1:t-1 ) indicates that the crop is in state p at the previous time t-1 given the previous t-1 observation sequence i The probability of (t-1), a t is the observation data from images, environment and sensors, a 1:t is the sequence of all observations from time 1 to time t, p i (t) represents the state of the i-th possible crop growth stage at time t, and outputs the current inferred crop type stage: in is the predicted crop growth stage of the i-th micro-block c at time t, It means that among all possible crop growth stages p, the posterior probability O(p|a 1:t ) is the largest stage, O(p|a 1:t ) is a given observation sequence a1, a2, ..., a at time t t The probability that the crop is in growth stage p under the premise of 1:t is the sequence of all observations from time 1 to time t.

[0085] The crop type stage corresponds to a set of most suitable environmental parameter intervals, and a mapping table is stored in the form of a database. After looking up the table, the corresponding control target vector is dynamically loaded. where g i (t) represents the environmental control target parameter vector for the i-th micro-block c at time t, It represents the optimal environmental temperature and humidity value that needs to be maintained in micro-block c for the current crop type and stage. Indicates the target light intensity required for the current crop growth stage. represents the soil parameter control item, Indicates fan control items, [·] ⊥ is the transpose symbol, Represents vector g i (t) is a crop class identified by and growth stage The control parameter space of the decision is extracted, is the predicted crop growth stage of the i-th micro-block c at time t, Represents the crop prediction category corresponding to the micro-block c at time t, and then the multimodal data x of the actual environmental state isi (t) and the control target vector g i (t) to compare the deviation and generate the control input sequence in the future Δt period, expressed as: in is the optimization operation, x i (χ) is the system state vector of the i-th micro-block c at time χ, I i (χ) is the expected environmental state corresponding to the current time χ, represents the weighted square difference between the current state and the target state, d i (χ) is the control amount applied to the micro-block c at time χ, is the weighted square of the control input, It is the overall optimization of all states and control inputs for the next P steps from the current time t, (st)x i (χ+1)=f{x i (χ),d i (χ)} describes how to i (χ) and control input d i (x) Deducing the next state, the local end actuator adjusts the local control execution end in real time after receiving it, executes the sprinkler irrigation, fill light, and fan operations, and periodically updates the state feedback to trigger incremental re-optimization;

[0086] The operation process of the automatic identification module includes:

[0087] The multimodal data includes but is not limited to a continuously acquired image modality sequence, an environmental modality vector sequence, and crop types. The multimodal data are aligned and constructed into a fusion tensor: in is the temporal feature input obtained by fusion of the image and sensor of the i-th micro-block c at time t, is a real matrix space indicating that is a size of ε×(Z q′ +Z e′ ), ε represents the time length, Z q′ is the image feature dimension, Z e′ is the environmental feature dimension, Concat[·] means concatenating two feature matrices into a whole temporal input in the feature dimension, CNN(q c ′(·)) for the continuous images q of micro-block c c ′(·) is input to apply convolutional neural network (CNN), is the continuous environmental data of the i-th micro-block c. The image feature extraction of the image modality sequence adopts a multi-scale convolutional feature encoder. Combined with the growth model, a neural differential system is introduced to capture the nonlinear time-varying feature evolution process. The hidden state at the current moment is obtained by ODE-Solver, and a local environment graph is constructed at the same time. Each node of the local environment graph is a local micro-block. The hidden state of the local micro-block is updated by graph convolution, and the expression formula is: in is the hidden state of the i′th node in the l+1th layer at time t, φ(·) is the GELU activation function, It is a summation operation for each node j′ in the set of all neighbor nodes N(i′) belonging to node j′. Indicates that at time t, the importance of node j' to node i' satisfies H l is the learnable linear transformation matrix of the lth layer, Represent the feature representation of the j′th adjacent node in the lth layer at time t, and build a "crop growth stage automatic identification algorithm system" driven by time series.

[0088] Based on the hidden state, a multi-class classifier is used to output the current stage of the crop. Then, the stage of the crop is further determined, which is expressed as:

[0089] in Represents the current temporal fusion feature of the known micro-block c and crop type prediction results Under the premise of the crop's growth stage is the probability of the k′th stage in the m′th class, and softmax(·) maps the output vector to a probability distribution. It is the feature vector of micro-block c at time t that integrates image, temperature, humidity, lighting and other data. is the crop type prediction result, Targeted at specific crop types The designed linear classifier weight matrix, For crop types The bias vector defined is is the result of the growth phase determination of the i-th micro-block c at time t, It means finding the stage number m′ that makes the subsequent probability value the largest. Indicates that the current fusion feature of the given block Under the premise that the probability value of the crop belongs to the k'th stage under the m'th category, is the probability of the k′th stage in the m′th category. The “crop growth stage automatic identification algorithm system” introduces a confidence dynamic smoothing mechanism to perform time series dynamic correction and adaptive confidence filtering on the stage of the crop.

