Plant monitoring and watering maintenance system and method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI POLYTECHNIC UNIV MECHANICAL & ELECTRICAL COLLEGE
- Filing Date
- 2026-04-10
- Publication Date
- 2026-06-09
Smart Images
Figure CN122176529A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of plant monitoring and maintenance technology, specifically a plant monitoring and watering maintenance system and method. Background Technology
[0002] With the rapid development of agricultural automation technology, precision agriculture has become a key area for improving crop yield and resource utilization efficiency. In recent years, with the widespread application of the Internet of Things (IoT), sensor technology, and artificial intelligence (AI), plant health monitoring and automatic irrigation systems have gradually shifted from traditional manual management to intelligent control.
[0003] In the field of agricultural automation, the accuracy and real-time performance of plant monitoring and maintenance systems directly affect crop yield and resource utilization.
[0004] However, traditional systems often use single-channel image acquisition (such as RGB or single spectrum) and lack effective image registration and noise reduction processing, resulting in low dataset quality.
[0005] Furthermore, existing technologies generally use the standard MQTT protocol to transmit control commands. However, in high-load scenarios (such as large-scale greenhouse clusters), the backlog of protocol messages can cause watering response delays of several seconds to tens of seconds. This delay can cause irreversible damage to plants when they experience sudden drought. According to agricultural experimental data, watering operations delayed by more than 5 seconds can increase crop yield reduction by more than 15%.
[0006] Based on this, a plant monitoring and watering maintenance system and method are provided, which can eliminate the drawbacks of existing equipment. Summary of the Invention
[0007] The purpose of this invention is to provide a plant monitoring and watering maintenance system and method to solve the problems in the prior art.
[0008] To achieve the above objectives, the present invention provides the following technical solution: A plant monitoring and watering maintenance system includes an image acquisition module, an edge computing node, and a watering execution module; The image acquisition module includes a multispectral camera, which is used to acquire image data of the target plant leaves. The image data includes RGB channel images and near-infrared channel images of the target plant leaves. The edge computing node includes an image preprocessing module, a feature extraction module, a parameter calculation module, a state recognition module, and a control command generation module. The image preprocessing module preprocesses image data, including image registration and noise filtering, to generate an aligned multi-channel image dataset. The feature extraction module includes a convolutional neural network model and a global average pooling model. The convolutional neural network model extracts features from the multi-channel image dataset and, based on the extracted features, generates a final feature sequence signal. The global average pooling model globally averages the final extracted feature sequence signal to obtain an overall image feature vector signal. The parameter calculation module calculates parameter data based on the overall image feature vector signal. The state recognition module generates plant health abnormality status labels based on the parameter data. The control command generation module generates control commands based on the plant health abnormality status labels. The watering execution module includes a watering actuator, a water pump, and a valve assembly. The watering actuator receives control commands to control the switching of the water pump and the valve assembly.
[0009] Based on the above technical solutions, the present invention also provides the following optional technical solutions: A method for a plant monitoring and watering maintenance system includes the following steps: Step S1: Acquire image data of the target plant leaves at preset time intervals using the multispectral camera in the image acquisition module, and transmit the acquired image data to the edge computing node; Step S2: The image preprocessing module at the edge computing node performs image registration and noise filtering on the acquired image data to generate an aligned multi-channel image dataset; Step S3: The convolutional neural network model at the edge computing node extracts features from the multi-channel image dataset and outputs the final feature sequence signal. Finally, the global average pooling model at the edge computing node is used to obtain the overall image feature vector signal. Step S4: The parameter calculation module of the edge computing node calculates parameter data based on the overall image feature vector signal. The parameter data includes leaf color distribution parameters, texture complexity parameters, morphological curling parameters, and vegetation index parameters. Step S5: Input the parameter data into the status recognition module in the edge computing node, and the status recognition module outputs a label for abnormal plant health status. Step S6: The control instruction generation module of the edge computing node generates control instructions based on the abnormal plant health status tags and transmits them to the watering execution module; Step S7: At the edge computing node, a hash chain log is generated for the transmitted image data, extracted features, and control commands, and the integrity of the log is confirmed through a blockchain verification mechanism; Step S8: The image preprocessing module at the edge computing node receives feedback data from the watering process in step S5, and updates the parameters of the convolutional neural network model online through incremental learning based on the feedback data.
