Hyperspectral dynamic imaging monitoring method for facility agriculture

Abstract

The invention relates to the technical field of communication data processing, in particular to a facility agriculture hyperspectral dynamic imaging monitoring method which comprises the steps that a mechanical arm carries a hyperspectral imaging unit to collect crop canopy images and synchronously collect environment sensing data, and the motion path of the mechanical arm is planned through a reinforcement learning algorithm; environmental data are fused through a filtering algorithm, and the light supplementing intensity is adjusted through photoelectric monitoring feedback in the collection process; transmitting the collected data through a software defined network architecture, analyzing and processing the data, inputting features into a digital twin model to invert canopy physiological parameters, feeding back an inversion result, adjusting the collection frequency, and correcting the model; generating a growth situation evaluation index through the fuzzy logic decision tree, and pushing the index to a user terminal for display; and an inversion result is fed back to an acquisition system to form closed-loop control, so that the problem of real-time sensing misalignment caused by environmental interference and data delay is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of communication data processing technology, and in particular to a hyperspectral dynamic imaging monitoring method for facility agriculture. Background Technology

[0002] Existing monitoring technologies for facility agriculture suffer from the following technical challenges: Handheld spectrometers are susceptible to operator tremors and angular deviations, making it difficult to maintain stable optical distance and axis, resulting in spatial distortion and spectral aberration in the collected spectral data; Fixed monitoring nodes, limited by their field of view and installation height, cannot adapt to the dynamic changes in crop morphology during growth, causing monitoring blind spots to expand with the growth period; Existing mobile robot platforms mostly employ centralized data processing architectures, requiring raw hyperspectral data to be transmitted back to remote servers for computation. Constrained by fluctuations in wireless transmission bandwidth and network latency within the greenhouse, these platforms cannot meet the real-time inversion requirements for canopy nutrient indicators, pest and disease patches, and water stress conditions, delaying critical decision-making windows for fertilization, irrigation, and disease control. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture. This invention solves the technical problem of inaccurate real-time perception of crop growth status caused by interference from the complex environment of facility agriculture and data transmission and processing delays of existing monitoring methods.

[0004] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows:

[0005] This invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture, comprising:

[0006] Step 1: A robotic arm carrying a hyperspectral imaging unit acquires hyperspectral images of the crop canopy and simultaneously collects environmental sensor data. The movement path of the robotic arm is dynamically planned using a reinforcement learning algorithm, and the environmental sensor data is fused and processed using a filtering algorithm. During the acquisition process, the intensity of supplementary lighting is adjusted based on photoelectric monitoring feedback.

[0007] Step 2: Transmit the hyperspectral image data and environmental sensing data acquired in Step 1 through a software-defined network architecture. Prioritize the hyperspectral image data stream according to the vegetation index and apply error correction coding. The transmission protocol dynamically adjusts the window parameters according to the wireless channel status. The environmental data stream is aligned with the image data stream using a timestamp synchronization mechanism.

[0008] Step 3: Receive the hyperspectral image data and environmental sensing data transmitted in Step 2, perform strip noise suppression and atmospheric correction on the hyperspectral image at the distributed processing node, enhance the spatial resolution of the image through generative adversarial network, and extract crop growth characteristics by fusing time series environmental data and hyperspectral data, including nutrient stress characteristics and pest and disease patch characteristics.

[0009] Step 4: Input the crop growth characteristics extracted in Step 3 into the crop growth digital twin model to retrieve canopy physiological parameters. The retrieval results are fed back to the robotic arm control system to adjust the acquisition frequency. At the same time, the digital twin model is corrected through system identification methods.

[0010] Step 5: The canopy physiological parameters obtained in step 4 are used to generate crop growth status assessment indicators through a fuzzy logic decision tree. The assessment indicators are then pushed to the user terminal for visualization via a message queue telemetry transmission protocol.

[0011] The inversion results are fed back from step 4 to step 1, and the robotic arm's data acquisition strategy is adjusted to form a closed-loop control system.

[0012] Furthermore, in the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention, step 1 employs the Q-learning algorithm to construct a robotic arm motion path planning model. The model's state space is a set of three-dimensional coordinates of the crop canopy, and its action space is the robotic arm motion vector. The reward function includes an imaging quality score and an energy consumption index. This model is used to plan the robotic arm's motion path. The filtering algorithm employs the Kalman filter algorithm to perform a prediction-update loop on the environmental sensing data to eliminate noise. The supplementary light intensity adjustment uses a photodiode to monitor the ambient light intensity, and the monitoring results are used to dynamically adjust the LED driving current based on the spectral reflectance characteristics.

