Deep learning recognition method and system for plant growth state based on plant lights

Abstract

The present invention relates to the field of deep learning recognition technology, and discloses a deep learning recognition method and system for plant growth status based on plant lights. The method comprises: using a track-mounted multispectral camera array paired with a fixed LED plant fill light to perform multimodal data acquisition on the plant growth area, obtaining a plant image dataset and an environmental parameter dataset; performing feature extraction to obtain a plant growth feature vector; performing plant growth status recognition and fill light demand prediction based on a multi-task deep learning model with an improved Transformer architecture, obtaining plant growth stage discrimination results and fill light demand spatiotemporal distribution data; constructing an initial fill light parameter optimization model based on the plant growth stage discrimination results and fill light demand spatiotemporal distribution data, and obtaining an LED plant fill light illumination parameter sequence through deep Q-network iterative optimization. The present invention achieves efficient optimization and dynamic adjustment of fill light parameters, significantly improving the accuracy of fill light control.

Description

Technical Field

[0001] The present invention relates to the field of deep learning recognition technology, and in particular to a deep learning recognition method and system for plant growth status based on plant lamps. Background Art

[0002] Traditional LED plant lighting systems rely on pre-set lighting parameters and are difficult to adjust in real time based on plant growth status, resulting in suboptimal lighting effects and low energy efficiency. Furthermore, existing lighting control methods rely primarily on manual experience or simple environmental parameter feedback, lacking precise identification of plant growth status and accurate prediction of dynamic lighting needs. Consequently, they are unable to meet the differentiated lighting requirements of plants at different growth stages.

[0003] In recent years, deep learning technology has achieved significant breakthroughs in computer vision and intelligent control, providing new technical means for identifying plant growth states and optimizing supplemental lighting parameters. However, the application of deep learning technology to plant supplemental lighting control research still faces many challenges. The plant growth process involves multiple physiological parameters, and effectively integrating multimodal data and extracting key features remains a challenge. Furthermore, plant growth exhibits significant temporal and spatial heterogeneity, making it difficult for traditional deep learning models to simultaneously capture information from these two dimensions. Optimizing supplemental lighting parameters is a complex dynamic decision-making process, and developing efficient online learning and optimization mechanisms requires further research. Summary of the Invention

[0004] The present invention provides a deep learning recognition method and system for plant growth status based on plant lamps, which is used to achieve efficient optimization and dynamic adjustment of fill light parameters, significantly improving the accuracy of fill light control.

[0005] In a first aspect, the present invention provides a deep learning recognition method for plant growth status based on plant lamps, the deep learning recognition method for plant growth status based on plant lamps comprising:

[0006] A track-mounted multispectral camera array paired with a fixed LED plant light was used to collect multimodal data from the plant growth area, obtaining a plant image dataset and an environmental parameter dataset containing track location points.

[0007] Performing feature extraction on the plant image dataset and the environmental parameter dataset to obtain a plant growth feature vector;

[0008] The plant growth feature vector is input into a multi-task deep learning model based on an improved Transformer architecture to perform plant growth status recognition and supplementary lighting demand prediction, thereby obtaining plant growth stage discrimination results and supplementary lighting demand spatiotemporal distribution data;

[0009] An initial fill light parameter optimization model is constructed based on the plant growth stage discrimination result and the spatiotemporal distribution data of the fill light demand, and an LED plant fill light illumination parameter sequence is obtained through deep Q network iterative optimization.

[0010] In a second aspect, the present invention provides a deep learning recognition system for plant growth status based on plant lamps, the deep learning recognition system for plant growth status based on plant lamps comprising:

[0011] The acquisition module is used to collect multimodal data of the plant growth area using a track-mounted multispectral camera array equipped with a fixed LED plant fill light, obtaining a plant image dataset and an environmental parameter dataset containing track location points;

[0012] A feature extraction module is used to extract features from the plant image dataset and the environmental parameter dataset to obtain a plant growth feature vector;

[0013] A prediction module is used to input the plant growth feature vector into a multi-task deep learning model based on an improved Transformer architecture to perform plant growth status recognition and supplementary lighting demand prediction, thereby obtaining plant growth stage discrimination results and supplementary lighting demand spatiotemporal distribution data;

[0014] The iterative module is used to construct an initial fill light parameter optimization model based on the plant growth stage judgment result and the spatiotemporal distribution data of the fill light demand, and obtain the LED plant fill light illumination parameter sequence through deep Q network iterative optimization.

[0015] The third aspect of the present invention provides a deep learning recognition device for plant growth status based on plant lamps, comprising: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the deep learning recognition device for plant growth status based on plant lamps executes the above-mentioned deep learning recognition method for plant growth status based on plant lamps.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned deep learning recognition method for plant growth status based on plant lights.

[0017] In the technical solution provided by the present invention, by mounting a multi-spectral camera array on an orbital LED fill light and combining a variety of environmental sensors, multi-dimensional and omni-directional data collection of the plant growth area is achieved. An improved image processing and feature extraction algorithm is used to accurately extract features such as plant leaf morphology, temperature distribution, and photosynthesis, and a comprehensive characterization of plant growth features is obtained through feature fusion. Based on a multi-task deep learning model with an improved Transformer architecture, accurate identification of the plant growth state and accurate prediction of the fill light requirements are realized, and the model has good spatio-temporal feature capture ability. By constructing a fill light parameter optimization model through a deep Q network and combining an experience replay mechanism, efficient optimization and dynamic adjustment of the fill light parameters are achieved, significantly improving the accuracy of fill light control. An industrial bus communication and multi-channel drive control scheme is adopted to accurately execute the fill light parameters, ensuring the reliable transmission and execution of control instructions. The fill light effect is evaluated in real time through a deep learning evaluation model, and the optimization model is updated online based on the evaluation results to ensure the continuous optimization of the fill light control strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a schematic flowchart of a method for deep learning recognition of plant growth state based on a plant light provided by an embodiment of the present application;

[0020] Figure 2 It is a schematic block diagram of the structure of a system for deep learning recognition of plant growth state based on a plant light provided by an embodiment of the present application;

[0021] Figure 3 It is a schematic block diagram of the structure of a device for deep learning recognition of plant growth state based on a plant light provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0023] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may change based on actual circumstances.

[0024] It should also be understood that the terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It should be further understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0026] The following describes some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments may be combined with each other.

[0027] See also Figure 1 , Figure 1 A flow chart of a method for deep learning and identifying plant growth status based on plant lights provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the deep learning recognition method for plant growth status based on plant lights provided in the embodiment of the present application includes steps S100 to S600.

[0028] Step S100: Using a track-mounted multispectral camera array equipped with a fixed LED plant fill light to collect multimodal data from the plant growth area, obtaining a plant image dataset and an environmental parameter dataset containing track location points;

[0029] It is understandable that the execution subject of the present invention can be a deep learning recognition system for plant growth status based on plant lights, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0030] Specifically, the moving orbit of the orbital multi-spectral camera array is calibrated with an electronic map, and mapping data of the orbital coordinates is constructed through an accurate calibration process. During this process, the geometric information of the orbit, the movement path of the device, and the spatial reference points are recorded to ensure the positioning accuracy and repeatability of subsequent data collection. Based on the orbital coordinate mapping data, the acquisition positions of the multi-spectral camera array are divided, and N orbital position points are determined according to the preset orbital step size and coverage range. Each orbital position point corresponds to a specific coordinate, which is used to guide the specific acquisition position of the multi-spectral camera in space. During the multi-spectral data acquisition process, for each orbital position point, through the systematic multi-spectral data acquisition trigger control mechanism, different types of cameras are triggered in an orderly manner for synchronous acquisition. The visible light camera is responsible for collecting the apparent characteristics of plants, including information such as leaf color, morphology, and texture; the infrared camera is used to collect the temperature distribution characteristics of plants, and reflects its thermal condition and transpiration state by recording the infrared radiation intensity on the plant surface; the chlorophyll fluorescence camera is specifically used to collect the characteristics related to plant photosynthesis, and reveals the photosynthesis efficiency and light response state of plants by detecting chlorophyll fluorescence signals. The image data collected by these different types of cameras constitutes the initial multi-modal image data set. The initial data is processed for image noise reduction, and image denoising algorithms are used, such as methods based on Gaussian filtering and median filtering, to remove random noise in the images. At the same time, considering that the multi-spectral data comes from different types of cameras and there are differences in their spatial resolution and field of view, the data is spatially registered. By matching feature points and applying affine transformation or perspective transformation, different types of images are aligned in space to generate a high-quality plant image data set containing orbital position points. In terms of environmental parameter data collection, for the N orbital position points, multiple environmental data are collected through sensors arranged on the orbital device. The temperature and humidity sensors continuously monitor the environmental temperature and humidity data at each position point, recording the changes in the water vapor content and temperature in the air; the carbon dioxide sensor is used to collect the carbon dioxide concentration data in the environment, which helps to analyze plant photosynthesis and respiration; the illuminance sensor collects the environmental light data, recording the light intensity distribution in the plant growth area. The original environmental parameter data is subjected to outlier detection, and through statistical methods or threshold-based detection mechanisms, abnormal data that does not conform to the actual situation is identified and removed. The data is standardized to ensure that different parameters have a unified scale, and an environmental parameter data set is obtained.