[0010] In one alternative: in steps S1 and S2, the image data includes RGB channel images and near-infrared channel images; In one alternative: In step S2, the image preprocessing module inputs the image data into the affine transformation matrix operation unit to complete image registration, and the image preprocessing module applies a Gaussian filter kernel through the convolution filter module to complete noise filtering and generate an aligned multi-channel image dataset.
[0011] In one alternative: in step S3, the convolutional neural network model includes a ResNet-50 backbone network and a Vision Transformer attention module, wherein the ResNet-50 backbone network is used to extract features from a multi-channel image dataset.
[0012] In one alternative: the Vision Transformer attention module includes a multi-head self-attention layer and a feedforward network layer. The multi-head self-attention layer calculates attention weight signals based on the extracted features, and the feedforward network layer outputs the final feature sequence signal.
[0013] In one alternative: the plant health abnormality status tags include drought stress tags, waterlogging stress tags, and nutrient deficiency tags; the control instruction generation module of the edge computing node generates control instructions based on the plant health abnormality status tags, wherein the control instruction parameters include water output values. , calculated as ,in This is a preset proportional coefficient. This represents the quantified value of the drought stress label. This is the quantification value of the excessive moisture stress label. This is a quantitative value for nutritional deficiency labels.
[0014] In one alternative: the quantification value of the drought stress label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of drought stress state in the current leaf image, where 0 indicates no drought stress and 1 indicates complete drought stress; The quantization value of the excessive moisture stress label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of the over-wet stress state in the current leaf image, where 0 indicates no over-wet stress and 1 indicates complete over-wet stress. The quantitative value of the nutrient deficiency label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of nutrient deficiency in the current leaf image, where 0 indicates no nutrient deficiency and 1 indicates complete nutrient deficiency.
[0015] In one alternative: the sum of the above three quantization values satisfies .
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention simultaneously acquires RGB and near-infrared dual-channel images using a multispectral camera, and generates a multi-channel dataset through image registration and Gaussian filtering for noise reduction. This effectively improves the accuracy of feature extraction and enhances the robustness of identifying abnormal plant health conditions. Combined with the ResNet-50 model, it achieves accurate quantification of leaf color distribution, texture complexity, morphological curling, and vegetation index, significantly improving the classification accuracy of abnormal plant health labels and exhibiting better adaptability under complex lighting conditions.
[0017] 2. This invention encapsulates and transmits control commands through an enhanced MQTT communication protocol, and activates the water pump and valve components based on the activation duration calculated by the water output value in the watering actuator. This effectively solves the response delay problem, breaks through the transmission limitations of traditional protocols in high-load scenarios, and achieves low-latency control of watering linkage, thereby effectively improving the real-time performance in the field of agricultural automation.
[0018] 3. This invention effectively solves the problems of insufficient data integrity and model adaptability by generating hash chain logs and using blockchain verification mechanisms to confirm log integrity, and by updating parameters online incrementally based on watering feedback data. It breaks through the bottlenecks of easy data tampering and static models in existing technologies, realizes the system's data security and continuous model optimization capabilities, thereby effectively improving the long-term reliability of the plant monitoring system. Attached Figure Description
[0019] Figure 1 This is a flowchart of steps S1 to S3 of the method of the present invention.
[0020] Figure 2 This is a flowchart of steps S3 to S6 of the method of the present invention.
[0021] Figure 3 This is a flowchart of steps S6 to S8 of the method of the present invention.
[0022] Figure 4 This is a flowchart of step S8 of the method of the present invention.