[0013] Furthermore, in the facility agriculture hyperspectral dynamic imaging monitoring method of the present invention, in step 2, the software-defined network architecture collects network status information through the OpenFlow protocol, and the collected network status information is input into the Dijkstra algorithm to calculate the transmission path; during the priority division process, Reed-Solomon coding is applied to the low-frequency components of the hyperspectral image data, and discrete wavelet transform is used to compress the high-frequency components; the transmission protocol uses the TCP Westwood algorithm to dynamically estimate the bandwidth, and the estimated bandwidth result is adjusted by an adaptive retransmission timeout mechanism to adjust the transmission window parameters.

[0014] Furthermore, in the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention, step 3 employs an improved U-Net architecture for strip noise suppression. Residual connections are introduced at the encoder end to extract image features, and transposed convolution is used at the decoder end to reconstruct the features, generating a noise-suppressed hyperspectral image. Atmospheric correction is performed on the noise-suppressed hyperspectral image. The atmospheric correction process integrates a radiative transfer model, and atmospheric transmittance is calculated using a lookup table method to obtain the corrected hyperspectral image. Feature extraction is then performed on the corrected hyperspectral image. Long Short-Term Memory (LSTM) networks are used to analyze the environmental data time series. An attention mechanism is introduced into the input gate of the LTM network to identify the time nodes where stress features appear. Simultaneously, principal component analysis (PCA) is used to reduce the dimensionality of the hyperspectral data and extract crop growth features.

[0015] Furthermore, in the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention, in step 4, the crop growth digital twin model uses the Richards equation to describe the canopy growth dynamics, and the crop growth features extracted in step 3 are assimilated into the digital twin model through extended Kalman filtering; the inversion process uses a particle swarm optimization algorithm, the inertia weight of the particle swarm optimization algorithm adopts a linear decreasing strategy, and the social learning factor adopts a random topological structure, and the canopy physiological parameters are inverted through this algorithm; the system identification uses recursive least squares to update the model parameters online, and the forgetting factor of recursive least squares is adaptively adjusted according to the prediction error of the digital twin model, and the updated model parameters are used to correct the digital twin model.

[0016] Furthermore, in the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention, the fuzzy logic decision tree in step 5 adopts the Takagi-Sugeno model. The canopy physiological parameters obtained in step 4 are input into the fuzzy logic decision tree. The membership function of the fuzzy logic decision tree is generated through fuzzy C-means clustering. The rule base is constructed based on the expert knowledge graph. Crop growth status assessment indicators are generated through this decision tree. The crop growth status assessment indicators include nitrogen deficiency level, water stress index, and disease occurrence probability. Among them, the nitrogen deficiency level is quantitatively inverted by fusing the spectral index NDRE and MCARI through a support vector regression model.

[0017] Furthermore, the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention further includes: setting a federated learning framework between step 3 and step 4; the federated learning framework uses the FedAvg algorithm to aggregate the model parameters of distributed processing nodes; during the aggregation process, Laplace noise is added using differential privacy technology to protect data security; the updated model parameters of the federated learning framework are fed back to the robotic arm motion path planning model in step 1 to optimize the acquisition strategy, and simultaneously transmitted to the fuzzy logic decision tree in step 5 to update the evaluation rules.

[0018] Beneficial effects of this invention:

[0019] This invention effectively solves the technical problem of inaccurate real-time perception of crop growth status caused by interference from complex environments in facility agriculture and data transmission and processing delays in existing monitoring methods. It adopts a robotic arm to dynamically plan the acquisition path to avoid spatial distortion and spectral distortion caused by manual operation. It uses a software-defined network architecture and priority transmission mechanism to cope with fluctuations in wireless transmission bandwidth. It uses distributed processing nodes and digital twin models to achieve efficient data processing and accurate inversion. Combined with a federated learning framework, it continuously optimizes system performance and ultimately achieves real-time and accurate perception and decision support of crop growth status. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a hyperspectral dynamic imaging monitoring method for facility agriculture provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0023] Please see Figure 1 The present invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture, comprising:

[0024] Step 1: A robotic arm carrying a hyperspectral imaging unit acquires hyperspectral images of the crop canopy and simultaneously collects environmental sensor data. The movement path of the robotic arm is dynamically planned using a reinforcement learning algorithm, and the environmental sensor data is fused and processed using a filtering algorithm. During the acquisition process, the intensity of supplementary lighting is adjusted based on photoelectric monitoring feedback.

