Under-forest plant growth monitoring method and system based on image recognition

Abstract

The invention discloses an under-forest plant growth monitoring method and system based on image recognition, and belongs to the field of plant growth monitoring, and the monitoring method specifically comprises the following steps: I, selecting an under-forest plant growth region, and carrying out the region division according to the factors of terrain, vegetation type and ecological sensitivity; iI, acquiring various environmental data such as soil parameters and meteorological data of a plant growth area in real time, capturing plant image data at regular time, and constructing data pairs accurately matched with timestamps; iII, uploading the collected data to a cloud platform, carrying out unified storage and classified arrangement on the environment and image data, and preprocessing the image data at the same time; according to the invention, simultaneous acquisition, transmission and storage of various data of environmental soil and plant images can be realized, the accuracy of state identification is improved, the ability to understand and predict a complex growth process is enhanced, automatic identification and instant early warning of an abnormal state are realized, and the management response efficiency of the system is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of plant growth monitoring, in particular to an under-forest plant growth monitoring method and system based on image recognition. BACKGROUND

[0002] As an important part of the forest ecosystem, the growth state of under-forest plants directly affects the ecological balance and forestry production quality. Due to the influence of factors such as light shading, complex soil conditions, etc. in the under-forest environment, the growth process of plants shows strong spatio-temporal heterogeneity. Currently, there are soil temperature sensors, humidity sensors, pH sensors, nitrogen, phosphorus and potassium sensors, and environment-related atmospheric temperature and humidity sensors, as well as ultraviolet intensity sensors that can be purchased off the shelf. The entire system is generally referred to as a "multi-functional weather station", which is powered by a solar panel and stores electricity in a battery. This is the first part of the platform. The second part is "real-time monitoring of plant images and image recognition". The camera is purchased off the shelf and assembled to connect to the solar panel and battery, and can also realize remote real-time monitoring and transmission. The current industry generally refers to the first part of the multi-functional weather station or only the second part. In order to monitor the growth state of the planted plants, a large amount of data of various parameters such as environment and soil is collected, transmitted and stored through the first part. After a large number of plant images are captured through the second part, modeling is performed through post-image recognition technology, the growth trend and growth efficiency of the plants are analyzed in combination with the environment and soil data of the first part, the growth efficiency of the planted plants is predicted, and measures are taken to handle the plants according to the model calculation results, so that the plants can grow better. Therefore, we propose an under-forest plant growth monitoring method and system based on image recognition. SUMMARY

[0003] The purpose of the present application is to solve the defects in the prior art and propose an under-forest plant growth monitoring method and system based on image recognition.

[0004] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0005] An under-forest plant growth monitoring method based on image recognition, the specific steps of the monitoring method are as follows:

[0006] I. Select the under-forest plant growth area, and divide the area according to topography, vegetation type and ecological sensitivity factors;

[0007] II. Real-time collection of plant growth area soil parameters and meteorological data, environment data, and simultaneous timed capture of plant image data, construction of timestamp accurately matched data pairs;

[0008] III. Upload the collected data to the cloud platform, and uniformly store and classify the environmental and image data, while preprocessing the image data;

[0009] IV. Based on the processed image data and soil data at the corresponding time point, a structured multi-dimensional data set is constructed, and a plant growth model is established;

[0010] V. By establishing a plant growth model, the newly collected image and environmental data are analyzed to automatically determine the current growth state of the plant;

[0011] VI. By comparing and analyzing the historical growth data and the current state of the plant, if there is an abnormal growth trend, the warning mechanism is triggered, and the warning information is sent to the management personnel;

[0012] VII. Periodically retrain and optimize the plant growth model, and summarize the monitoring results to generate a multi-dimensional plant growth report, and upload it to the cloud platform.

[0013] As a further scheme of the present application, the soil parameters in step II specifically include temperature, humidity, pH value, electrical conductivity and nutrient content information; the meteorological data specifically include light, wind speed and air temperature and humidity information.