[0090] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A planting management method for smart agriculture, characterized in that: include: S1, acquire multimodal data by configuring an independent perception module and a local control execution terminal for each micro-block; S2. Using a low-power edge computing device combined with image recognition to identify the current crop type in real time based on the multimodal data; S3. The local terminal adjusts the microenvironment based on the identified crop type by retrieving its phased control parameters; S4. Based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed to dynamically determine the crop stage.

2. The planting management method of smart agriculture according to claim 1, characterized in that: In step S1, by configuring an independent perception module and a local control execution terminal for each micro-block, the method for acquiring multimodal data is as follows: The micro-block is equipped with an independent sensing module to obtain environmental temperature and humidity, light, and soil parameters. The environmental temperature and humidity are collected by an infrared thermal array, and the light is measured by a multi-channel spectroscopic reflectometer. The soil parameters represent volumetric moisture content, electrical conductivity, pH, and nutrient concentration. An all-in-one soil probe is used to continuously monitor soil parameters. A weighted time synchronization fusion model is introduced to fuse the environmental temperature and humidity, light, and soil parameters to obtain multimodal data x i (t), expression formula: where x c (t) is the multimodal data of micro-block c at time t, M is the total number of perceptual channels, q j (t) is the dynamic weight coefficient of the jth mode, is the data of the jth mode after the lag time, ɑ cj is the time delay of the jth mode on the micro-block c relative to the unified time t. The local control execution end includes but is not limited to sprinkler irrigation, fill light, and fan. The controller of the local control execution end generates control instructions through a nonlinear state-behavior mapping function. The control instructions control the sprinkler irrigation, fill light, and fan to operate respectively.

3. The planting management method of smart agriculture according to claim 2, characterized in that: In step S2, a method for real-time identification of the current crop type based on the multimodal data using a low-power edge computing device combined with image recognition is as follows: The low-power edge computing device obtains image frames from the camera at every moment, maps them into high-dimensional image feature vectors through the visual encoder, and at the same time obtains the multimodal data to construct a joint semantic embedding space: Q m (t), in the joint semantic embedding space, crop species recognition is regarded as a multi-category image-environment joint classification task. The image recognition calculates the posterior probability of crop species through a multimodal Bayesian inference framework, which is expressed as: Where W(w k |Q m (t)) represents the given feature Q m (t) belongs to the wth k Class probability, Q m (t) Joint semantic embedding space, β k represents the mean vector of category k, ∑ k represents the covariance matrix of category k, represents the inverse of the covariance matrix of category k, β u represents the mean vector of category u, represents the covariance matrix of category u, It means to sum up the Gaussian likelihood values of all categories, K represents the number of all categories, and the recognition result is: where Q m (t) Joint semantic embedding space, represents the crop prediction category corresponding to the micro-block c at time t, w c Indicates that category c is an element in the set E, W(w k |Q m (t)) represents the given joint semantic embedding space Q at time t m (t) Micro-block c belongs to category w k The probability of E represents the set of all possible categories, which is the set of all crop types and growth stages.

4. The planting management method of smart agriculture according to claim 3, characterized in that: In step S3, the local end adjusts the microenvironment by retrieving the phased control parameters of the identified crop type: The crop species can be divided into multiple key growth stages. The multiple key growth stages include but are not limited to germination, jointing, heading, and filling. The hidden variables of the growth stages are defined and time series prediction is performed using a hidden Markov model (HMM). The expression formula is: Where O(p i (t)|a 1:t ) is the total observation sequence a from time 1 to t given at time t 1:t Under the premise that the crop state is p i (t), A is the normalization factor, O(a t |p i (t)) indicates that if the crop is in state p i (t) Then we observe the current observation a at time t t The possibility of pi(t-1) Indicates that a weighted sum is to be made for all possible previous states, O(p i (t)|p i (t-1)) indicates that the crop changes from state p i (t-1) transfer to the current state p i The probability of (t), O(p i (t-1)|a 1:t-1 ) indicates that the crop is in state p at the previous time t-1 given the previous t-1 observation sequence i The probability of (t-1), a t is the observation data from images, environment and sensors, a 1:t is the sequence of all observations from time 1 to time t, p i (t) represents the state of the i-th possible crop growth stage at time t, and outputs the current inferred crop type stage: in is the predicted crop growth stage of the i-th micro-block c at time t, It means that among all possible crop growth stages p, the posterior probability O(p|a 1:t ) is the largest stage, O(p|a 1:t ) is a given observation sequence a1, a2, ..., a at time t t The probability that the crop is in growth stage p under the premise of 1:t is the sequence of all observations from time 1 to time t.