[0023] Figure 5This is a flowchart illustrating the operation of the system of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0025] In one embodiment, such as Figures 1-3 As shown, a plant monitoring and watering maintenance system includes an image acquisition module, an edge computing node, and a watering execution module. The image acquisition module includes a multispectral camera, which is used to acquire image data of the target plant leaves. The image data includes RGB channel images and near-infrared channel images of the target plant leaves. The edge computing node includes an image preprocessing module, a feature extraction module, a parameter calculation module, a state recognition module, and a control command generation module. The image preprocessing module preprocesses image data, including image registration and noise filtering, to generate an aligned multi-channel image dataset. The feature extraction module includes a convolutional neural network model and a global average pooling model. The convolutional neural network model extracts features from the multi-channel image dataset and, based on the extracted features, generates a final feature sequence signal. The global average pooling model globally averages the final extracted feature sequence signal to obtain an overall image feature vector signal. The parameter calculation module calculates parameter data based on the overall image feature vector signal. The state recognition module generates plant health abnormality status labels based on the parameter data. The control command generation module generates control commands based on the plant health abnormality status labels. The watering execution module includes a watering actuator, a water pump, and a valve assembly. The watering actuator receives control commands to control the switching of the water pump and the valve assembly.
[0026] In this embodiment, RGB (visible light) and near-infrared (NIR) images of plant leaves are simultaneously acquired by a multispectral camera. RGB provides phenotypic information such as color and texture, while NIR reflects physiological states such as photosynthetic activity and water stress, thus achieving phenotypic-physiological dual-dimensional data fusion.
[0027] It integrates an image preprocessing module, a Transformer-enhanced convolutional neural network model, a global average pooling module, a state recognition module, and a control command generation module. Control commands are generated through feature extraction, parameter calculation (such as quantification values of drought / overwater / nutrient deficiency labels), and state classification.
[0028] The watering actuator drives the water pump and valve assembly to achieve precise water volume control according to the control command. After the execution is completed, it sends back a confirmation signal through the enhanced MQTT protocol to trigger a new round of data acquisition and model update.
[0029] It employs a hash chain log generation + blockchain verification mechanism to ensure that data transmission is tamper-proof throughout the entire process; it also supports online incremental learning to continuously optimize model parameters.
[0030] A method for a plant monitoring and watering maintenance system includes the following steps: Step S1: Acquire image data of the target plant leaves at preset time intervals using the multispectral camera in the image acquisition module, and transmit the acquired image data to the edge computing node; Step S2: The image preprocessing module at the edge computing node performs image registration and noise filtering on the acquired image data to generate an aligned multi-channel image dataset; Step S3: The convolutional neural network model at the edge computing node extracts features from the multi-channel image dataset and outputs the final feature sequence signal. Finally, the global average pooling model at the edge computing node is used to obtain the overall image feature vector signal. Step S4: The parameter calculation module of the edge computing node calculates parameter data based on the overall image feature vector signal. The parameter data includes leaf color distribution parameters, texture complexity parameters, morphological curling parameters, and vegetation index parameters. Step S5: Input the parameter data into the status recognition module in the edge computing node, and the status recognition module outputs a label for abnormal plant health status. Step S6: The control instruction generation module of the edge computing node generates control instructions based on the abnormal plant health status tags and transmits them to the watering execution module; Step S7: At the edge computing node, a hash chain log is generated for the transmitted image data, extracted features, and control commands, and the integrity of the log is confirmed through a blockchain verification mechanism; Step S8: The image preprocessing module at the edge computing node receives feedback data from the watering process in step S5, and updates the parameters of the convolutional neural network model online through incremental learning based on the feedback data.
[0031] In step S1, the preset time interval is a fixed interval point that is evenly distributed every 24 hours, and the number of interval points is 60, so as to ensure continuous monitoring.
[0032] In one embodiment, in steps S1 and S2, the image data includes RGB channel images and near-infrared channel images; The acquired RGB channel images are defined as three-dimensional tensors. ,in Indicates the image height. Indicates the image width; The acquired near-infrared channel images are defined as two-dimensional tensors. ,in Indicates the image height. Indicates the image width; The acquired RGB channel images and near-infrared channel images are combined into four-channel image data. ,in Indicates the image height. Indicates the image width.