[0014] As a further scheme of the present application, the specific steps of preprocessing the image data in step III are as follows:

[0015] S1.1: The real-time plant image data of each plant growth area uploaded to the cloud platform in batches is subjected to pixel normalization processing, so that each pixel value in each plant image data is unified to the interval [0, 1], and then each plant image data is enhanced by rotation, flipping and contrast adjustment, and each plant image data is adjusted to a predetermined uniform resolution;

[0016] S1.2: The processed plant image data is input into the CNN model, and the CNN model uses the forward propagation algorithm to transfer each plant image data layer by layer, and from shallow to deep, extracts feature maps of each plant image data layer by layer through multiple convolution layers, and then uses the ReLU function to perform nonlinear activation on the feature maps output by each convolution layer;

[0017] S1.3: After nonlinear activation, each feature map is subjected to max-pooling processing to reduce the size of the feature map, and convolution, nonlinear activation and pooling processing are repeatedly performed until a predetermined processing round is reached, and then the processing is stopped, and the final feature map of each plant is output, and the generated final feature map is transmitted to the backbone network ResNet of the CNN model, and ResNet performs convolution processing on the input image through convolution kernels of different scales, and outputs multi-scale feature maps;

[0018] S1.4: generating multiple groups of candidate boxes at each pixel position of the feature map of different scales, predicting the offset of each candidate box or the positioning center point, then predicting the plant object class contained in each candidate box position, then decoding each candidate box, and adjusting each candidate box to the actual plant structure position based on the predicted offset, to obtain the position and size of each candidate box in the original image data;

[0019] S1.5: removing the overlapping candidate boxes that exist redundantly on the original image data, and taking the remaining candidate boxes as target boxes, while outputting the detection results of each target box, wherein the detection results include center coordinates, width and height dimensions, class labels and confidence scores, and then extracting the morphological parameters of plants in each target box region, such as area, aspect ratio and color index.

[0020] As a further scheme of the present application, the specific steps of constructing the structured multi-dimensional data set in step IV are as follows:

[0021] S2.1: selecting the time stamp with the minimum absolute difference between the image data collection time and the environment sampling time, synchronously matching the plant growth information and the environment data, and performing numerical standardization processing on the plant growth information through Z-score standardization;

[0022] S2.2: extracting soil temperature, soil humidity, soil pH value, light intensity and air humidity from each environment data, respectively, and establishing an environment parameter vector in a unified format, then integrating the standardized plant growth information, image data and corresponding environment parameters at each time into a unified fusion vector;

[0023] S2.3: using the interpolation filling method to unify the collection time of each data in the fusion vector, then adding the manually observed plant growth state label to the corresponding plant growth information, and grouping the fusion vector containing each group of plant growth information, image data and environment parameters at each time point into a final multi-dimensional data set.

[0024] As a further scheme of the present application, the specific steps of establishing the plant growth model in step IV are as follows:

[0025] S3.1: according to the preset time interval, collecting the fusion feature vectors in multiple consecutive time windows from the multi-dimensional data set, dividing each group of collected fusion feature vectors into a training set, a test set and a validation set, and then establishing a plant growth model based on a Bi-GRU network architecture;

[0026] S3.2: divide the training set into multiple batches of small batches of training subsets, and sequentially input each training subset into the plant growth model, the plant growth model input layer receives each training subset and is transmitted to the forward GRU layer and the reverse GRU layer respectively through forward propagation;

[0027] S3.3: update the hidden state on the forward GRU layer to capture the historical evolution of the plant growth process and output the forward hidden state, perform GRU update on the reverse GRU layer to capture the potential counteraction of the future state of the plant on the current time and output the reverse hidden state, and splice the forward and reverse hidden states at each time step to form the corresponding bidirectional features;

[0028] S3.4: generate the corresponding overall representation vector by averaging the generated bidirectional features, and then perform classification processing on the overall representation vector through the feedforward fully connected layer of the plant growth model, and output the plant growth state prediction result through the output layer;