[0031] Step S200: Extract features from the plant image data set and the environmental parameter data set to obtain a plant growth feature vector;

[0032] Specifically, perform image segmentation preprocessing on the visible light images in the plant image dataset. Use deep learning segmentation algorithms (such as U-Net or Mask R-CNN) to perform pixel-level segmentation on the plant images, generating an image feature mapping array containing the main structural features of the plants. Based on the image feature mapping array, focus on the leaf area of the plants, extract the leaf area through the segmentation algorithm, and generate mask data for the leaf area. Perform morphological processing and edge detection on the mask data of the plant leaf area. Remove noise and enhance boundary details by applying morphological algorithms such as dilation, erosion, opening, and closing operations. Then, combine Canny edge detection or Sobel operator to extract the edge contour information of the plant leaves, generating a set of contour feature points. Based on the set of contour feature points, calculate the key morphological parameters of the plants, including the leaf area, number, plant height, and crown width. The leaf area is calculated by multiplying the number of pixel points in the mask data by the spatial resolution, and the number of leaves is counted through the connected component analysis algorithm; while the plant height and crown width need to be measured by identifying the upper and lower boundaries and left and right boundaries of the plants in combination with the scale information marked in the image. These calculation results constitute the plant morphological feature data, reflecting the external geometric morphological information of the plants. At the same time, for the infrared image data in the plant image dataset, it is necessary to extract and analyze the temperature distribution characteristics of the plants. Generate a temperature distribution feature mapping matrix through the infrared imaging data, and each pixel point in the mapping matrix represents the temperature value distribution on the plant surface. Perform statistical analysis on the temperature distribution feature mapping matrix, such as calculating the mean, maximum, minimum, and standard deviation of the temperature, to extract the feature data reflecting the overall and local heat distribution conditions of the plants, revealing the transpiration, environmental adaptability, and potential heat stress conditions of the plants. For the chlorophyll fluorescence images obtained from the plant image dataset, extract features for their photosynthesis characteristics. Generate a photosynthesis feature mapping matrix by analyzing the chlorophyll fluorescence signal, and each pixel point in the matrix represents the photosynthesis efficiency of the plant leaves in different spatial regions. Combine the spatial feature analysis of photosynthesis to extract the difference data of the photosynthesis efficiency of different parts of the plants, such as analyzing the light response characteristics and the distribution pattern of the fluorescence signal strength at the top and bottom of the leaves. These feature data reveal the photosynthetic capacity of the plants and their adaptation to the light environment, constituting the plant photosynthetic feature data. Extract the temporal features of the temperature data, humidity data, carbon dioxide concentration data, and illuminance data in the environmental parameter dataset. Use temporal analysis methods (such as short-time Fourier transform or long short-term memory network, LSTM) to extract periodic, trend, and mutation features from these time series data. For example, identify the daily and nightly variation patterns of the light intensity, the long-term fluctuation patterns of the temperature and humidity, and the sudden increase phenomenon of the carbon dioxide concentration. The extracted environmental feature data can help understand the dynamic response of the plants in a specific growth environment.Feature - encode and fuse the plant morphological feature data, plant temperature distribution feature data, plant photosynthetic feature data, and environmental feature data to generate a plant growth feature vector. During the fusion process, the method of feature embedding is adopted to map data of different dimensions into the same vector space, and through weighted encoding, attention mechanism, or other feature fusion algorithms, the information association between multi - modal data is fully reflected.

[0033] Step S300: Input the plant growth feature vector into a multi - task deep learning model based on an improved Transformer architecture for plant growth state recognition and supplementary lighting requirement prediction, and obtain the plant growth stage discrimination result and the spatio - temporal distribution data of supplementary lighting requirements.

[0034] Specifically, the plant growth feature vector is input into the multi-head attention encoder in the multi-task deep learning model based on the improved Transformer architecture for parallel processing. The multi-head attention encoder contains 8 attention heads, and each attention head uses query vectors, key vectors, and value vectors of 64 dimensions. These vectors calculate the feature weights through the scaled dot-product attention mechanism. Through this mechanism, the model can identify important patterns in the input feature vector and dynamically adjust the weights. At the same time, the sequential information of the input feature vector is explicitly embedded into the encoder by combining sinusoidal positional encoding to generate an initial feature representation matrix. The initial feature representation matrix is input into the three-layer cascaded Transformer encoding block in the multi-task deep learning model for feature enhancement. Each layer of the Transformer encoding block consists of a 512-dimensional multi-head self-attention layer, a 2048-dimensional feed-forward neural network layer, and a LayerNorm normalization layer. In the multi-head self-attention layer, the model captures the global dependencies between features by parallel computing multiple attention mechanisms, while the feed-forward neural network layer extracts and processes feature patterns through non-linear activation functions to enhance the feature representation ability. The LayerNorm normalization layer is used to normalize the output of each layer to ensure numerical stability during the training process. After three-layer cascaded processing, the features are further extracted and integrated to obtain an enhanced feature matrix, which contains the deep information in the plant growth feature vector. The enhanced feature matrix is input into the feature splitting module in the multi-task deep learning model for task separation. The feature splitting module accurately divides the features and sends the task-related information into the growth state recognition branch and the supplementary light demand prediction branch respectively to form a dual-task feature representation. In the growth state recognition branch, the growth state features in the dual-task feature representation are input into a fully-connected neural network, which contains 4 hidden layers with hidden layer dimensions of 512, 256, 128, and 64 respectively. Each layer uses the ReLU activation function for non-linear mapping to extract the deep features of the plant growth state. The probability scores of the growth stages are calculated through the output layer to generate a multi-class probability distribution, and the Softmax function is used to normalize the probabilities to obtain the discrimination result of the plant growth stage, accurately describing the current growth stage of the plant. At the same time, the features of the supplementary light demand prediction branch in the dual-task feature representation are input into the spatio-temporal attention module, which combines the spatial attention map and the temporal attention vector to capture the dynamic change patterns of the supplementary light demand in the spatial and temporal dimensions. The spatial attention map dynamically assigns attention weights by focusing on specific spatial regions, such as the key parts of the plant growth state, while the temporal attention vector captures the temporal change trend of the light demand by analyzing the features at different time steps. The spatio-temporal attention module adopts a gated fusion mechanism and selectively fuses the spatial and temporal features through a gating function to generate a supplementary light demand feature map.Perform feature extraction operations on the supplementary light demand feature map, extract key supplementary light demand information from it, and generate a supplementary light demand prediction tensor. Through the reconstruction of this tensor, it is transformed into a data format with a clear spatio-temporal distribution structure to obtain the spatio-temporal distribution data of the supplementary light demand.

[0035] Step S400: Construct an initial supplementary light parameter optimization model based on the plant growth stage discrimination result and the spatio-temporal distribution data of the supplementary light demand, and iteratively optimize it through a deep Q-network to obtain a sequence of LED plant supplementary light illumination parameters.