[0025] Step 2: Transmit the hyperspectral image data and environmental sensing data acquired in Step 1 through a software-defined network architecture. Prioritize the hyperspectral image data stream according to the vegetation index and apply error correction coding. The transmission protocol dynamically adjusts the window parameters according to the wireless channel status. The environmental data stream is aligned with the image data stream using a timestamp synchronization mechanism.

[0026] Step 3: Receive the hyperspectral image data and environmental sensing data transmitted in Step 2, perform strip noise suppression and atmospheric correction on the hyperspectral image at the distributed processing node, enhance the spatial resolution of the image through generative adversarial network, and extract crop growth characteristics by fusing time series environmental data and hyperspectral data, including nutrient stress characteristics and pest and disease patch characteristics.

[0027] Step 4: Input the crop growth characteristics extracted in Step 3 into the crop growth digital twin model to retrieve canopy physiological parameters. The retrieval results are fed back to the robotic arm control system to adjust the acquisition frequency. At the same time, the digital twin model is corrected through system identification methods.

[0028] Step 5: The canopy physiological parameters obtained in step 4 are used to generate crop growth status assessment indicators through a fuzzy logic decision tree. The assessment indicators are then pushed to the user terminal for visualization via a message queue telemetry transmission protocol.

[0029] The inversion results are fed back from step 4 to step 1, and the robotic arm's data acquisition strategy is adjusted to form a closed-loop control system.

[0030] The present invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture, which specifically includes the following steps:

[0031] In step one, a robotic arm carrying a hyperspectral imaging unit acquires hyperspectral images of the crop canopy. The robotic arm's motion path is dynamically planned using a reinforcement learning algorithm, which employs Q-learning to construct the motion path planning model. Simultaneously, environmental sensor data is collected and fused using a filtering algorithm. This filtering algorithm uses a Kalman filter to perform a prediction-update loop on the environmental sensor data to eliminate noise. During the acquisition process, photoelectric monitoring feedback adjusts the supplementary lighting intensity. Photoelectric monitoring uses photodiodes to monitor ambient light intensity, and the monitoring results are used to dynamically adjust the LED driving current based on spectral reflectance characteristics.

[0032] In step two, the hyperspectral image data and environmental sensor data acquired in step one are transmitted via a software-defined network architecture. This architecture collects network status information using the OpenFlow protocol, and this information is input into the Dijkstra algorithm to calculate the transmission path. The hyperspectral image data stream is prioritized according to vegetation indices and error correction coding is applied. During the prioritization process, Reed-Solomon coding is applied to the low-frequency components of the hyperspectral image data, while discrete wavelet transform is used to compress the high-frequency components. The transmission protocol dynamically adjusts the window parameters based on the wireless channel status and uses the TCP Westwood algorithm to dynamically estimate bandwidth. The estimated bandwidth is then used to adjust the transmission window parameters through an adaptive retransmission timeout mechanism. The environmental data stream is synchronized with the image data stream using a timestamp synchronization mechanism.

[0033] In step three, the hyperspectral image data and environmental sensor data transmitted in step two are received. Strip noise suppression and atmospheric correction are performed on the hyperspectral image at the distributed processing node. Strip noise suppression employs an improved U-Net architecture, with residual connections introduced at the encoder to extract image features, and transposed convolution used at the decoder to reconstruct the features, generating a noise-suppressed hyperspectral image. Atmospheric correction is then performed on the noise-suppressed hyperspectral image. The atmospheric correction process integrates a radiative transfer model and calculates atmospheric transmittance using a lookup table method to obtain the corrected hyperspectral image. Feature extraction is then performed on the corrected hyperspectral image. Long Short-Term Memory (LSTM) networks are used to analyze the environmental data time series. An attention mechanism is introduced into the input gate of the LTM network to identify the time nodes where stress features appear. Simultaneously, principal component analysis (PCA) is used to reduce the dimensionality of the hyperspectral data and extract crop growth features, including nutrient stress features and pest and disease patch features.

[0034] In step four, the crop growth characteristics extracted in step three are input into the crop growth digital twin model to invert canopy physiological parameters. The crop growth digital twin model uses the Richards equation to describe canopy growth dynamics, and the crop growth characteristics are assimilated into the digital twin model through extended Kalman filtering. The inversion process uses a particle swarm optimization algorithm, with the inertia weight employing a linear decreasing strategy and the social learning factor using a random topology structure. The inversion results are fed back to the robotic arm control system to adjust the acquisition frequency. Simultaneously, the digital twin model is corrected using a system identification method. The system identification method uses recursive least squares to update the model parameters online. The forgetting factor of the recursive least squares method is adaptively adjusted according to the prediction error of the digital twin model. The updated model parameters are used to correct the digital twin model.