[0033] In one embodiment, in step S2, the image preprocessing module inputs the image data into the affine transformation matrix operation unit to complete image registration, and the image preprocessing module applies a Gaussian filter kernel through the convolution filter module to complete noise filtering and generate an aligned multi-channel image dataset.
[0034] right RGB channel images and near-infrared channel images Image registration is performed using an affine transformation matrix. Near-infrared channel images Align to RGB channel image Using the coordinate system, the aligned near-infrared channel image is obtained. , where the affine transformation matrix It is a 3×3 matrix, including rotation, translation, and scaling parameters; For RGB channel images and aligned near-infrared channel images Noise removal is achieved by applying Gaussian filtering, where the Gaussian filter kernel is defined as follows: , The filtered RGB channel image is obtained by setting the standard deviation. and filtered near-infrared channel images The filtered RGB channel image and filtered near-infrared channel images Combined into an aligned multi-channel image dataset ,in Indicates the image height. Indicates the image width.
[0035] In one embodiment, in step S3, the convolutional neural network model includes a ResNet-50 backbone network and a Vision Transformer attention module, wherein the ResNet-50 backbone network is used to extract features from a multi-channel image dataset.
[0036] Will Divided into a sequence of non-overlapping image blocks, where each image block has a size of [size missing]. , Preset patch size, number of image patches To obtain the image patch sequence Each of them and each image block Flattened into a vector Then, position-encoded embeddings are added to form the input token sequence. Each of them , For the embedding dimension, a ResNet-50 backbone network is applied to the input token sequence. Convolutional feature extraction is performed. The ResNet-50 backbone network consists of multiple residual blocks that sequentially process the token sequence to output a preliminary feature sequence. Each of them , For the embedded dimension.
[0037] The Vision Transformer attention module includes a multi-head self-attention layer and a feedforward network layer. The multi-head self-attention layer calculates attention weight signals based on the extracted features, and the feedforward network layer outputs the final feature sequence signal.
[0038] Preliminary feature sequence Input the Vision Transformer attention module, which consists of a multi-head self-attention layer and a feedforward network layer, where the multi-head self-attention is computed as follows: , , , , For the projection matrix, For the key dimension, the feedforward network layer consists of two fully connected layers, outputting the final extracted feature sequence. Each of them , For the embedding dimension, the global average pooling model extracts the final feature sequence. Perform global average pooling to obtain the overall image feature vector. ,in For the embedded dimension.
[0039] In step S4, the parameter calculation module at the edge computing node calculates the feature vector from the overall image. Extracting RGB color correlation sub-features ,in For the color sub-dimension, calculate the leaf color distribution parameters, including calculating the mean of the RGB three channels. , , ,in Indicates the image height. Indicates the image width. The filtered RGB channel image, and the calculation of variance. , , From the overall image feature vector Extract texture-related sub-features ,in For texture sub-dimensions, and to compute texture complexity parameters using the Gray-Level Co-occurrence Matrix (GLCM), where GLCM elements... This indicates that gray levels i and j are at distance d and angle d. Co-occurrence frequencies under the given conditions are used to calculate contrast. Where d is the preset distance 1, The degree is 0, from the overall image feature vector Extracting morphologically correlated sub-features ,in For the morphological sub-dimension, calculate the morphological curl parameter using a custom curl index. ,in This represents the change in height after detecting the leaf edge. The change in blade width is determined by the Canny edge detection algorithm from the filtered RGB channel image. After extracting the leaf contour, the feature vector from the overall image is calculated. Extracting spectral correlation sub-features ,in For the vegetation index sub-dimension, calculate the vegetation index parameter NDVI using the formula. ,in This is the filtered near-infrared channel image. The red channel of the filtered RGB channel image is used as the input for the state recognition module. The calculated leaf color distribution parameters, texture complexity parameters, morphological curling parameters, and vegetation index parameters are input into the state recognition module, which is a fully connected layer network that outputs plant health abnormality status labels.