[0036] Specifically, the plant growth stage discrimination results and the spatio-temporal distribution data of the supplementary lighting requirements are input into the state encoder, and the state encoder encodes the input information into a high-dimensional supplementary lighting environment state vector. Feature partitioning is performed on the supplementary lighting environment state vector to construct an action space. The action space includes three key dimensions: light intensity, spectral ratio, and irradiation angle, and each dimension directly affects the supplementary lighting effect and energy consumption efficiency. To facilitate model processing, this multi-dimensional continuous action space is discretized to generate a discrete action matrix. Each row of the action matrix represents a possible combination of supplementary lighting parameters, forming a complete set of supplementary lighting parameter actions. Through this discretized action set, the supplementary lighting problem is transformed into a tractable reinforcement learning task. Based on the supplementary lighting environment state vector and the supplementary lighting parameter action set, a deep Q-network is constructed to optimize the supplementary lighting strategy. The deep Q-network adopts a two-stream architecture, where one branch calculates the state value function of the supplementary lighting action, and the other branch calculates the advantage function of the specific action. The two are combined to evaluate the Q-value of each action, forming an initial supplementary lighting parameter optimization model. At this stage, the network finds the optimal parameter selection under different supplementary lighting environments through the mechanism of reinforcement learning. In the specific optimization process, for the supplementary lighting environment state vector at each orbital position point, the deep Q-network calculates the Q-values of different supplementary lighting parameter actions. The Q-value represents the total reward obtained after selecting a certain supplementary lighting parameter action in the current state. To balance exploring new strategies and exploiting existing strategies, an ε-greedy strategy is used to select the optimal action. This strategy selects random actions for exploration with a certain probability and selects the action with the highest current Q-value in the remaining cases. In this way, the initial supplementary lighting parameter combination for each orbital position point is obtained. The initial supplementary lighting parameter combination is stored in the experience replay buffer, and through the mechanism of experience replay, batches of data are randomly sampled from the buffer at regular intervals for training the deep Q-network. During training, the temporal difference algorithm is used to calculate the target Q-value. The weight parameters of the network are updated by calculating the gradient, enabling the model to gradually learn the ability to optimize the supplementary lighting parameters. The updated deep Q-network outputs the target supplementary lighting parameter combination. Due to the discontinuity problem caused by discretization in the action parameters between orbital position points, these parameters are smoothed. The bilinear interpolation method is used to make the parameters continuous in time series and space between orbital position points, generating a smooth supplementary lighting parameter curve to ensure the continuity and stability of the lighting conditions during the supplementary lighting process. The continuous supplementary lighting parameter curve is reorganized according to the orbital position points, and the light intensity, spectral ratio, and irradiation angle are normalized to ensure that the parameters are within the same numerical range and avoid affecting subsequent applications due to scale differences. Finally, an LED plant supplementary lighting parameter sequence is obtained, which can dynamically adjust the supplementary lighting strategy according to time and orbital position changes to meet the real-time needs of plant growth while maximizing energy utilization efficiency and plant growth benefits.

[0037] Perform protocol conversion on the illumination parameter sequence of the LED plant supplementary light, and send the control parameters of illumination intensity, spectral ratio, and irradiation angle to the LED supplementary light controller through the industrial bus to generate corresponding supplementary light control instructions. In the supplementary light control instructions, the illumination intensity parameter is converted into a PWM duty cycle modulation signal, and this signal adjusts the output power of LEDs in different bands through four constant current sources in the LED drive circuit to achieve precise control of the target illumination intensity. To achieve dynamic adjustment of different spectra, calculate the drive current ratios of red, blue, and far-red LEDs according to the spectral ratio parameters in the supplementary light control instructions. Precisely adjust the current values of LEDs in each band through a multi-channel LED drive circuit to make the LED light source output according to the target spectral ratio and achieve the response of plants to specific spectral requirements. The adjustment of the irradiation angle is based on the pitch angle parameters in the supplementary light control instructions. After analyzing these parameters, the stepping motor control system drives the pitch angle adjustment mechanism of the LED supplementary light. The stepping motor has high control precision and can ensure that the LED supplementary light realizes the adjustment of the target irradiation angle to provide directional supplementary light for different plant areas. This movement trajectory can cover each track position point in the plant growth area to achieve comprehensive supplementary light, and at the same time optimize the energy use efficiency through the control of the supplementary light time at the track position points. To ensure the real-time monitoring and dynamic adjustment of the supplementary light system, use voltage sensors, current sensors, and temperature sensors to collect the electrical parameters of the LED supplementary light, and combine spectral sensors to collect the spectral output data of the LED supplementary light to construct the real-time working state data of the LED supplementary light. Combine the plant growth state indicators, and evaluate the real-time working state data and plant growth indicators through deep learning technology to calculate the supplementary light effect score. The evaluation model is based on the matching degree between the plant growth characteristic data and the actual light output, and combines the change trend of the growth state to comprehensively generate the supplementary light effect score. If the supplementary light effect score is lower than the set preset threshold, it means that the current supplementary light parameters do not meet the optimal requirements of plant growth. At this time, perform an online update on the initial supplementary light parameter optimization model. During the online update process, use the collected real-time data as new input samples, and by dynamically adjusting the parameters of the deep Q network, enable the model to re-optimize the supplementary light strategy and generate a target supplementary light parameter optimization model that more meets the actual requirements.

[0038] Group the real-time working state data according to the orbital position points to ensure that each group of data can reflect the supplementary lighting state and environmental conditions at a specific position point. The grouped data includes multiple parameters such as light intensity, spectral ratio, irradiation angle, voltage and current, and temperature. Standardize these data to eliminate the differences in dimension and scale, and obtain the supplementary lighting state feature sequence, which describes the operating state of the supplementary lighting lamp at each orbital position point. Based on the supplementary lighting state feature sequence, construct the supplementary lighting execution feature vector. The supplementary lighting execution feature vector is a high-dimensional representation of the actual execution result of the supplementary lighting system, which can capture the correlation between the supplementary lighting state and plant growth. At the same time, in order to quantify the plant growth effect, comprehensively evaluate the plant growth indicators. By analyzing the three key indicators of the plant growth rate change rate, chlorophyll content change rate, and photosynthesis efficiency change rate, and performing weighted calculations on them, a comprehensive plant growth evaluation index is generated. During the weighted calculation process, the weights of different indicators are adjusted according to actual needs. For example, the photosynthesis efficiency has a higher weight in some application scenarios to highlight its impact on plant health. Calculate the supplementary lighting effect score according to the supplementary lighting execution feature vector and the comprehensive plant growth evaluation index. The supplementary lighting effect score reflects the degree of influence of the supplementary lighting parameters on plant growth in actual execution. By comparing this score with a preset score threshold, it is judged whether the current supplementary lighting parameters need to be adjusted. When the supplementary lighting effect score is lower than the preset threshold, it indicates that the existing supplementary lighting strategy fails to achieve the desired effect. At this time, the initial supplementary lighting parameter optimization model is updated online. The first step of the online update is to calculate the gradient value of the supplementary lighting effect score with respect to the parameters of the initial supplementary lighting parameter optimization model. The calculation of the gradient value is based on the backpropagation process of the loss function, which accurately reflects the direction and magnitude of the current parameter adjustment. Apply the Adam optimizer to optimize the calculated parameter update gradient. The Adam optimizer is an optimization algorithm with an adaptive learning rate, which can effectively combine the first-order momentum and second-order momentum information to quickly find the parameter optimization direction. Under the processing of the Adam optimizer, a parameter optimization direction is generated, which is used to guide the adjustment of the weights of the deep Q network of the initial supplementary lighting parameter optimization model. As the core model for supplementary lighting parameter optimization, the weight update of the deep Q network can directly affect the model's ability to evaluate the value of the supplementary lighting environment state, thereby adjusting the supplementary lighting strategy. After the weights of the deep Q network are optimized, the updated model parameters are reloaded into the deep Q network, and the parameters of its value evaluation network are updated. The value evaluation network is a key component of the deep Q network. By learning the value relationship between the environmental state and the supplementary lighting parameter actions, it provides an accurate reference for the supplementary lighting decision at each orbital position point. The value evaluation network after online update has higher decision-making accuracy and environmental adaptability, and finally generates the target supplementary lighting parameter optimization model.

[0039] In the embodiments of the present invention, by mounting a multi-spectral camera array on an orbital LED fill light and combining multiple environmental sensors, multi-dimensional and omni-directional data collection of the plant growth area is achieved. An improved image processing and feature extraction algorithm is used to accurately extract features such as plant leaf morphology, temperature distribution, and photosynthesis, and comprehensive plant growth feature representations are obtained through feature fusion. Based on a multi-task deep learning model with an improved Transformer architecture, accurate identification of the plant growth state and accurate prediction of the fill light requirements are realized, and the model has good spatio-temporal feature capture capabilities. By constructing a fill light parameter optimization model through a deep Q-network and combining an experience replay mechanism, efficient optimization and dynamic adjustment of the fill light parameters are achieved, significantly improving the accuracy of fill light control. An industrial bus communication and multi-channel drive control scheme is adopted to accurately execute the fill light parameters, ensuring the reliable transmission and execution of control instructions. The fill light effect is evaluated in real time through a deep learning evaluation model, and the optimization model is updated online based on the evaluation results to ensure the continuous optimization of the fill light control strategy.