[0035] In step five, the canopy physiological parameters obtained in step four are used to generate crop growth status assessment indicators through a fuzzy logic decision tree. The fuzzy logic decision tree adopts the Takagi-Sugeno model, and the membership function is generated through fuzzy C-means clustering. The rule base is constructed based on an expert knowledge graph. The crop growth status assessment indicators include nitrogen deficit level, water stress index, and disease occurrence probability. The nitrogen deficit level is quantitatively retrieved by fusing the spectral index NDRE and MCARI using a support vector regression model. The assessment indicators are pushed to the user terminal for visualization via a message queue telemetry transmission protocol. The inversion results are fed back from step four to step one, adjusting the robotic arm's data acquisition strategy to form a closed-loop control system.

[0036] Furthermore, a federated learning framework is set up between steps three and four. This framework uses the FedAvg algorithm to aggregate the model parameters from the distributed processing nodes. During the aggregation process, Laplace noise is added using differential privacy technology to protect data security. The updated model parameters from the federated learning framework are fed back to the robotic arm motion path planning model in step one to optimize the data acquisition strategy, and simultaneously transmitted to the fuzzy logic decision tree in step five to update the evaluation rules.

[0037] Specifically, in the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention, step 1 uses the Q-learning algorithm to construct a robotic arm motion path planning model. The state space of the model is a set of three-dimensional coordinates of the crop canopy, the action space is the robotic arm motion vector, and the reward function includes an imaging quality score and an energy consumption index. The robotic arm motion path is planned through this model. The filtering algorithm uses the Kalman filter algorithm to perform a prediction-update loop on the environmental sensing data to eliminate noise. The supplementary light intensity adjustment uses a photodiode to monitor the ambient light intensity, and the monitoring results are used to dynamically adjust the LED driving current based on the spectral reflectance characteristics.

[0038] This invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture. In step one, a reinforcement learning algorithm using the Q-learning algorithm is employed to construct a robotic arm motion path planning model. The model's state space is a set of three-dimensional coordinates of the crop canopy, and its action space is the robotic arm motion vector. The reward function includes an image quality score and an energy consumption index. This model is used to plan the robotic arm's motion path. A Kalman filter algorithm is used to perform a prediction-update loop on the environmental sensor data to eliminate noise. The prediction-update loop includes a time update step and a measurement update step. Supplemental lighting intensity adjustment is achieved by monitoring ambient light intensity using photodiodes. The monitoring results are used to dynamically adjust the LED driving current based on spectral reflectance characteristics, which are obtained by analyzing hyperspectral image data.

[0039] Specifically, in the facility agriculture hyperspectral dynamic imaging monitoring method of the present invention, in step 2, the software-defined network architecture collects network status information through the OpenFlow protocol, and the collected network status information is input into the Dijkstra algorithm to calculate the transmission path; during the priority division process, Reed-Solomon coding is applied to the low-frequency components of the hyperspectral image data, and discrete wavelet transform is used to compress the high-frequency components; the transmission protocol uses the TCP Westwood algorithm to dynamically estimate the bandwidth, and the estimated bandwidth result is adjusted by an adaptive retransmission timeout mechanism to adjust the transmission window parameters.

[0040] This invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture. In step two, a software-defined network architecture collects network status information via the OpenFlow protocol. The collected network status information is input into the Dijkstra algorithm to calculate the transmission path. The network status information includes network bandwidth utilization and transmission delay data. During priority allocation, Reed-Solomon coding is applied to the low-frequency components of the hyperspectral image data, and discrete wavelet transform is used to compress the high-frequency components. The low-frequency components include the main structural information of the image. The transmission protocol uses the TCP Westwood algorithm to dynamically estimate the bandwidth. The estimated bandwidth result adjusts the transmission window parameters through an adaptive retransmission timeout mechanism, which dynamically adjusts the retransmission time threshold according to the network congestion level.