[0029] S3.5: calculate the loss value between the plant growth state prediction result and the corresponding plant growth state label through the cross-entropy function, and input the loss value from the output layer of the plant growth model based on the chain rule, and calculate the gradient value of the loss value for each network layer of the plant growth model, and then update the model parameter information using the Adam optimizer;

[0030] S3.6: each training subset is used, and a training period is completed, after each training period, the verification set is input into the plant growth model, and the loss value of the plant growth model under the verification set is calculated, if the loss value does not reach the preset threshold, the plant growth model is retrained and adjusted until the loss value converges to the preset threshold range;

[0031] S3.7: after verification, the test set is input into the trained plant growth model, and based on the loss value of the plant growth model under the test set, the accuracy, precision, recall and F1 value are evaluated, if the evaluation result reaches the preset expectation, the plant growth model is deployed to each monitoring device, otherwise, the plant growth model is retrained.

[0032] A forest plant growth monitoring system based on image recognition, comprising a planning and deployment module, an environment acquisition module, an image acquisition module, a data transmission module, an image processing module, an identification and extraction module, a data fusion module, a state evaluation module, a detection and early warning module, a storage management module, a training and optimization module and a display report module;

[0033] The planning and deployment module is used for planning the monitoring area, and deploying and initializing each acquisition device according to the planning result;

[0034] The environment collecting module is used for collecting the environment data of the under-forest plant growth area at a fixed point and time;

[0035] The image collecting module is used for shooting the under-forest plant image to obtain the plant shape, growth posture, and leaf color information;

[0036] The data transmission module uploads each group of data collected by the environment collecting module and the image collecting module to the cloud platform through wireless communication;

[0037] The image processing module is used for preprocessing each plant image received;

[0038] The recognition and extraction module is used for recognizing the plant area in each plant image after preprocessing and establishing the digital plant features;

[0039] The data fusion module is used for integrating the plant image features and the synchronously collected environment data to establish the structured multi-dimensional data set;

[0040] The state evaluation module automatically identifies the current growth state of the plant according to the structured multi-dimensional data set;

[0041] The detection and early warning module monitors the change trend of the growth state, compares the historical data and the set threshold, automatically identifies the abnormal situation, and immediately generates the early warning information;

[0042] The storage management module is used for uniformly storing all original data and processing results in the database;

[0043] The training and optimization module periodically re-trains and optimizes the parameters of the state evaluation module;

[0044] The display and report module is used for presenting the collection and analysis results to the user in the forms of charts, heat maps, and time series charts, and generating the corresponding detection report.

[0045] As a further scheme of the application, the specific steps of the detection and early warning module for monitoring the change trend of the growth state and comparing the historical data and the set threshold are as follows:

[0046] S4.1: The newly collected image and environment data are subjected to the preprocessing process to establish the latest fusion vector, and the fusion vector is input into the plant growth model after training or parameter updating, the input layer of the plant growth model performs dimension reduction or dimension increase processing on each feature information in the input fusion vector, and the input dimensions are unified;

[0047] S4.2: the corresponding forward and reverse hidden states of the fusion vector are calculated through the forward GRU layer and the reverse GRU layer respectively, then the hidden state at the current moment is obtained through splicing, and the hidden state is input into the output layer for state classification processing, and the growth state of the current plant is determined according to the output result;

[0048] S4.3: the growth state sequence in a preset time interval is extracted as a historical trend line, then the deviation of the growth state of the current plant from the historical average and the average rate of historical state change are calculated, and whether the plant growth state is in an abnormal growth state is judged;

[0049] S4.4: if the deviation value or the average rate is higher than a preset threshold, it is judged that the current plant is in an abnormal growth state, and a warning is triggered, and the abnormal state is recorded, and a warning notification is sent to the user end through the server, otherwise, it indicates that the current plant is in a normal growth state.