[0040] In a specific embodiment, the process of executing step S100 may specifically include the following steps:

[0041] Perform electronic map calibration on the moving track of the orbital multi-spectral camera array to obtain track coordinate mapping data, and divide the acquisition positions of the multi-spectral camera array based on the track coordinate mapping data to obtain N track position points;

[0042] Perform trigger control for multi-spectral data acquisition on the N track position points, collect plant apparent features through a visible light camera, collect plant temperature distribution features through an infrared camera, and collect plant photosynthetic features through a chlorophyll fluorescence camera to obtain original multi-modal image data, and perform image denoising and spatial registration on the original multi-modal image data to obtain a plant image dataset containing track position points;

[0043] Collect environmental data for the N track position points, collect environmental temperature and humidity data through a temperature and humidity sensor, collect carbon dioxide concentration data through a carbon dioxide sensor, and collect environmental light data through an illuminance sensor to obtain original environmental parameter data, and perform outlier detection and data standardization processing on the original environmental parameter data to obtain an environmental parameter dataset.

[0044] Specifically, perform electronic map calibration on the moving track of the orbital multi-spectral camera array. Measure and record the spatial coordinates of each acquisition position on the track. By arranging multiple reference points with known positions on the track, such as using a high-precision laser rangefinder or total station, collect the key position coordinates of the track and establish a global coordinate system for the track. These coordinate data are represented as , where are respectively the track points Position in three-dimensional space. The discrete points are fitted by a curve fitting or interpolation algorithm (such as spline interpolation) to generate a continuous path of the entire orbit. The electronic map of the orbit is represented in the form of a function , where is a parametric variable on the orbit (such as distance) and is mapped to the three-dimensional coordinate points on the orbit. Based on the calibration data of the electronic map of the orbit, the orbit is segmented into uniform or adaptive orbit position points to guide the acquisition task of the multispectral camera array. For example, assume the total length of the orbit is . If it needs to be evenly divided into position points, then the interval of each position point is . The coordinates of the position points are represented as , where is the orbit parameter of the th position point. For multispectral data acquisition, a multispectral camera array is used to accurately image the plants. At each orbit position point, the visible light camera, infrared camera, and chlorophyll fluorescence camera are synchronously controlled through a trigger control system to collect the apparent characteristics, temperature distribution characteristics, and photosynthesis characteristics of the plants respectively. The visible light camera is used to capture the color, shape, and texture characteristics of the plants, and its imaging data is represented in matrix form , where and are the height and width of the image, and is the number of color channels. The infrared camera collects the thermal radiation data of the plants and obtains the temperature distribution matrix through thermal imaging, where each pixel represents the temperature on the plant surface. The chlorophyll fluorescence camera is used to detect the fluorescence signal related to plant photosynthesis, and its output is the fluorescence intensity matrix . Since the collected multimodal image data is affected by noise, the original data is denoised. The Gaussian filtering algorithm is used, and the formula is:

[0045] ;

[0046] where is the original image, is the pixel window, is the standard deviation of the Gaussian kernel, and the image is smoothed by this method to remove random noise. Since the viewing angles and resolutions of the multispectral camera array are different, spatial registration is performed, and the transformation matrix between the cameras is calculated through feature point matching (such as the SIFT algorithm), and the images are aligned using affine transformation:

[0047] ;

[0048] where is the pixel coordinate of the target image, is the pixel coordinate of the original image. After completing the image data processing, environmental data is collected. The temperature and humidity data of the track position points are collected through temperature and humidity sensors, and its output is a temperature sequence and a humidity sequence . The carbon dioxide concentration data recorded by the carbon dioxide sensor is represented as , while the light intensity recorded by the illuminance sensor is . Outlier detection is performed on the original environmental parameter data, and the three - standard - deviation method is used to define the outlier range:

[0049] ;

[0050] where is the parameter mean, is the parameter standard deviation. For data points outside the range, they are replaced with the mean or interpolation value. To unify the data scale, all environmental parameters are standardized, and the formula is:

[0051] ;

[0052] where is the original data value, is the mean, is the standard deviation. The standardized data constitutes the environmental parameter dataset , which is combined with the plant image dataset to provide high - quality multi - modal data input for subsequent analysis.

[0053] In a specific embodiment, the process of executing step S200 may specifically include the following steps:

[0054] Perform image segmentation pre - processing on the visible light images in the plant image dataset to obtain an image feature mapping array, and perform plant leaf area segmentation on the image feature mapping array to obtain plant leaf area mask data;

[0055] Perform morphological processing and edge detection on the plant leaf area mask data to obtain a set of contour feature points, and based on the set of contour feature points, calculate the plant morphological parameters such as leaf area, number of leaves, plant height, and crown width to obtain plant morphological feature data;

[0056] Perform temperature distribution feature calculation on the infrared image data in the plant image dataset to obtain a temperature distribution feature mapping matrix, and perform statistical analysis on the temperature distribution feature mapping matrix to obtain plant temperature distribution feature data;

[0057] Extract photosynthesis features from chlorophyll fluorescence images in plant image datasets to obtain photosynthesis feature mapping matrix, and perform spatial feature analysis on the photosynthesis feature mapping matrix to obtain plant photosynthesis feature data;

[0058] Perform time series feature extraction on the temperature data, humidity data, carbon dioxide concentration data, and illumination data in the environmental parameter data set to obtain environmental feature data;

[0059] The plant morphological characteristic data, plant temperature distribution characteristic data, plant photosynthetic characteristic data and environmental characteristic data are fused by feature coding to obtain a plant growth characteristic vector.

[0060] Specifically, the visible light images in the plant image dataset are preprocessed by image segmentation. By using a deep learning segmentation algorithm (such as U-Net or DeepLab), the original visible light images are Divided into different areas, and are the height and width of the image, is the number of color channels. The goal of segmentation is to extract the main part of the plant while excluding the background and irrelevant areas. The result of image segmentation is represented as an image feature map array , where the value of each pixel represents the category it belongs to (such as leaf, stem or background). The image feature map array is processed by region growing or connected domain analysis algorithm, focusing on plant leaf region segmentation to generate plant leaf region mask data ,in Represents pixels Belongs to the leaf area, The mask data of the plant leaf area is morphologically processed to remove isolated noise points through dilation and erosion operations, while filling small holes in the leaf area. The mask data after morphological processing is used for edge detection, and the boundary points of the leaf are extracted using Sobel or Canny operators to generate a set of contour feature points. ,in For the The coordinates of the contour feature points, is the total number of boundary points. Based on these contour feature points, the geometric parameters of the leaf are calculated. For example, the leaf area It is obtained by calculating the total number of all leaf pixels in the mask data and multiplying it by the area of a single pixel. The formula is:

[0061] ;

[0062] in, is the area of a single pixel in the actual scene. The number of leaves is calculated by connecting domain analysis to count the leaf area. Then, by calculating the vertical distance from the top to the bottom of the plant and the crown width which is the horizontal distance between the left and right boundaries of the plant, these parameters are combined to form the morphological feature data of the plant. Calculate the temperature distribution characteristics of the infrared image data in the plant image dataset. Each pixel value in the infrared image represents the temperature on the surface of the plant. Extract the key characteristics of the temperature distribution through statistical analysis, including the average temperature , the maximum temperature , the minimum temperature and the standard deviation . The specific calculation formulas are as follows:

[0063] ;

[0064] ;

[0065] These characteristic data reflect the thermal distribution state of the plant, reveal its transpiration and potential heat stress conditions, and constitute the temperature distribution characteristic data of the plant. For the chlorophyll fluorescence images in the plant image dataset, use specific photosynthesis parameter extraction techniques to generate a photosynthesis characteristic mapping matrix . The fluorescence signal intensity reflects the activity degree of plant photosynthesis. Commonly used fluorescence parameters include the maximum fluorescence value , the minimum fluorescence value and the photochemical efficiency . The calculation formula for the photochemical efficiency is:

[0066] ;

[0067] where and represent the maximum and minimum values of the fluorescence signal respectively. Through spatial feature analysis of the photosynthesis characteristic mapping matrix, identify the distribution pattern of plant photosynthesis efficiency, and extract key features as the photosynthetic characteristic data of the plant. At the same time, extract the temporal features of the temperature, humidity, carbon dioxide concentration and illuminance data in the environmental parameter dataset. Assume that the time series of these parameters are , where is the number of time steps. Extract the frequency domain features through short-time Fourier transform, and calculate the moving average and change rate using a sliding window to obtain the characteristic data reflecting the dynamic changes of the environment. Feature encode and fuse the above-extracted plant morphological feature data, plant temperature distribution characteristic data, plant photosynthetic characteristic data and environmental characteristic data to generate a plant growth feature vector , where is the dimension of the fused feature vector. The fusion method adopts feature concatenation, attention mechanism or the embedding layer of a deep learning model to ensure the mutual correlation of multi-modal features.