5. The planting management method of smart agriculture according to claim 4, characterized in that: In step S3, the local end adjusts the microenvironment by retrieving the phased control parameters of the identified crop type: The crop type stage corresponds to a set of most suitable environmental parameter intervals, and a mapping table is stored in the form of a database. After looking up the table, the corresponding control target vector is dynamically loaded. where g i (t) represents the environmental control target parameter vector for the i-th micro-block c at time t, It represents the optimal environmental temperature and humidity value that needs to be maintained in micro-block c for the current crop type and stage. Indicates the target light intensity required for the current crop growth stage. represents the soil parameter control item, Indicates fan control items, [·] ⊥ is the transpose symbol, Represents vector g i (t) is a crop class identified by and growth stage The control parameter space of the decision is extracted, is the predicted crop growth stage of the i-th micro-block c at time t, Represents the crop prediction category corresponding to the micro-block c at time t, and then the multimodal data x of the actual environmental state is i (t) and the control target vector g i (t) to compare the deviation and generate the control input sequence in the future Δt period, expressed as: in is the optimization operation, x i (χ) is the system state vector of the i-th micro-block c at time χ, I i (χ) is the expected environmental state corresponding to the current time χ, represents the weighted square difference between the current state and the target state, d i (χ) is the control amount applied to the micro-block c at time χ, is the weighted square of the control input, It is the overall optimization of all states and control inputs for the next P steps from the current time t, (st)x i (χ+1)=f{x i (χ),d i (χ)} describes how to i (χ) and control input d i (χ) Deducing the next state, the local end actuator adjusts the local control execution end in real time after receiving it, executes the work of sprinkler irrigation, fill light, and fan, and periodically updates the state feedback to trigger incremental re-optimization.

6. The planting management method of smart agriculture according to claim 5, characterized in that: In step S4, based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed. The method for dynamically determining the crop stage is as follows: The multimodal data includes but is not limited to a continuously acquired image modality sequence, an environmental modality vector sequence, and crop types. The multimodal data are aligned and constructed into a fusion tensor: in is the temporal feature input obtained by fusion of the image and sensor of the i-th micro-block c at time t, is a real matrix space indicating that is a size of ε×(Z q′ +Z e′ ), ε represents the time length, Z q′ is the image feature dimension, Z e′ is the environmental feature dimension, Concat[·] means concatenating two feature matrices into a whole temporal input in the feature dimension, CNN(q c ′(·)) for the continuous images q of micro-block c c ′(·) is input to apply convolutional neural network (CNN), is the continuous environmental data of the i-th micro-block c. The image feature extraction of the image modality sequence adopts a multi-scale convolutional feature encoder. Combined with the growth model, a neural differential system is introduced to capture the nonlinear time-varying feature evolution process. The hidden state at the current moment is obtained by ODE-Solver, and a local environment graph is constructed at the same time. Each node of the local environment graph is a local micro-block. The hidden state of the local micro-block is updated by graph convolution, and the expression formula is: in is the hidden state of the i′th node in the l+1th layer at time t, φ(·) is the GELU activation function, It is a summation operation for each node j′ in the set of all neighbor nodes N(i′) belonging to node j′. Indicates that at time t, the importance of node j' to node i' satisfies H l is the learnable linear transformation matrix of the lth layer, Represent the feature representation of the j′th adjacent node in the lth layer at time t, and build a "crop growth stage automatic identification algorithm system" driven by time series.

7. A planting management method for smart agriculture according to claim 6, characterized in that: In step S4, based on the multimodal data, combined with the crop types and growth models, a time series-driven "crop growth stage automatic identification algorithm system" is constructed. The method for dynamically determining the crop stage is as follows: Based on the hidden state, a multi-class classifier is used to output the current stage of the crop. Then, the stage of the crop is further determined, which is expressed as: in Represents the current temporal fusion feature of the known micro-block c and crop type prediction results Under the premise of the crop's growth stage is the probability of the k′th stage in the m′th class, and softmax(·) maps the output vector to a probability distribution. It is the feature vector of micro-block c at time t that integrates image, temperature, humidity, lighting and other data. is the crop type prediction result, Targeted at specific crop types The designed linear classifier weight matrix, For crop types The bias vector defined is is the result of the growth phase determination of the i-th micro-block c at time t, It means finding the stage number m′ that makes the subsequent probability value the largest. Indicates that the current fusion feature of the given block Under the premise that the probability value of the crop belongs to the k'th stage under the m'th category, is the probability of the k′th stage in the m′th category. The “crop growth stage automatic identification algorithm system” introduces a confidence dynamic smoothing mechanism to perform time series dynamic correction and adaptive confidence filtering on the crop stage.