[0041] Specifically, in the facility agriculture hyperspectral dynamic imaging monitoring method of the present invention, step 3 of the strip noise suppression adopts an improved U-Net architecture. The encoder end introduces residual connections to extract image features, and the decoder end uses transposed convolution to reconstruct the features to generate a noise-suppressed hyperspectral image. Atmospheric correction is performed on the noise-suppressed hyperspectral image. The atmospheric correction process integrates the radiative transfer model and calculates atmospheric transmittance using a lookup table method to obtain the corrected hyperspectral image. Feature extraction is performed on the corrected hyperspectral image. Feature extraction uses a long short-term memory network to analyze the environmental data time series. The input gate of the long short-term memory network introduces an attention mechanism to identify the time nodes when stress features appear. At the same time, principal component analysis algorithm is used to reduce the dimensionality of the hyperspectral data and extract crop growth features.

[0042] This invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture. In step three, stripe noise suppression employs an improved U-Net architecture. Residual connections are introduced at the encoder end to extract image features, and transposed convolution is used at the decoder end to reconstruct the features, generating a noise-suppressed hyperspectral image. Atmospheric correction is performed on the noise-suppressed hyperspectral image. The atmospheric correction process integrates a radiative transfer model, and atmospheric transmittance is calculated using a lookup table method to obtain the corrected hyperspectral image. Feature extraction is performed on the corrected hyperspectral image. Feature extraction uses a Long Short-Term Memory (LSTM) network to analyze the environmental data time series. An attention mechanism is introduced into the input gate of the LSM network to identify the time nodes where stress features appear. Simultaneously, principal component analysis (PCA) is used to reduce the dimensionality of the hyperspectral data and extract crop growth features.

[0043] Specifically, in the facility agriculture hyperspectral dynamic imaging monitoring method of the present invention, in step 4, the crop growth digital twin model uses the Richards equation to describe the canopy growth dynamics, and the crop growth features extracted in step 3 are assimilated into the digital twin model through extended Kalman filtering; the inversion process uses the particle swarm optimization algorithm, the inertia weight of the particle swarm optimization algorithm adopts a linear decreasing strategy, and the social learning factor adopts a random topology structure, and the canopy physiological parameters are inverted through this algorithm; the system identification uses the recursive least squares method to update the model parameters online, and the forgetting factor of the recursive least squares method is adaptively adjusted according to the prediction error of the digital twin model, and the updated model parameters are used to correct the digital twin model.

[0044] This invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture. In step four, the crop growth digital twin model uses the Richards equation to describe canopy growth dynamics, and the crop growth features extracted in step three are assimilated into the digital twin model using extended Kalman filtering. The inversion process employs a particle swarm optimization algorithm, with a linearly decreasing inertia weight and a stochastic topology for the social learning factor. This algorithm is used to invert canopy physiological parameters. System identification uses recursive least squares to update model parameters online. The forgetting factor of the recursive least squares method is adaptively adjusted according to the prediction error of the digital twin model, and the updated model parameters are used to correct the digital twin model.

[0045] Specifically, in the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention, the fuzzy logic decision tree in step 5 adopts the Takagi-Sugeno model. The canopy physiological parameters obtained in step 4 are input into the fuzzy logic decision tree. The membership function of the fuzzy logic decision tree is generated through fuzzy C-means clustering. The rule base is constructed based on the expert knowledge graph. Crop growth status assessment indicators are generated through this decision tree. The crop growth status assessment indicators include nitrogen deficiency level, water stress index, and disease occurrence probability. Among them, the nitrogen deficiency level is quantitatively inverted by fusing the spectral index NDRE and MCARI through a support vector regression model.

[0046] This invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture. In step five, the fuzzy logic decision tree adopts the Takagi-Sugeno model. The canopy physiological parameters obtained in step four are input into the fuzzy logic decision tree. The membership function of the fuzzy logic decision tree is generated through fuzzy C-means clustering. The rule base is constructed based on an expert knowledge graph. Crop growth status assessment indicators are generated through this decision tree. The crop growth status assessment indicators include nitrogen deficiency level, water stress index, and disease occurrence probability. Among them, the nitrogen deficiency level is quantitatively retrieved from the canopy physiological parameters by fusing the spectral index NDRE and MCARI through a support vector regression model.

[0047] Specifically, the hyperspectral dynamic imaging monitoring method for facility agriculture described in this invention further includes: setting up a federated learning framework between step 3 and step 4; the federated learning framework uses the FedAvg algorithm to aggregate the model parameters of distributed processing nodes; and adding Laplace noise to protect data security during the aggregation process using differential privacy technology; the updated model parameters of the federated learning framework are fed back to the robotic arm motion path planning model in step 1 to optimize the acquisition strategy, and simultaneously transmitted to the fuzzy logic decision tree in step 5 to update the evaluation rules.