[0068] In a specific embodiment, the process of executing step S300 may specifically include the following steps:

[0069] Input the plant growth feature vector into the multi-head attention encoder in the multi-task deep learning model based on the improved Transformer architecture for parallel processing. The multi-head attention encoder contains 8 attention heads, and each attention head uses 64-dimensional query vectors, key vectors, and value vectors. Calculate the feature weights through the scaled dot-product attention mechanism and combine with the sine position encoding to obtain the initial feature representation matrix;

[0070] Input the initial feature representation matrix into 3 cascaded Transformer encoding blocks in the multi-task deep learning model. Each layer of the Transformer encoding block contains a 512-dimensional multi-head self-attention layer, a 2048-dimensional feed-forward neural network layer, and a LayerNorm normalization layer to obtain the enhanced feature matrix;

[0071] Input the enhanced feature matrix into the feature splitting module in the multi-task deep learning model for task separation, and obtain the dual-task feature representation through the growth state recognition branch and the light supplement demand prediction branch;

[0072] Input the growth state recognition branch features in the dual-task feature representation into the fully connected neural network in the multi-task deep learning model. The fully connected neural network contains 4 hidden layers, and the dimensions of the hidden layers are 512, 256, 128, and 64 respectively to obtain the growth state feature vector, and calculate the probability score of the growth stage according to the growth state feature vector to obtain the discrimination result of the plant growth stage;

[0073] Input the light supplement demand prediction branch features in the dual-task feature representation into the spatio-temporal attention module in the multi-task deep learning model, calculate the spatial attention map and the temporal attention vector, and adopt the gated fusion mechanism to integrate the spatio-temporal dimension information to obtain the light supplement demand feature map;

[0074] Extract features from the light supplement demand feature map to obtain the light supplement demand prediction tensor, and reconstruct the light supplement demand prediction tensor into a spatio-temporal distribution structure to obtain the spatio-temporal distribution data of the light supplement demand.

[0075] Specifically, the plant growth feature vector is represented as a matrix with features, where the dimension of each feature is Take The input is processed in parallel by a multi - head attention encoder, which consists of 8 attention heads. Each attention head independently calculates the query vector , the key vector and the value vector . These vectors are generated from the input feature vectors through linear transformations, and the transformation formulas are:

[0076] ;

[0077] where are the weight matrices of the query, key, and value vectors. The feature weights are calculated through the scaled dot - product attention mechanism, and the formula is:

[0078] ;

[0079] where is the scaling factor to prevent the softmax gradient from vanishing due to an overly large dot - product value, is the dimension of the key vector. The multi - head attention is achieved by concatenating the outputs of the 8 attention heads and performing another linear transformation, resulting in the initial feature representation matrix . To capture the sequential information of the input features, sine - wave positional encoding is added to , and the positional encoding formula is:

[0080] ;

[0081] where pos is the position index, is the feature dimension index. The initial feature representation matrix is input into 3 cascaded Transformer encoding blocks. Each layer contains a multi - head self - attention layer with 512 dimensions, a feed - forward neural network layer with 2048 dimensions, and a LayerNorm normalization layer. In each layer, the global feature relationships are captured through the multi - head self - attention layer, then the non - linear features are extracted through the feed - forward neural network, and finally LayerNorm is applied to normalize the output of each layer to ensure the numerical stability during the training process. After 3 - layer processing, an enhanced feature matrix is obtained. The enhanced feature matrix is input into a feature splitting module for task separation, separated into the growth - state recognition branch feature and the supplementary - light - demand prediction branch feature . The growth - state recognition branch feature is input into a fully - connected neural network, which contains 4 hidden layers with dimensions of 512, 256, 128, and 64 respectively. Non - linear transformations are performed between layers through the ReLU activation function. The network output is the growth - state feature vector , and finally the probability scores of the growth stages are calculated through the Softmax layer, where is the number of categories in the growth stage, is defined as:

[0082] ;

[0083] where is the network output score for the corresponding category. The supplementary lighting demand prediction branch feature is input into the spatio-temporal attention module to calculate the spatial attention map and the temporal attention vector where are the spatial resolution and the number of time steps respectively. The spatial attention is calculated by a convolutional neural network, and the temporal attention extracts the time series features through LSTM. After combining the two, a gated fusion mechanism is adopted to generate the supplementary lighting demand feature map and the gated fusion formula is:

[0084] ;

[0085] where is the Sigmoid activation function, is the fusion weight. The supplementary lighting demand feature map is processed through a convolutional layer and upsampling to extract the supplementary lighting demand prediction tensor . Through reconstruction, is converted into the supplementary lighting demand data with a spatio-temporal distribution structure, where the value of each point represents the supplementary lighting demand intensity at the corresponding spatial position and time point.

[0086] In a specific embodiment, the process of executing step S400 may specifically include the following steps:

[0087] Input the plant growth stage discrimination result and the supplementary lighting demand distribution data into the state encoder for state space encoding to obtain the supplementary lighting environment state vector;

[0088] Perform feature partitioning on the supplementary lighting environment state vector, construct an action space including light intensity, spectral ratio, irradiation angle, etc., and discretize the action space into an action matrix to obtain the supplementary lighting parameter action set;

[0089] Based on the supplementary lighting environment state vector and the supplementary lighting parameter action set, construct a deep Q network, and the deep Q network includes a value evaluation network with a two-stream architecture to obtain an initial supplementary lighting parameter optimization model;

[0090] For the supplementary lighting environment state vector within the irradiation range of the LED plant supplementary light, calculate the Q values of different supplementary lighting parameter actions through the deep Q network, and select the optimal action using the ε-greedy strategy to obtain the initial supplementary lighting parameter combination;

[0091] Input the initial supplementary lighting parameter combination into the experience replay buffer pool, randomly sample batch data from the experience replay buffer pool, calculate the target Q value through the temporal difference algorithm, and obtain the parameter update gradient;

[0092] Optimize and update the weight parameters of the deep Q network according to the parameter update gradient to obtain the target supplementary lighting parameter combination;

[0093] Normalize the light intensity, spectral ratio, and irradiation angle of the target supplementary lighting parameter combination to obtain the LED plant supplementary lighting parameter sequence.