[0040] In one embodiment, the plant health abnormality status tags include drought stress tags, waterlogging stress tags, and nutrient deficiency tags; the control instruction generation module of the edge computing node generates control instructions based on the plant health abnormality status tags, wherein the control instruction parameters include water output values. , calculated as ,in This is a preset proportional coefficient. This represents the quantified value of the drought stress label. This is the quantification value of the excessive moisture stress label. This is a quantitative value for nutritional deficiency labels.
[0041] The quantification value of the drought stress label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of drought stress state in the current leaf image, where 0 indicates no drought stress and 1 indicates complete drought stress; The quantization value of the excessive moisture stress label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of the over-wet stress state in the current leaf image, where 0 indicates no over-wet stress and 1 indicates complete over-wet stress. The quantitative value of the nutrient deficiency label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of nutrient deficiency in the current leaf image, where 0 indicates no nutrient deficiency and 1 indicates complete nutrient deficiency.
[0042] The sum of the above three quantization values satisfies .
[0043] In step S5, the plant health abnormality status label and control command parameters are encapsulated into an enhanced MQTT message payload, which is a JSON format structure {"status":[L_{drought},L_{overwet},L_{nutrition}],"command":Q}\; In step S6, the status recognition module uses the enhanced MQTT communication protocol to publish the encapsulated message payload to the preset topic "plant / irrigation / control". The enhanced MQTT communication protocol includes QoS level 2 to ensure at least one delivery. The published message payload is transmitted to the watering actuator via a wireless network, and the watering actuator receives the enhanced MQTT message payload. The watering actuator parses the message payload and extracts the water volume output value. According to the water output value Calculate the pump activation time ,in To preset the water pump flow rate, the duration of water pump activation is set. ,in , To preset the water pump flow rate, the valve assembly simultaneously opens the channel corresponding to the root region of the target plant. The valve assembly includes multiple solenoid valves, each corresponding to the root region of a single plant, and operates for a specified duration. After completion, shut off the water pump and valve assembly to achieve the desired water output value. To the root region of the target plant; Edge computing nodes collect and transmit image data from multispectral cameras. ,in Indicates the image height. Represents the image width, and is the extracted overall image feature vector. ,in For embedding dimensions, and control command parameters For image data Calculate hash value SHA256 is a 256-bit function of the secure hash algorithm, which is applied to the overall image feature vector. Calculate hash value SHA256 is a 256-bit function of the secure hash algorithm, used for control command parameters. Calculate hash value SHA256 is a 256-bit function of the secure hash algorithm, which generates a hash chain log and records the hash value. , , Connected to form a chain structure , where || denotes a concatenation operation. Given the hash chain of the previous log, calculate the root hash of the current hash chain log. SHA256 is a 256-bit function of the secure hash algorithm, which uses a blockchain verification mechanism to determine the root hash. Submit to the distributed ledger node and query the corresponding record in the ledger to confirm that the log is intact and has not been tampered with; New RGB channel images and near-infrared channel images acquired by a multispectral camera are defined as new combined image data. ,in Indicates the image height. Indicates the image width for the newly combined image data. Perform the same preprocessing as the image preprocessing to generate a feedback multichannel image dataset. ,in Indicates the image height. Representing the image width, a Transformer-enhanced convolutional neural network model is applied to the feedback multi-channel image dataset. Feature extraction is performed to obtain the overall image feature vector. ,in For the embedding dimension, based on the feedback overall image feature vector Calculate the actual plant health anomaly state labels, including actual drought stress labels, actual waterlogging stress labels, and actual nutrient deficiency labels. Calculate the loss value between the predicted plant health anomaly state labels and the actual plant health anomaly state labels using the cross-entropy loss function. ,in The number of categories (drought stress, excessive moisture stress, nutrient deficiency). One-hot encoding of the actual label. To predict the softmax probability of the label, the Adam optimizer is used based on the loss value. Calculate the gradient and update the parameters of the Transformer-enhanced convolutional neural network model, including the ResNet-50 backbone network weights and the Vision Transformer attention module weights, using the following update formula: ,in For model parameters, With a preset learning rate, it is updated every 24 hours and has the technical advantages of improving feature extraction accuracy through multimodal data fusion, reducing transmission latency through edge computing and enhanced MQTT protocol, and ensuring data integrity through blockchain mechanism. Specific Implementation A large-scale intelligent greenhouse agricultural base, covering approximately 500 acres, mainly grows cash crops such as roses, pothos, and peppers. The main problem the base faces is that plant health monitoring and irrigation management rely on traditional soil moisture sensors and manual observation. This leads to hidden problems such as drought stress, overwatering stress, or nutrient deficiency in plants under variable weather conditions. Especially at night or when unattended, the base cannot respond to changes in plant status in real time, resulting in water waste and reduced crop yields. The base management hopes to solve these problems through an automated system to improve plant survival rates and resource utilization efficiency.