[0048] This invention provides a hyperspectral dynamic imaging monitoring method for facility agriculture. Between steps three and four, a federated learning framework is established. This framework uses the FedAvg algorithm to aggregate model parameters from distributed processing nodes. During the aggregation process, Laplace noise is added using differential privacy technology to protect data security. The updated model parameters from the federated learning framework are fed back to the robotic arm motion path planning model in step one to optimize the acquisition strategy, and simultaneously transmitted to the fuzzy logic decision tree in step five to update the evaluation rules. Through periodic model parameter aggregation and distribution, the federated learning framework enables collaborative learning and model optimization among multiple distributed processing nodes.

[0049] This invention addresses the technical problem of inaccurate real-time crop growth monitoring caused by interference from complex environments in facility agriculture and data transmission delays in existing monitoring methods. It employs a closed-loop control system. In the data acquisition phase, a robotic arm carrying a hyperspectral imaging unit is used. A reinforcement learning algorithm dynamically plans the movement path, enabling the acquisition device to adapt to changes in crop canopy morphology, maintaining optimal imaging distance and angle, and avoiding spatial distortion and spectral aberration caused by manual operation. Simultaneously collected environmental sensor data is fused using a filtering algorithm to eliminate environmental fluctuations and ensure data acquisition stability.

[0050] In the data transmission stage, a software-defined network architecture is adopted. Hyperspectral image data streams are prioritized and error-correcting codes are applied based on vegetation indices. The transmission protocol dynamically adjusts window parameters according to the wireless channel status, effectively addressing the issue of fluctuating wireless transmission bandwidth within the greenhouse. The environmental data stream uses a timestamp synchronization mechanism to align with the image data stream, achieving spatiotemporal consistency of multi-source data and reducing data asynchrony issues caused by transmission delays.

[0051] In the data processing stage, distributed processing nodes are used to perform stripe noise suppression and atmospheric correction on hyperspectral images. Generative adversarial networks are used to enhance the spatial resolution of images and improve data quality. Crop growth characteristics are extracted by fusing time-series environmental data and hyperspectral data, achieving accurate extraction of nutrient stress features and pest and disease patch features.

[0052] In the data inversion stage, the extracted crop growth characteristics are input into the crop growth digital twin model to invert canopy physiological parameters. The digital twin model is then corrected using a system identification method to improve the model's predictive accuracy. The inversion results are fed back to the robotic arm control system to adjust the acquisition frequency, forming a closed-loop mechanism of forward acquisition and backward correction.

[0053] In the decision-making output stage, the inverted canopy physiological parameters are used to generate crop growth status assessment indicators through a fuzzy logic decision tree. These assessment indicators are then pushed to the user terminal for visualization via a message queue telemetry transmission protocol, enabling real-time output of monitoring results. Simultaneously, the inversion results are fed back from the digital twin model to the robotic arm data acquisition system, adjusting the acquisition strategy to form a complete closed-loop control system. This effectively solves the problem of inaccurate real-time perception caused by environmental interference and data latency.

[0054] The model parameters of distributed processing nodes are aggregated by a federated learning framework, and differential privacy technology is used to protect data security. The updated model parameters are fed back to the robotic arm motion path planning model to optimize the acquisition strategy, and at the same time transmitted to the fuzzy logic decision tree to update the evaluation rules, further improving the system's adaptability and real-time performance.

[0055] The specific implementation of this invention is as follows:

[0056] In facility agriculture monitoring scenarios, to address issues such as the instability of handheld devices, blind spots at fixed monitoring nodes, and latency in centralized data processing, a robotic arm carrying a hyperspectral imaging unit is used to acquire crop canopy images. The robotic arm's motion path is dynamically planned using the Q-learning algorithm. This algorithm constructs a state space including a set of three-dimensional coordinates of the crop canopy, and the action space is defined as the robotic arm's motion vectors. The reward function comprehensively considers image quality scores and energy consumption indicators, enabling the robotic arm to adapt to changes in crop growth morphology. Simultaneously acquired environmental sensor data is fused using a Kalman filter algorithm, and a prediction-update loop is executed to eliminate noise interference. During acquisition, ambient light intensity is monitored using photodiodes, and the LED drive current is dynamically adjusted based on real-time spectral reflectance characteristics to maintain imaging stability.