[0094] Specifically, input the plant growth stage discrimination result and the spatio-temporal distribution data of the supplementary lighting demand into the state encoder. By encoding these data, convert them into a supplementary lighting environment state vector suitable for the deep reinforcement learning model. Assume that the plant growth stage discrimination result is a classification result vector , where is the number of categories of the growth stage, each value in represents the probability that the plant is in a certain stage. At the same time, the spatio-temporal distribution data of the supplementary lighting demand is represented as a three-dimensional tensor , where , and represent the horizontal and vertical resolutions of the spatial position and the time step respectively. The state encoder encodes these input data through a multi-layer perceptron to generate the supplementary lighting environment state vector , and the formula is:

[0095] ;

[0096] where Flatten flattens the three-dimensional tensor into a one-dimensional vector, and are the weight matrix and bias vector of the state encoder respectively, is the activation function (such as ReLU). Divide the supplementary lighting environment state vector according to the orbital position points, and each position point corresponds to a local state vector , where represents the th position point on the orbit. Based on the state vector of each position point, construct an action space, and the action space includes the light intensity , spectral ratio[[ID=5l]] (corresponding to the ratio of red light, blue light, and far red light), and irradiation angle . Through discretization, the value of each parameter is divided into several discrete points. Combine all parameter combinations to form an action matrix , where is the number of actions, is the action dimension. A Deep Q-Network (DQN) is constructed based on the fill light environmental state vector and the action matrix. The DQN adopts a two-stream architecture, including a state value stream and an action advantage stream . The Q value is calculated by the following formula:

[0097] ;

[0098] where is the th action, is all actions. The DQN uses this as the initial fill light parameter optimization model, and selects actions through the -greedy strategy, that is, randomly selects actions with probability , and selects the action with the largest current Q value with probability , to obtain the initial fill light parameter combination. The initial fill light parameter combination is stored in the experience replay buffer . During the training process, a batch of data is randomly sampled from , where is the reward value, is the next state. The target Q value is updated through the Temporal Difference (TD) algorithm:

[0099] ;

[0100] where is the discount factor, indicating the importance of future rewards. The loss function is calculated as:

[0101] ;

[0102] The loss function is minimized by the Adam optimizer to obtain the parameter update gradient and optimize the weight parameters of the DQN. The optimized target fill light parameter combination is organized into a continuous curve of orbital position points. The bilinear interpolation method is used to smooth the parameters between the orbital position points. Let the parameters of two adjacent points on the orbit be and , and the interpolation parameter is:

[0103] ;

[0104] where is the interpolation point, is the time index of the orbital point. The fill light parameter curve is normalized to scale the values of light intensity, spectral ratio, and irradiation angle to a unified range, generating a sequence of LED plant fill light parameters.

[0105] In a specific embodiment, the method for deep learning recognition of plant growth status based on plant lights further includes the following steps:

[0106] Perform protocol conversion on the illumination parameter sequence of the LED plant supplementary light, and send the control parameters of illumination intensity, spectral ratio, and irradiation angle to the LED supplementary light controller through the industrial bus to obtain a supplementary light control instruction;

[0107] Convert the illumination intensity parameter in the supplementary light control instruction into a PWM duty cycle modulation signal, and control the output power of different bands of LEDs through the four-channel constant current source of the LED drive circuit to obtain the target illumination intensity;

[0108] Calculate the drive current ratio of red, blue, and far-red LEDs according to the spectral ratio parameter in the supplementary light control instruction, and adjust the current values of LEDs in each band through the multi-channel LED drive circuit to obtain the target spectral ratio;

[0109] Analyze the irradiation angle parameter in the supplementary light control instruction, and drive the pitch angle adjustment mechanism of the LED supplementary light through the stepper motor control system to obtain the target irradiation angle;

[0110] Collect the electrical parameters of the LED supplementary light through a voltage sensor, a current sensor, and a temperature sensor, and collect the spectral output data of the LED supplementary light through a spectral sensor to obtain the real-time working state data of the LED supplementary light;

[0111] Perform deep learning evaluation on the real-time working state data and plant growth indicators, calculate the supplementary light effect score, and when the supplementary light effect score is lower than the preset threshold, perform online parameter update on the initial supplementary light parameter optimization model to obtain the target supplementary light parameter optimization model.

[0112] Specifically, for the illumination intensity , spectral ratio , and irradiation angle in the illumination parameter sequence of the LED plant supplementary light, perform protocol conversion, convert the parameter format into a frame structure that conforms to the industrial bus communication standard. Transmit it to the LED supplementary light controller through the industrial bus and generate a supplementary light control instruction. For the illumination intensity parameter , the controller converts it into a PWM (pulse width modulation) duty cycle signal. The duty cycle DutyCycle of the PWM signal is proportional to the target illumination intensity, and its calculation formula is:

[0113] ;

[0114] Where is the maximum illumination intensity of the LED light source. The PWM signal passes through the four-channel constant current source module of the LED driving circuit to control the output power of the red, blue, far-red, and white LEDs respectively, generating the target illumination intensity . The spectral ratio parameter is used to adjust the driving current ratio of the LEDs in each band. Assuming the total current is , then the driving current for each band , and can be calculated as follows:

[0115] ;

[0116] By adjusting the driving current of each band through the multi-channel LED driving circuit respectively, it is ensured that the spectral output conforms to the target ratio. For the irradiation angle parameter , after the controller analyzes the parameter, it adjusts the pitch angle of the LED supplementary light through the stepping motor control system. The rotation angle of the stepping motor is calculated according to the following formula:

[0117] ;

[0118] where is the single-step angle of the stepping motor, and StepsPerRevolution is the number of steps per revolution of the stepping motor. In order to monitor the system operation status in real time, electrical parameters , and of the LED supplementary light are collected by using voltage sensors, current sensors, and temperature sensors. At the same time, the spectral sensor records the real-time spectral output data . These data constitute the real-time working status data of the LED supplementary light. Deep learning evaluation is performed on the real-time working status data and the plant growth indicators (such as chlorophyll content, photosynthesis efficiency) to calculate the supplementary light effect score . Online update is based on deep reinforcement learning, and the weight parameters of the deep Q network are optimized by collecting new state-action-reward data. The optimized target supplementary light parameters are reloaded into the controller, generating an updated dynamic supplementary light parameter curve and then sent to the system for execution again to achieve dynamic and closed-loop supplementary light control.

[0119] In a specific embodiment, the process of performing deep learning evaluation on the real-time working status data and the plant growth indicators, calculating the supplementary light effect score, and when the supplementary light effect score is lower than the preset threshold, performing online parameter update on the initial supplementary light parameter optimization model to obtain the target supplementary light parameter optimization model may specifically include the following steps:

[0120] Group the real-time working state data to obtain multiple groups of working state data, and standardize the light intensity, spectral ratio, irradiation angle, voltage and current, and temperature parameters in each group of working state data to obtain a supplementary light state feature sequence;

[0121] Construct a supplementary light execution feature vector based on the supplementary light state feature sequence. At the same time, perform weighted calculation on the growth rate change rate, chlorophyll content change rate, and photosynthesis efficiency change rate in the plant growth indicators to obtain a comprehensive plant growth evaluation index;

[0122] Calculate the supplementary light effect score according to the supplementary light execution feature vector and the comprehensive plant growth evaluation index, and compare the supplementary light effect score with a preset score threshold. When the supplementary light effect score is lower than the preset threshold, calculate the gradient value of the supplementary light effect score with respect to the initial supplementary light parameter optimization model parameters to obtain a parameter update gradient;

[0123] Perform Adam optimizer processing on the parameter update gradient to obtain a parameter optimization direction, and perform online update on the weights of the deep Q network of the initial supplementary light parameter optimization model based on the parameter optimization direction to obtain updated model parameters;

[0124] Reload the updated model parameters into the deep Q network, and update the parameters of the value evaluation network of the deep Q network to obtain a target supplementary light parameter optimization model.

[0125] Specifically, group the real-time working state data according to the track position points, and classify the multi-dimensional data such as light intensity, spectral ratio, irradiation angle, voltage and current, and temperature from different position points into independent data groups. Assume that the track position points on the track are , and the real-time working state data corresponding to each position point includes light intensity , spectral ratio , irradiation angle , voltage , current , and temperature . Through the grouping operation, organize these data into multiple groups of working state data , where represents the index of the track position point. For each group of working state data, standardize each parameter to eliminate the differences in dimension and numerical range and ensure the stability of subsequent analysis. The standardization processing formula is as follows:

[0126] ;

[0127] where is the original parameter value, is the mean value of this parameter, is the standard deviation, is the parameter value after standardization. By performing standardization processing on , in each group of data in sequence, a supplementary lighting state feature sequence is obtained. Based on the supplementary lighting state feature sequence, a supplementary lighting execution feature vector is constructed. Let the supplementary lighting execution feature vector be , which contains standardized multi-dimensional features and represents the supplementary lighting execution state of the orbital position point . At the same time, the growth rate change rate , the chlorophyll content change rate , and the photosynthesis efficiency change rate in the plant growth indicators are weighted and calculated to generate a comprehensive plant growth evaluation index . Its calculation formula is:

[0128] ;

[0129] where is the weight coefficient used to balance the importance of different growth indicators. According to the supplementary lighting execution feature vector and the comprehensive plant growth evaluation index , the supplementary lighting effect score is calculated. The scoring model is implemented through regression analysis or deep learning methods. For example, a linear model is adopted:

[0130] ;

[0131] where is the model weight and is the bias value. If is lower than the preset scoring threshold , then it is necessary to calculate the gradient value of the supplementary lighting effect score with respect to the initial supplementary lighting parameter optimization model parameters to guide the optimization direction of the model. The calculation of the gradient value is based on the loss function , and the mean squared error loss function is defined as:

[0132] ;

[0133] where is the target score value. Through the backpropagation algorithm, the gradient of the loss function with respect to the model parameters is obtained. The Adam optimizer is used to optimize the gradient value. The Adam optimizer combines the first-order and second-order momenta, and the formula is:

[0134] ;

[0135] ;

[0136] ;

[0137] wherein is the momentum parameter, is the learning rate, is the numerical stability parameter, is the optimized model parameter. The updated model parameter is reloaded into the deep Q-network to perform online update on the value evaluation part in the network, ensuring that the network's value estimation of the state-action pair is more accurate. After multiple iterations of optimization, the target supplementary lighting parameter optimization model is obtained.