[0050] Compared with the prior art, the beneficial effects of the present application are:

[0051] The present application processes the under-forest plant images uploaded in batches on the cloud through pixel normalization, enhancement and uniform resolution, extracts key structural features and candidate box information, and obtains detection results of the center coordinates, size, category and confidence of the plant, extracts the morphological parameters of the plant region, synchronously fuses the environmental soil data closest to the collection time, constructs a structured multi-dimensional data set, then completes data preprocessing by using Z-score standardization, interpolation filling and artificial labeling, trains the fusion feature vectors in the continuous time sequence based on the plant growth model, extracts the growth state evolution features and completes the state classification, after the training is completed, the system predicts the new data, calculates the deviation and rate of the current state from the historical trend to judge whether it is abnormal, if it exceeds the threshold, a warning notification is triggered, which can realize the simultaneous collection, transmission and storage of environmental soil data and plant images, improve the accuracy of state recognition, enhance the understanding and prediction ability of complex growth process, realize the automatic recognition and real-time warning of abnormal state, and improve the management response efficiency of the system. BRIEF DESCRIPTION OF DRAWINGS

[0052] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application.

[0053] Figure 1 A flow chart of an under-forest plant growth monitoring method based on image recognition is provided for the present application;

[0054] Figure 2 A system block diagram of an under-forest plant growth monitoring system based on image recognition is provided for the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application.

[0056] Embodiment 1

[0057] Referring to Figure 1 The present embodiment discloses a forest plant growth monitoring method based on image recognition, and the specific steps of the monitoring method are as follows:

[0058] Select a forest plant growth area, and divide the area according to factors such as terrain, vegetation type and ecological sensitivity.

[0059] Collect real-time plant growth area soil parameters and meteorological data and other environmental data, and simultaneously capture plant image data to construct a timestamp-accurately-matched data pair.

[0060] It should be further explained that the soil parameters specifically include temperature, humidity, pH value, conductivity and nutrient content information; and the meteorological data specifically includes light, wind speed and air temperature and humidity information.

[0061] Upload the collected data to a cloud platform, and uniformly store, classify and arrange the environmental and image data, while preprocessing the image data.

[0062] Specifically, the real-time plant image data of each plant growth area uploaded to the cloud platform in batches is subjected to pixel normalization processing, so that each pixel value in each plant image data is unified to the interval [0, 1], and then each plant image data is subjected to enhancement processing through rotation, flipping and contrast adjustment enhancement, while each plant image data is adjusted to a preset uniform resolution. The processed plant image data is input into the CNN model, which uses the forward propagation algorithm to pass each plant image data layer by layer, and extracts feature maps of each plant image data layer by layer through multiple convolution layers from shallow to deep. Then, the ReLU function is used to perform nonlinear activation on the feature maps output by each convolution layer. After nonlinear activation, each feature map is subjected to max-pooling processing to reduce the size of the feature map. The convolution, nonlinear activation and pooling processing are repeated until the preset processing round is reached, and then the processing is stopped and the final feature map of each plant is output. The generated final feature map is transmitted to the backbone network ResNet of the CNN model as an input image. ResNet performs convolution processing on the input image through convolution kernels of different scales, and outputs multi-scale feature maps. A plurality of candidate boxes are generated at each pixel position of the different scale feature maps, and the offset of each candidate box or positioning center point is predicted. Then, the class of the plant object contained in each candidate box is predicted. Then, each candidate box is decoded, and each candidate box is adjusted to the actual plant structure position based on the predicted offset to obtain the position and size of each candidate box in the original image data. The redundant overlapping candidate boxes in the original image data are removed, and the remaining candidate boxes are used as target boxes. The detection results of each target box are output, including the center coordinates, width and height, class label and confidence score. Then, the morphological parameters of the plants in each target box region, such as area, aspect ratio and color index, are extracted as plant growth information.

[0063] Based on the processed image data and the soil data at the corresponding time point, a structured multi-dimensional data set is constructed, and a plant growth model is established.