[0057] The acquired hyperspectral image data and environmental sensor data are transmitted via a software-defined network architecture. The OpenFlow protocol is used to collect network status information, including network bandwidth utilization and transmission latency. This collected network status information is input into the Dijkstra algorithm to calculate the optimal transmission path. The hyperspectral image data stream is prioritized according to vegetation indices. Reed-Solomon coding is applied to low-frequency components, and discrete wavelet transform is used to compress high-frequency components. The transmission protocol uses the TCP Westwood algorithm to dynamically estimate bandwidth and adjusts the transmission window parameters through an adaptive retransmission timeout mechanism to effectively cope with wireless channel fluctuations. The environmental data stream uses a timestamp synchronization mechanism to align with the image data stream, ensuring spatiotemporal consistency of multi-source data.

[0058] The transmitted data is processed on distributed processing nodes. An improved U-Net architecture is used for stripe noise suppression. Residual connections are introduced at the encoder to extract image features, and transposed convolution is used at the decoder to reconstruct the image, generating a noise-suppressed hyperspectral image. Atmospheric correction is performed on the noise-suppressed image, and a radiative transfer model is fused. Atmospheric transmittance is calculated using a lookup table method to obtain the corrected hyperspectral image. Feature extraction employs a long short-term memory network to analyze the environmental data time series. An attention mechanism is introduced into the input gate to identify the time nodes where stress features appear. Principal component analysis is used to reduce the dimensionality of the hyperspectral data, extracting crop growth features including nutrient stress and pest / disease patch characteristics.

[0059] Extracted crop growth features are input into a crop growth digital twin model, which uses the Richards equation to describe canopy growth dynamics. Extended Kalman filtering assimilates the features into the model. The inversion process employs a particle swarm optimization algorithm, with a linear decreasing strategy for inertia weights and a stochastic topology for the social learning factor, yielding canopy physiological parameters. System identification uses recursive least squares to update model parameters online, and the forgetting factor is adaptively adjusted based on the digital twin model's prediction error. The updated parameters are used to correct the model. The inversion results are fed back to the robotic arm control system to adjust the acquisition frequency, forming an acquisition-correction closed loop.

[0060] Canopy physiological parameters were used to generate growth status assessment indicators using a fuzzy logic decision tree. The decision tree employed a Takagi-Sugeno model, and the membership function was generated through fuzzy C-means clustering. The rule base was constructed based on an expert knowledge graph. Assessment indicators included nitrogen deficit level, water stress index, and disease occurrence probability. The nitrogen deficit level was quantitatively retrieved using a support vector regression model that fused the spectral index NDRE and MCARI. Assessment results were pushed to user terminals for visualization via a message queue telemetry transmission protocol.

[0061] A federated learning framework is set up between feature extraction and model inversion. The FedAvg algorithm is used to aggregate model parameters from distributed nodes. During the aggregation process, Laplace noise is added using differential privacy technology to protect data security. The updated model parameters are fed back to the robotic arm motion path planning model optimization acquisition strategy and simultaneously transmitted to the fuzzy logic decision tree to update the evaluation rules, thereby achieving continuous optimization of the overall system performance.

Claims

1. A method for monitoring facility agriculture using hyperspectral dynamic imaging, characterized in that, include: Step 1: A robotic arm carrying a hyperspectral imaging unit acquires hyperspectral images of the crop canopy and simultaneously collects environmental sensor data. The movement path of the robotic arm is dynamically planned using a reinforcement learning algorithm, and the environmental sensor data is fused and processed using a filtering algorithm. During the acquisition process, the intensity of supplementary lighting is adjusted based on photoelectric monitoring feedback. Step 2: Transmit the hyperspectral image data and environmental sensing data acquired in Step 1 through a software-defined network architecture. Prioritize the hyperspectral image data stream according to the vegetation index and apply error correction coding. The transmission protocol dynamically adjusts the window parameters according to the wireless channel status. The environmental data stream is aligned with the image data stream using a timestamp synchronization mechanism. Step 3: Receive the hyperspectral image data and environmental sensing data transmitted in Step 2, perform strip noise suppression and atmospheric correction on the hyperspectral image at the distributed processing node, enhance the spatial resolution of the image through generative adversarial network, and extract crop growth characteristics by fusing time series environmental data and hyperspectral data, including nutrient stress characteristics and pest and disease patch characteristics. Step 4: Input the crop growth characteristics extracted in Step 3 into the crop growth digital twin model to retrieve canopy physiological parameters. The retrieval results are fed back to the robotic arm control system to adjust the acquisition frequency. At the same time, the digital twin model is corrected through system identification methods. Step 5: The canopy physiological parameters obtained in step 4 are used to generate crop growth status assessment indicators through a fuzzy logic decision tree. The assessment indicators are then pushed to the user terminal for visualization via a message queue telemetry transmission protocol.