[0138] Please refer to Figure 2 , Figure 2 which is the schematic block diagram of the structure of the deep learning recognition system 200 for plant growth state based on plant lights provided by the embodiment of the present application. As Figure 2 shown, the deep learning recognition system 200 for plant growth state based on plant lights includes:

[0139] An acquisition module 210, configured to perform multimodal data acquisition on a plant growth area by using an orbital multispectral camera array matching a fixed LED plant supplementary light, and obtain a plant image data set and an environmental parameter data set including orbital position points;

[0140] A feature extraction module 220, configured to extract features from the plant image data set and the environmental parameter data set to obtain a plant growth feature vector;

[0141] A prediction module 230, configured to input the plant growth feature vector into a multitask deep learning model based on an improved Transformer architecture for plant growth state recognition and supplementary light demand prediction, and obtain a plant growth stage discrimination result and supplementary light demand spatio-temporal distribution data;

[0142] An iteration module 240, configured to construct an initial supplementary lighting parameter optimization model based on the plant growth stage discrimination result and the supplementary light demand spatio-temporal distribution data, and iteratively optimize through a deep Q-network to obtain a sequence of LED plant supplementary light illumination parameters.

[0143] Through the collaborative cooperation of the above-mentioned various components, by mounting a multi-spectral camera array on an orbital LED fill light and combining multiple environmental sensors, multi-dimensional and all-round data collection of the plant growth area is achieved. An improved image processing and feature extraction algorithm is used to accurately extract features such as plant leaf morphology, temperature distribution, and photosynthesis, and a comprehensive plant growth feature representation is obtained through feature fusion. Based on a multi-task deep learning model with an improved Transformer architecture, accurate identification of the plant growth state and accurate prediction of the fill light requirements are achieved, and the model has good spatio-temporal feature capture ability. By constructing a fill light parameter optimization model with a deep Q network and combining an experience replay mechanism, efficient optimization and dynamic adjustment of the fill light parameters are realized, significantly improving the accuracy of fill light control. An industrial bus communication and multi-channel drive control scheme is adopted to accurately execute the fill light parameters, ensuring the reliable transmission and execution of control instructions. The fill light effect is evaluated in real time through a deep learning evaluation model, and the optimization model is updated online based on the evaluation results, ensuring the continuous optimization of the fill light control strategy.

[0144] Please refer to Figure 3 , Figure 3 FIG. 3 is a schematic block diagram of the structure of a deep learning recognition device 300 for plant growth state based on a plant light provided by an embodiment of the present application. The deep learning recognition device 300 for plant growth state based on a plant light includes a processor 301 and a memory 302, and the processor 301 and the memory 302 are connected through a system bus 303. Among them, the memory 302 may include a non-volatile storage medium and an internal memory.

[0145] The non-volatile storage medium can store a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor 301, the processor 301 can be made to execute any one of the above-mentioned deep learning recognition methods for plant growth state based on a plant light.

[0146] The processor 301 is used to provide computing and control capabilities to support the operation of the entire deep learning recognition device 300 for plant growth state based on a plant light.

[0147] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can be made to execute any one of the above-mentioned deep learning recognition methods for plant growth state based on a plant light.

[0148] Those skilled in the art can understand that Figure 3The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the plant growth state deep learning recognition device 300 based on plant lights involved in the solution of this application. Specifically, the plant growth state deep learning recognition device 300 based on plant lights may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0149] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0150] It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described plant growth state deep learning recognition device 300 based on plant lights can refer to the corresponding process of the plant growth state deep learning recognition method based on plant lights described above, and will not be repeated here.

[0151] The embodiment of this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by one or more processors, the one or more processors are caused to implement the plant growth state deep learning recognition method provided by the embodiment of this application.

[0152] Among them, the computer-readable storage medium may be an internal storage unit of the plant growth state deep learning recognition device 300 based on plant lights in the foregoing embodiment, such as the hard disk or memory of the plant growth state deep learning recognition device 300 based on plant lights. The computer-readable storage medium may also be an external storage device of the plant growth state deep learning recognition device 300 based on plant lights, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped with the plant growth state deep learning recognition device 300 based on plant lights.

[0153] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, systems, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0154] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0155] As described above, the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do 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 this application.

Claims

1. A deep learning recognition method for plant growth status based on plant lights, characterized in that, include: A track-mounted multispectral camera array paired with a fixed LED plant light was used to collect multimodal data from the plant growth area, obtaining a plant image dataset and an environmental parameter dataset containing track location points. Performing feature extraction on the plant image dataset and the environmental parameter dataset to obtain a plant growth feature vector; The plant growth feature vector is input into a multi-task deep learning model based on an improved Transformer architecture to perform plant growth status recognition and fill light demand prediction, and obtain plant growth stage discrimination results and fill light demand spatiotemporal distribution data; specifically comprising: inputting the plant growth feature vector into a multi-head attention encoder in a multi-task deep learning model based on an improved Transformer architecture for parallel processing, wherein the multi-head attention encoder comprises 8 attention heads, each attention head adopts a 64-dimensional query vector, a key vector and a value vector, calculates feature weights by a scaled dot product attention mechanism, and combines with sinusoidal position encoding to obtain an initial feature representation matrix; inputting the initial feature representation matrix into a 3-layer series-connected Transformer encoding block in the multi-task deep learning model, wherein each layer of the Transformer encoding block comprises a 512-dimensional multi-head self-attention layer, a 2048-dimensional feedforward neural network layer and a LayerNorm normalization layer to obtain an enhanced feature matrix; and inputting the enhanced feature representation matrix into a multi-task deep learning model; The feature matrix is input into the feature diversion module in the multi-task deep learning model to perform task separation, and a dual-task feature representation is obtained through the growth state recognition branch and the fill light demand prediction branch; the growth state recognition branch feature in the dual-task feature representation is input into the fully connected neural network in the multi-task deep learning model, and the fully connected neural network includes 4 hidden layers, and the hidden layer dimensions are 512, 256, 128, and 64 respectively, to obtain a growth state feature vector, and the probability score of the growth stage is calculated according to the growth state feature vector to obtain a plant growth stage discrimination result; the fill light demand prediction branch feature in the dual-task feature representation is input into the spatiotemporal attention module in the multi-task deep learning model, and the spatial attention map and the temporal attention vector are calculated, and the spatiotemporal dimension information is integrated using a gated fusion mechanism to obtain a fill light demand feature map; feature extraction is performed on the fill light demand feature map to obtain a fill light demand prediction tensor, and the fill light demand prediction tensor is reconstructed into a spatiotemporal distribution structure to obtain fill light demand spatiotemporal distribution data; An initial fill light parameter optimization model is constructed based on the plant growth stage discrimination result and the spatiotemporal distribution data of the fill light demand, and an LED plant fill light illumination parameter sequence is obtained through deep Q network iterative optimization.