[0045] In the application of this invention at the greenhouse base, a multispectral camera is first deployed at a height of 1 meter above each plant. Leaf images are captured every 24 minutes within a 24-hour period. The image resolution is 1024×1024 pixels, the RGB channel depth is 8 bits, and the near-infrared channel wavelength range is 700-1100nm. Immediately after acquisition, the images are transmitted wirelessly to a computing node deployed at the edge of the greenhouse, which is equipped with an NVIDIA Jetson processor. The Nano processor and 4GB of memory ensure data processing latency is less than 50 milliseconds. Next, the received images are preprocessed on edge computing nodes. A 3×3 affine transformation matrix T is used to align the near-infrared image with the RGB image. The matrix parameters include a rotation angle of 5 degrees, a translation vector of (2,3) pixels, and a scaling factor of 1.01. Then, a Gaussian filter kernel with a standard deviation σ=1.5 is applied to remove noise. The resulting multi-channel image dataset is used for subsequent analysis. In the feature extraction stage, the image is divided into 16×16 pixel patches, with a total of 4096 patches. Each patch is flattened into a vector and a sinusoidal positional code is added, with an embedding dimension D=512. This vector is then fed into a ResNet-50 backbone network, which contains 50 residual blocks, each including a 3×3 convolutional kernel and a batch normalization layer. The output preliminary feature sequence is then fed into Vision. The Transformer attention module has 12 multi-head self-attention layers, each with a key dimension d_k=64. The feedforward network has a hidden layer dimension of 2048. Attention weights are calculated using softmax to achieve global feature association. In the parameter calculation stage, sub-features are extracted from the overall feature vector F, and the RGB color mean μ_r=180, μ_g=150, μ_b=120 and variance σ_r²=2500, σ_g²=2000, σ_b²=1800 are calculated. The gray-level co-occurrence matrix (GLCM) is used to calculate the contrast (Contrast=45) at a distance d=1 and an angle θ=0 degrees, and the curvature index (Curvature) is also calculated. With an index of 12, leaf contours were extracted using the Canny edge detection threshold (100, 200), and height variation ΔH = 5 pixels and width variation ΔW = 42 pixels were calculated. The vegetation index NDVI = 0.65. These parameters were input into a fully connected layer classifier with three output neurons. The sigmoid activation function was used to identify drought stress quantification values L_drought = 0.8, overwet stress L_overwet = 0.2, and nutrient deficiency L_nutrition = 0.4. Based on these tags, a control command parameter Q=14 ml is generated, with a preset proportionality coefficient k=10. The tags and Q are encapsulated into a JSON payload and published via the enhanced MQTT protocol at QoS level 2 to the topic "plant / irrigation / control". This payload is then transmitted to the watering actuator, which is connected to an electromagnetic pump with a flow rate r=100 ml / min and a multi-channel valve. After receiving the payload, the actuator parses Q, calculates the activation duration t=8.4 seconds, activates the pump, and opens the corresponding valve channel at the plant root. The channel diameter is 5 mm. Ensure water is accurately delivered to a depth of 10 cm in the root zone soil, and shut down the equipment after completion to avoid waste; during the data verification phase, calculate the image data hash h1=SHA256(I_processed), feature vector h2=SHA256(F), and instruction h3=SHA256(Q), concatenate them to form a chain chain=h1||h2||h3||prev_chain, calculate the root hash root_h=SHA256(chain), and submit it to the Ethereum private chain distributed ledger, checking every 10 seconds to confirm no tampering; During the feedback update phase, new images are acquired 5 minutes after watering as feedback. Preprocessing and feature extraction are repeated, and the cross-entropy loss L=0.12 is calculated. The Adam optimizer, including ResNet-50 weights and the Transformer module, is used to update the model parameters every 24 hours to ensure the model adapts to the dynamic environment of the facility, where light levels change from 600 lux in the morning to 2000 lux at noon. Through this application, the facility achieves closed-loop automation from image acquisition to watering execution, solving the problems of traditional methods such as monitoring accuracy being affected by light, response delays of up to 5 minutes, and data tampering leading to decision-making errors. In actual operation, the system processes data for a single plant in 200 milliseconds, the water volume control error is less than 5%, and the accuracy rate for detecting abnormal plant health reaches 96%, reducing water consumption by 30% compared to manual management.