[0064] Specifically, the time stamp with the minimum absolute difference between the image data collection time and the environment sampling time is selected to synchronize and match the plant growth information and the environment data. The plant growth information is subjected to numerical standardization processing by Z-score standardization. Soil temperature, soil humidity, soil pH, light intensity and air humidity are extracted from each environment data, and an environment parameter vector in a unified format is established. Each set of plant growth information, image data and environment parameter at the corresponding time point is integrated into a unified fusion vector. The interpolation filling method is used to unify the collection time of each data in each fusion vector. Then, the manually observed plant growth state labels are added to the corresponding plant growth information. Each time point fusion vector containing each set of plant growth information, image data and environment parameter forms a final multi-dimensional data set.

[0065] Specifically, according to a preset time interval, a plurality of fusion feature vectors in a plurality of continuous time windows are collected from the multi-dimensional data set, and each group of collected fusion feature vectors is divided into a training set, a test set and a validation set. Then, based on the Bi-GRU network architecture, a plant growth model is established. The training set is divided into a plurality of batches of small batch training subsets, and each training subset is sequentially input into the plant growth model. The plant growth model input layer receives each training subset and is transmitted to the forward GRU layer and the reverse GRU layer through forward propagation, respectively. The hidden state is updated on the forward GRU layer to capture the historical evolution of the plant growth process and output the forward hidden state. The GRU update is performed on the reverse GRU layer to capture the potential counteraction of the future state of the plant on the current time and output the reverse hidden state. The forward and reverse hidden states of each time step are spliced to form the corresponding bidirectional features. Each bidirectional feature generated by the average pooling processing is used to generate the corresponding overall representation vector. The overall representation vector is classified by the feedforward fully connected layer of the plant growth model, and the plant growth state prediction result is output by the output layer. The loss value between the plant growth state prediction result and the corresponding plant growth state label is calculated by the cross-entropy function. The loss value is input from the output layer of the plant growth model based on the back propagation algorithm. The loss value is transmitted layer by layer based on the chain rule, and the gradient value of the loss value for each network layer of the plant growth model is calculated. The model parameter information is updated by using the Adam optimizer. Each training subset is used to complete a training period. After each training period, the validation set is input into the plant growth model, and the loss value of the plant growth model under the validation set is calculated. If the loss value does not reach the preset threshold, the plant growth model is retrained and adjusted until the loss value converges to the preset threshold range. After the verification is completed, the test set is input into the trained plant growth model, and the loss value of the plant growth model under the test set is used to evaluate its accuracy, precision, recall and F1 value. If the evaluation result reaches the preset expectation, the plant growth model is deployed to each monitoring device. Otherwise, the plant growth model is retrained.

[0066] By establishing the plant growth model, the newly collected images and environmental data are analyzed to automatically determine the current growth state of the plant.

[0067] The historical growth data and the current state of the plant are compared and analyzed. If there is an abnormal growth trend, a warning mechanism is triggered, and warning information is sent to the management personnel.

[0068] The plant growth model is periodically retrained and optimized, and the monitoring results are summarized to generate a multi-dimensional plant growth report and uploaded to the cloud platform.

[0069] Embodiment 2

[0070] Reference Figure 2 The embodiment discloses an image recognition-based under-forest plant growth monitoring system, which comprises a planning and deployment module, an environment acquisition module, an image acquisition module, a data transmission module, an image processing module, an identification and extraction module, a data fusion module, a state evaluation module, a detection and early warning module, a storage management module, a training and optimization module, and a display and report module.

[0071] The planning and deployment module is used for planning a monitoring area and deploying and initializing various acquisition devices according to a planning result; the environment acquisition module is used for acquiring environment data of the under-forest plant growth area at fixed points and fixed times; the image acquisition module is used for shooting images of under-forest plants to obtain various information of plant morphology, growth posture, and leaf color; and the data transmission module is used for uploading various groups of data acquired by the environment acquisition module and the image acquisition module to a cloud platform through wireless communication.