2. The method for deep learning recognition of plant growth state based on plant lights according to claim 1, wherein The multi-modal data acquisition of the plant growth area using a track-mounted multispectral camera array matched with a fixed LED plant fill light is performed to obtain a plant image dataset and an environmental parameter dataset containing track location points, including: Calibrate the moving orbit of the orbiting multi-spectral camera array with an electronic map to obtain orbit coordinate mapping data, and divide the acquisition positions of the multi-spectral camera array based on the orbit coordinate mapping data to obtain N orbit position points; Perform multi-spectral data acquisition trigger control on the N orbit position points, collect plant apparent characteristics through a visible light camera, collect plant temperature distribution characteristics through an infrared camera, and collect plant photosynthetic characteristics through a chlorophyll fluorescence camera to obtain original multi-modal image data, and perform image noise reduction and spatial registration on the original multi-modal image data to obtain a plant image data set containing orbit position points; Collect the environmental data of the N orbit position points, collect environmental temperature and humidity data through a temperature and humidity sensor, collect carbon dioxide concentration data through a carbon dioxide sensor, and collect environmental light data through an illuminance sensor to obtain original environmental parameter data, and perform outlier detection and data standardization processing on the original environmental parameter data to obtain an environmental parameter data set.

3. The method for deep learning recognition of plant growth state based on plant lights according to claim 2, wherein Extract features from the plant image data set and the environmental parameter data set to obtain a plant growth feature vector, including: Perform image segmentation preprocessing on the visible light images in the plant image data set to obtain an image feature mapping array, and perform plant leaf area segmentation on the image feature mapping array to obtain plant leaf area mask data; Perform morphological processing and edge detection on the plant leaf area mask data to obtain a set of contour feature points, and calculate plant morphological parameters such as leaf area, number of leaves, plant height, and crown width based on the set of contour feature points to obtain plant morphological feature data; Calculate the temperature distribution characteristics of the infrared image data in the plant image data set to obtain a temperature distribution characteristic mapping matrix, and perform statistical analysis on the temperature distribution characteristic mapping matrix to obtain plant temperature distribution characteristic data; Extract photosynthesis characteristics from the chlorophyll fluorescence images in the plant image data set to obtain a photosynthesis characteristic mapping matrix, and perform spatial feature analysis on the photosynthesis characteristic mapping matrix to obtain plant photosynthetic characteristic data; Extract temporal characteristics from the temperature data, humidity data, carbon dioxide concentration data, and illuminance data in the environmental parameter data set to obtain environmental characteristic data; Perform feature coding fusion on the plant morphological feature data, the plant temperature distribution characteristic data, the plant photosynthetic characteristic data, and the environmental characteristic data to obtain a plant growth feature vector.

4. The method for deep learning recognition of plant growth state based on plant lights according to claim 1, wherein Construct an initial supplementary lighting parameter optimization model based on the plant growth stage discrimination result and the spatio-temporal distribution data of the supplementary lighting requirement, and iteratively optimize it through a deep Q network to obtain a sequence of LED plant supplementary lighting illumination parameters, including: Input the plant growth stage discrimination result and the supplementary lighting requirement distribution data into a state encoder for state space encoding to obtain a supplementary lighting environment state vector; Perform feature partitioning on the supplementary lighting environment state vector, construct an action space including light intensity, spectral ratio, and irradiation angle, and discretize the action space into an action matrix to obtain a supplementary lighting parameter action set; Construct a deep Q-network based on the supplementary lighting environment state vector and the supplementary lighting parameter action set. The deep Q-network includes a value evaluation network with a dual-stream architecture to obtain an initial supplementary lighting parameter optimization model; For the supplementary lighting environment state vector within the irradiation range of the LED plant supplementary light, calculate the Q-values of different supplementary lighting parameter actions through the deep Q-network, and select the optimal action using the ε-greedy strategy to obtain an initial supplementary lighting parameter combination; Input the initial supplementary lighting parameter combination into the experience replay buffer pool, randomly sample batch data from the experience replay buffer pool, and calculate the target Q-value through the temporal difference algorithm to obtain a parameter update gradient; Optimize and update the weight parameters of the deep Q-network according to the parameter update gradient to obtain a target supplementary lighting parameter combination; Normalize the light intensity, spectral ratio, and irradiation angle of the target supplementary lighting parameter combination to obtain an LED plant supplementary lighting parameter sequence; 5. The method for deep learning recognition of plant growth state based on plant lights according to claim 4, characterized in that The deep learning recognition method for plant growth state based on plant lights further includes: Perform protocol conversion on the LED plant supplementary lighting parameter sequence, and send the control parameters of light intensity, spectral ratio, and irradiation angle to the LED supplementary light controller through the industrial bus to obtain a supplementary lighting control instruction; Convert the light intensity parameter in the supplementary lighting control instruction into a PWM duty cycle modulation signal, and control the output power of different bands of LEDs through the four-channel constant current source of the LED driving circuit to obtain the target light intensity; Calculate the drive current ratio of red, blue, and far-red LEDs according to the spectral ratio parameter in the supplementary lighting control instruction, and adjust the current values of each band of LEDs through the multi-channel LED driving circuit to obtain the target spectral ratio; Parse the irradiation angle parameter in the supplementary lighting control instruction, and drive the pitch angle adjustment mechanism of the LED supplementary light through the stepper motor control system to obtain the target irradiation angle; Collect the electrical parameters of the LED supplementary light through a voltage sensor, a current sensor, and a temperature sensor, and collect the spectral output data of the LED supplementary light through a spectral sensor to obtain the real-time working state data of the LED supplementary light; Perform deep learning evaluation on the real-time working state data and plant growth indicators, calculate the supplementary lighting effect score, and when the supplementary lighting effect score is lower than the preset threshold, perform online parameter update on the initial supplementary lighting parameter optimization model to obtain a target supplementary lighting parameter optimization model.

6. The method for deep learning recognition of plant growth state based on plant lights according to claim 5, characterized in that The performing deep learning evaluation on the real-time working state data and plant growth indicators, calculating the supplementary lighting effect score, and when the supplementary lighting effect score is lower than the preset threshold, performing online parameter update on the initial supplementary lighting parameter optimization model to obtain a target supplementary lighting parameter optimization model includes: The real-time working status data is grouped to obtain multiple groups of working status data, and the light intensity, spectral ratio, illumination angle, voltage and current, and temperature parameters in each group of working status data are standardized to obtain a fill light state characteristic sequence; Constructing a fill light execution feature vector based on the fill light state feature sequence, and performing weighted calculation on the growth rate change rate, chlorophyll content change rate, and photosynthesis efficiency change rate among the plant growth indicators to obtain a comprehensive plant growth evaluation index; calculating a fill light effect score based on the fill light execution feature vector and the comprehensive plant growth evaluation index, and comparing the fill light effect score with a preset score threshold; when the fill light effect score is lower than the preset threshold, calculating a gradient value of the fill light effect score relative to an initial fill light parameter optimization model parameter to obtain a parameter update gradient; Performing Adam optimizer processing on the parameter update gradient to obtain a parameter optimization direction, and performing online updating on the deep Q network weights of the initial fill light parameter optimization model based on the parameter optimization direction to obtain updated model parameters; The updated model parameters are reloaded into the deep Q network, and the parameters of the value evaluation network of the deep Q network are updated to obtain the target fill light parameter optimization model.

7. A deep learning recognition system for plant growth status based on plant lights, characterized in that, The method for performing deep learning recognition of plant growth status based on a plant lamp according to any one of claims 1 to 6 comprises: The acquisition module is used to collect multimodal data of the plant growth area using a track-mounted multispectral camera array equipped with a fixed LED plant fill light, obtaining a plant image dataset and an environmental parameter dataset containing track location points; A feature extraction module is used to extract features from the plant image dataset and the environmental parameter dataset to obtain a plant growth feature vector; A prediction module is used to input the plant growth feature vector into a multi-task deep learning model based on an improved Transformer architecture to perform plant growth status recognition and supplementary lighting demand prediction, thereby obtaining plant growth stage discrimination results and supplementary lighting demand spatiotemporal distribution data; The iterative module is used to construct an initial fill light parameter optimization model based on the plant growth stage judgment result and the spatiotemporal distribution data of the fill light demand, and obtain the LED plant fill light illumination parameter sequence through deep Q network iterative optimization.

8. A deep learning recognition device for plant growth status based on plant lights, characterized in that, The plant growth status deep learning recognition device based on plant lamps includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory so that the plant growth status deep learning recognition device based on the plant lamp executes the plant growth status deep learning recognition method based on the plant lamp as described in any one of claims 1-6.

9. A computer-readable storage medium, on which instructions are stored, characterized in that, When the instructions are executed by the processor, the deep learning recognition method for plant growth status based on plant lamps according to any one of claims 1 to 6 is implemented.

Images