[0046] In its application at this greenhouse, the specific operation of the method of this invention began at 6:00 AM. When the multispectral camera detected an NDVI value of 0.55 for the rose plant leaves, the system automatically identified it as drought stress and generated a command of Q=20 ml. The actuator completed watering within 15 seconds. After watering, the feedback image showed that the NDVI value had risen to 0.68. After the model was updated, the next detection threshold was adjusted to 0.58. At 12:00 PM, under a light intensity of 1500 lux, the leaf curling index of the pothos reached 15. The system calculated L_drought=0.9, triggering a watering of Q=18 ml. The valve precisely delivered water to the root zone, avoiding root rot caused by overwatering. At 3:00 PM, the texture contrast of the pepper plant was Contrast= 50, nutrient deficiency L_nutrition=0.6, instruction includes nutrient solution addition ratio of 20%, feedback loss after execution L=0.08, the model attention weight was optimized, improving the sensitivity to nutrient indicators; at 8 pm, when unattended, the system verifies the integrity of logs through blockchain, confirming that 1000 data entries for the day have not been tampered with, ensuring traceability of decisions, the whole process does not require manual intervention, base managers monitor the real-time status through mobile terminal, system log shows that the number of plants processed daily is 5000, the average abnormal response time is 30 seconds, the total activation time of water pumps is reduced by 25%, the number of valve opening and closing times is optimized to 200 times per day, compared with the previous manual watering method, the crop growth cycle is shortened by 5 days and the yield is increased by 15%. To verify the effectiveness, the base conducted a control experiment. One group used the method of this invention, while the other group used a traditional sensor system. The results showed that the plant survival rate in the group using the invention was 98%, compared to 92% in the traditional group; the water resource utilization rate was 85% vs. 65%; the anomaly detection accuracy was 96% vs. 78%; and there were 0 data security incidents vs. 3. These data demonstrate the effectiveness of the method in solving the problems of low monitoring accuracy, high latency, and poor data reliability.
[0047] Comparison Data Table of Application Effects in Greenhouse Base
[0048] Performance data table before and after model update
[0049] Watering Execution Efficiency Data Table
[0050] Any aspects of this invention not described in detail are well-known to those skilled in the art.
[0051] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A plant monitoring and watering maintenance system, characterized in that, Includes an image acquisition module, edge computing nodes, and a watering execution module; The image acquisition module includes a multispectral camera, which is used to acquire image data of the target plant leaves. The image data includes RGB channel images and near-infrared channel images of the target plant leaves. The edge computing node includes an image preprocessing module, a feature extraction module, a parameter calculation module, a state recognition module, and a control command generation module. The image preprocessing module preprocesses image data, including image registration and noise filtering, to generate an aligned multi-channel image dataset. The feature extraction module includes a convolutional neural network model and a global average pooling model. The convolutional neural network model extracts features from the multi-channel image dataset and, based on the extracted features, generates a final feature sequence signal. The global average pooling model globally averages the final extracted feature sequence signal to obtain an overall image feature vector signal. The parameter calculation module calculates parameter data based on the overall image feature vector signal. The state recognition module generates plant health abnormality status labels based on the parameter data. The control command generation module generates control commands based on the plant health abnormality status labels. The watering execution module includes a watering actuator, a water pump, and a valve assembly. The watering actuator receives control commands to control the switching of the water pump and the valve assembly.