2. The method for hyperspectral dynamic imaging monitoring of facility agriculture according to claim 1, characterized in that, In step 1, the reinforcement learning algorithm uses the Q-learning algorithm to construct a robotic arm motion path planning model. The model's state space is a set of three-dimensional coordinates of the crop canopy, and the action space is the robotic arm motion vector. The reward function includes an image quality score and an energy consumption index. The robotic arm's motion path is planned using this model. The filtering algorithm uses the Kalman filter algorithm to perform a prediction-update loop on the environmental sensing data to eliminate noise. The supplementary lighting intensity adjustment uses a photodiode to monitor the ambient light intensity, and the monitoring results are used to dynamically adjust the LED driving current based on the spectral reflectance characteristics.

3. The method for hyperspectral dynamic imaging monitoring of facility agriculture according to claim 1, characterized in that, In step 2, the software-defined network architecture collects network status information through the OpenFlow protocol, and the collected network status information is input into the Dijkstra algorithm to calculate the transmission path. During the priority division process, Reed-Solomon coding is applied to the low-frequency components of the hyperspectral image data, and discrete wavelet transform is used to compress the high-frequency components. The transmission protocol uses the TCP Westwood algorithm to dynamically estimate the bandwidth, and the estimated bandwidth result is adjusted to adjust the transmission window parameters through an adaptive retransmission timeout mechanism.

4. The method for hyperspectral dynamic imaging monitoring of facility agriculture according to claim 1, characterized in that, In step 3, strip noise suppression adopts an improved U-Net architecture. The encoder introduces residual connections to extract image features, and the decoder uses transposed convolution to reconstruct the features, generating a noise-suppressed hyperspectral image. Atmospheric correction is performed on the noise-suppressed hyperspectral image. The atmospheric correction process incorporates a radiative transfer model, and atmospheric transmittance is calculated using a lookup table method to obtain the corrected hyperspectral image. Feature extraction is then performed on the corrected hyperspectral image. Long Short-Term Memory (LSTM) network is used to analyze the environmental data time series. An attention mechanism is introduced into the input gate of the LSM network to identify the time nodes when stress features appear. At the same time, principal component analysis is used to reduce the dimensionality of the hyperspectral data and extract crop growth features.

5. The method for hyperspectral dynamic imaging monitoring of facility agriculture according to claim 1, characterized in that, In step 4, the crop growth digital twin model uses the Richards equation to describe canopy growth dynamics, and the crop growth features extracted in step 3 are assimilated into the digital twin model through extended Kalman filtering. The inversion process uses the particle swarm optimization algorithm, with the inertia weights employing a linear decreasing strategy and the social learning factor using a random topology structure. The canopy physiological parameters are inverted through this algorithm. The system identification uses the recursive least squares method to update the model parameters online. The forgetting factor of the recursive least squares method is adaptively adjusted according to the prediction error of the digital twin model, and the updated model parameters are used to correct the digital twin model.

6. The method for hyperspectral dynamic imaging monitoring of facility agriculture according to claim 1, characterized in that, In step 5, the fuzzy logic decision tree adopts the Takagi-Sugeno model. The canopy physiological parameters obtained in step 4 are input into the fuzzy logic decision tree. The membership function of the fuzzy logic decision tree is generated by fuzzy C-means clustering. The rule base is constructed based on the expert knowledge graph. The crop growth status assessment index is generated through this decision tree. The crop growth status assessment index includes nitrogen deficiency level, water stress index and disease occurrence probability. Among them, the nitrogen deficiency level is quantitatively inverted by fusing the spectral index NDRE and MCARI through the support vector regression model.

7. The method for hyperspectral dynamic imaging monitoring of facility agriculture according to claim 1, characterized in that, Also includes: Between steps 3 and 4, a federated learning framework is set up. The federated learning framework uses the FedAvg algorithm to aggregate the model parameters of the distributed processing nodes. During the aggregation process, Laplace noise is added to protect data security through differential privacy technology. The updated model parameters of the federated learning framework are fed back to the robotic arm motion path planning model in step 1 to optimize the acquisition strategy. At the same time, they are transmitted to the fuzzy logic decision tree in step 5 to update the evaluation rules.

Images