8. A smart agriculture planting management system, based on a smart agriculture planting management method according to any one of claims 1 to 7, characterized in that: include: The distributed micro-block management module acquires multimodal data by configuring an independent perception module and a local control execution terminal for each micro-block; The automatic crop identification module uses low-power edge computing devices combined with image recognition to identify the current crop type in real time based on the multimodal data; The strategy dynamic matching module, based on the identified crop type, retrieves its stage-specific control parameters to adjust the microenvironment and dynamically determines the crop stage; The automatic identification module constructs a time series-driven "crop growth stage automatic identification algorithm system" based on the multimodal data, combined with the crop type and growth model, to dynamically determine the crop stage.

9. The smart agriculture planting management system according to claim 8, characterized in that: The operation process of the distributed micro-block management module includes: The micro-block is equipped with an independent sensing module to obtain environmental temperature and humidity, light, and soil parameters. The environmental temperature and humidity are collected by an infrared thermal array, and the light is measured by a multi-channel spectroscopic reflectometer. The soil parameters represent volumetric moisture content, electrical conductivity, pH, and nutrient concentration. An all-in-one soil probe is used to continuously monitor soil parameters. A weighted time synchronization fusion model is introduced to fuse the environmental temperature and humidity, light, and soil parameters to obtain multimodal data x i (t), expression formula: where x c (t) is the multimodal data of micro-block c at time t, M is the total number of perceptual channels, q j (t) is the dynamic weight coefficient of the jth mode, is the data of the jth mode after the lag time, α cj is the time delay of the jth mode on the micro-block c relative to the unified time t. The local control execution end includes but is not limited to sprinkler irrigation, fill light, and fan. The controller of the local control execution end generates control instructions through a nonlinear state-behavior mapping function. The control instructions control the sprinkler irrigation, fill light, and fan to operate respectively; The operation process of the automatic crop identification module includes: The low-power edge computing device obtains image frames from the camera at every moment, maps them into high-dimensional image feature vectors through the visual encoder, and at the same time obtains the multimodal data to construct a joint semantic embedding space: Q m (t), in the joint semantic embedding space, crop species recognition is regarded as a multi-category image-environment joint classification task. The image recognition calculates the posterior probability of crop species through a multimodal Bayesian inference framework, which is expressed as: Where W(w k |Q m (t)) represents the given feature Q m (t) belongs to the wth k Class probability, Q m (t) Joint semantic embedding space, β k represents the mean vector of category k, ∑ k represents the covariance matrix of category k, represents the inverse of the covariance matrix of category k, β u represents the mean vector of category u, represents the covariance matrix of category u, It means to sum up the Gaussian likelihood values of all categories, K represents the number of all categories, and the recognition result is: where Q m (t) Joint semantic embedding space, represents the crop prediction category corresponding to the micro-block c at time t, w c Indicates that category c is an element in the set E, W(w k |Q m (t)) represents the given joint semantic embedding space Q at time t m (t) Micro-block c belongs to category w k The probability of E represents the set of all possible categories, which is the set of all crop types and growth stages.