2. The method for a plant monitoring and watering maintenance system according to claim 1, characterized in that, Includes the following steps: Step S1: Acquire image data of the target plant leaves at preset time intervals using the multispectral camera in the image acquisition module, and transmit the acquired image data to the edge computing node; Step S2: The image preprocessing module at the edge computing node performs image registration and noise filtering on the acquired image data to generate an aligned multi-channel image dataset; Step S3: The convolutional neural network model at the edge computing node extracts features from the multi-channel image dataset and outputs the final feature sequence signal. Finally, the global average pooling model at the edge computing node is used to obtain the overall image feature vector signal. Step S4: The parameter calculation module of the edge computing node calculates parameter data based on the overall image feature vector signal. The parameter data includes leaf color distribution parameters, texture complexity parameters, morphological curling parameters, and vegetation index parameters. Step S5: Input the parameter data into the status recognition module in the edge computing node, and the status recognition module outputs a label for abnormal plant health status. Step S6: The control instruction generation module of the edge computing node generates control instructions based on the abnormal plant health status tags and transmits them to the watering execution module; Step S7: At the edge computing node, a hash chain log is generated for the transmitted image data, extracted features, and control commands, and the integrity of the log is confirmed through a blockchain verification mechanism; Step S8: The image preprocessing module at the edge computing node receives feedback data from the watering process in step S5, and updates the parameters of the convolutional neural network model online through incremental learning based on the feedback data.
3. The method for a plant monitoring and watering maintenance system according to claim 2, characterized in that, In steps S1 and S2, the image data includes RGB channel images and near-infrared channel images.
4. The method for a plant monitoring and watering maintenance system according to claim 2, characterized in that, In step S2, the image preprocessing module inputs the image data into the affine transformation matrix operation unit to complete image registration. The image preprocessing module then applies a Gaussian filter kernel through the convolution filtering module to remove noise and generate an aligned multi-channel image dataset.
5. The plant monitoring and watering maintenance system and method according to claim 2, characterized in that, In step S3, the convolutional neural network model includes a ResNet-50 backbone network and a Vision Transformer attention module. The ResNet-50 backbone network is used to extract features from a multi-channel image dataset.
6. The plant monitoring and watering maintenance system and method according to claim 5, characterized in that, The Vision Transformer attention module includes a multi-head self-attention layer and a feedforward network layer. The multi-head self-attention layer calculates attention weight signals based on the extracted features, and the feedforward network layer outputs the final feature sequence signal.
7. The plant monitoring and watering maintenance system and method according to claim 2, characterized in that, The plant health abnormality status tags include drought stress tags, excessive water stress tags, and nutrient deficiency tags; the control command generation module of the edge computing node generates control commands based on the plant health abnormality status tags, wherein the control command parameters include water output values. , calculated as ,in This is a preset proportional coefficient. This represents the quantified value of the drought stress label. This is the quantification value of the excessive moisture stress label. This is a quantitative value for nutritional deficiency labels.
8. The plant monitoring and watering maintenance system and method according to claim 7, characterized in that, The quantification value of the drought stress label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of drought stress state in the current leaf image, where 0 indicates no drought stress and 1 indicates complete drought stress; The quantization value of the excessive moisture stress label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of the over-wet stress state in the current leaf image, where 0 indicates no over-wet stress and 1 indicates complete over-wet stress. The quantitative value of the nutrient deficiency label A floating-point number between 0 and 1, output by the fully connected layer state recognition module through the softmax function, representing the confidence probability of nutrient deficiency in the current leaf image, where 0 indicates no nutrient deficiency and 1 indicates complete nutrient deficiency.
9. A plant monitoring and watering maintenance system and method according to claim 8, characterized in that, The sum of the above three quantized values satisfies .