10. The smart agriculture planting management system according to claim 9, characterized in that: The operation process of the strategy dynamic matching module includes: The crop species can be divided into multiple key growth stages. The multiple key growth stages include but are not limited to germination, jointing, heading, and filling. The hidden variables of the growth stages are defined and time series prediction is performed using a hidden Markov model (HMM). The expression formula is: Where O(p i (t)|a 1:t ) is the total observation sequence a from time 1 to t given at time t 1:t Under the premise that the crop state is p i (t), A is the normalization factor, O(a t |p i (t)) indicates that if the crop is in state p i (t) Then we observe the current observation a at time t t the possibility of Indicates that a weighted sum is to be made for all possible previous states, O(p i (t)|p i (t-1)) indicates that the crop changes from state p i (t-1) transfer to the current state p i The probability of (t), O(p i (t-1)|a 1:t-1 ) indicates that the crop is in state p at the previous time t-1 given the previous t-1 observation sequence i The probability of (t-1), a t is the observation data from images, environment and sensors, a 1:t is the sequence of all observations from time 1 to time t, p i (t) represents the state of the i-th possible crop growth stage at time t, and outputs the current inferred crop type stage: in is the predicted crop growth stage of the i-th micro-block c at time t, It means that among all possible crop growth stages p, the posterior probability O(p|a 1:t ) is the largest stage, O(p|a 1:t ) is a given observation sequence a1, a2, ..., a at time t t The probability that the crop is in growth stage p under the premise of 1:t is the sequence of all observations from time 1 to time t. The crop type stage corresponds to a set of most suitable environmental parameter intervals, and a mapping table is stored in the form of a database. After looking up the table, the corresponding control target vector is dynamically loaded. where g i (t) represents the environmental control target parameter vector for the i-th micro-block c at time t, It represents the optimal environmental temperature and humidity value that needs to be maintained in micro-block c for the current crop type and stage. Indicates the target light intensity required for the current crop growth stage. represents the soil parameter control item, Indicates fan control items, [·] ⊥ is the transpose symbol, Represents vector g i (t) is a crop class identified by and growth stage The control parameter space of the decision is extracted, is the predicted crop growth stage of the i-th micro-block c at time t, Represents the crop prediction category corresponding to the micro-block c at time t, and then the multimodal data x of the actual environmental state is i (t) and the control target vector g i (t) to compare the deviation and generate the control input sequence in the future Δt period, expressed as: in is the optimization operation, x i (χ) is the system state vector of the i-th micro-block c at time χ, I i (χ) is the expected environmental state corresponding to the current time χ, represents the weighted square difference between the current state and the target state, d i (χ) is the control amount applied to the micro-block c at time χ, is the weighted square of the control input, It is the overall optimization of all states and control inputs for the next P steps from the current time t, (st)x i (χ+1)=f{x i (χ),d i (χ)} describes how to i (χ) and control input d i (x) Deducing the next state, the local end actuator adjusts the local control execution end in real time after receiving it, executes the sprinkler irrigation, fill light, and fan operations, and periodically updates the state feedback to trigger incremental re-optimization; The operation process of the automatic identification module includes: The multimodal data includes but is not limited to a continuously acquired image modality sequence, an environmental modality vector sequence, and crop types. The multimodal data are aligned and constructed into a fusion tensor: in is the temporal feature input obtained by fusion of the image and sensor of the i-th micro-block c at time t, is a real matrix space indicating that is a size of ε×(Z q′ +Z e′ ), ε represents the time length, Z q′ is the image feature dimension, Z e′ is the environmental feature dimension, Concat[·] means concatenating two feature matrices into a whole temporal input in the feature dimension, CNN(q c ′(·)) for the continuous images q of micro-block c c ′(·) is input to apply convolutional neural network (CNN), is the continuous environmental data of the i-th micro-block c. The image feature extraction of the image modality sequence adopts a multi-scale convolutional feature encoder. Combined with the growth model, a neural differential system is introduced to capture the nonlinear time-varying feature evolution process. The hidden state at the current moment is obtained by ODE-Solver, and a local environment graph is constructed at the same time. Each node of the local environment graph is a local micro-block. The hidden state of the local micro-block is updated by graph convolution, and the expression formula is: in is the hidden state of the i′th node in the l+1th layer at time t, φ(·) is the GELU activation function, It is a summation operation for each node j′ in the set of all neighbor nodes N(i′) belonging to node j′. Indicates that at time t, the importance of node j' to node i' satisfies H l is the learnable linear transformation matrix of the lth layer, Represent the feature representation of the j′th adjacent node in the lth layer at time t, and build a "crop growth stage automatic identification algorithm system" driven by time series. Based on the hidden state, a multi-class classifier is used to output the current stage of the crop. Then, the stage of the crop is further determined, which is expressed as: in Represents the current temporal fusion feature of the known micro-block c and crop type prediction results Under the premise of the crop's growth stage is the probability of the k′th stage in the m′th class, and softmax(·) maps the output vector to a probability distribution. It is the feature vector of micro-block c at time t that integrates image, temperature, humidity, lighting and other data. is the crop type prediction result, Targeted at specific crop types The designed linear classifier weight matrix, For crop types The bias vector defined is is the result of the growth phase determination of the i-th micro-block c at time t, It means finding the stage number m′ that makes the subsequent probability value the largest. Indicates that the current fusion feature of the given block Under the premise that the probability value of the crop belongs to the k'th stage under the m'th category, is the probability of the k′th stage in the m′th category. The “crop growth stage automatic identification algorithm system” introduces a confidence dynamic smoothing mechanism to perform time series dynamic correction and adaptive confidence filtering on the crop stage.

Images