[0072] The image processing module is used for preprocessing various plant images received; the identification and extraction module is used for identifying plant areas in the preprocessed plant images and establishing digital plant features; and the data fusion module is used for integrating plant image features and synchronously acquired environment data to establish a structured multi-dimensional data set.

[0073] The state evaluation module is used for automatically identifying a current growth state of plants according to the structured multi-dimensional data set; and the detection and early warning module is used for monitoring a change trend of the growth state, comparing historical data and a set threshold, automatically identifying an abnormal condition, and immediately generating early warning information.

[0074] Specifically, a newly acquired image and environment data are subjected to a preprocessing procedure to establish a latest fusion vector, which is input into a plant growth model whose training or parameter updating is completed; a plant growth model input layer performs dimension reduction or dimension increase processing on various feature information in the input fusion vector, unifies an input dimension, respectively calculates a forward hidden state and a reverse hidden state corresponding to the fusion vector through a forward GRU layer and a reverse GRU layer, then acquires a hidden vector at a current time through a splicing manner, and inputs the hidden vector into an output layer to perform state classification processing; a growth state of a current plant is determined according to an output result, a growth state sequence in a preset time interval is extracted as a historical trend line, a deviation of the growth state of the current plant from a historical average value and an average rate of a historical state change are calculated, whether the plant growth state is in an abnormal growth state is determined, if the deviation value or the average rate is higher than a preset threshold, it is determined that the current plant is in the abnormal growth state, an early warning is triggered, an abnormal state is recorded, and an early warning notification is sent to a user end through a server, and otherwise, it is indicated that the current plant is in a normal growth state.

[0075] The storage management module is configured to store all raw data and processing results in a database; the training optimization module is configured to periodically retrain and optimize parameters of the state evaluation module; and the display report module is configured to present the collection and analysis results to a user in various forms such as a chart, a heat map, and a time series chart, and generate a corresponding detection report.

Claims

1. A method for monitoring understory plant growth based on image recognition, characterized in that: The specific steps of this monitoring method are as follows: Ⅰ. Select the understory plant growth area and divide it into regions based on topography, vegetation type and ecological sensitivity; II. Real-time collection of soil parameters and meteorological data of the plant growth area and other environmental data, while also capturing plant image data at regular intervals to construct data pairs with precise timestamp matching; III. Upload the collected data to the cloud platform, store and classify the environment and image data, and pre-process the image data; IV. Construct a structured multidimensional dataset based on the processed image data and the soil data at the corresponding time points, and establish a plant growth model; V. By establishing a plant growth model, the newly collected images and environmental data are analyzed to automatically determine the current growth status of the plant; VI. Compare and analyze the plant's historical growth data with its current status. If there are any abnormal growth trends, an early warning mechanism will be triggered and a warning message will be sent to the management personnel; Ⅶ. Regularly retrain and optimize the plant growth model, summarize the monitoring results to generate a multi-dimensional plant growth report, and upload it to the cloud platform.

2. The method for monitoring understory plant growth based on image recognition according to claim 1, characterized in that: The specific steps of pre-processing each image data in step III are as follows: S1.1: Perform pixel normalization on the real-time plant image data of each plant growth area uploaded to the cloud platform so that the pixel values ​​in each plant image data are unified to the range [0, 1]. Then, enhance each plant image data through rotation, flipping, and contrast adjustment, and adjust each plant image data to a preset uniform resolution. S1.2: The processed plant image data is input into the CNN model. The CNN model uses the forward propagation algorithm to pass each plant image data layer by layer, and extracts the feature map of each plant image data layer by layer from shallow to deep through multiple convolutional layers. The ReLU function is then used to perform nonlinear activation on the feature map output by each convolutional layer. S1.3: After the nonlinear activation is completed, each feature map is subjected to maximum pooling to reduce the size of the feature map. Convolution, nonlinear activation, and pooling are repeated until the preset processing rounds are reached. The processing is stopped and the final feature map of each plant is output. The generated final feature map is used as the input image and transmitted to the backbone network ResNet of the CNN model. ResNet convolves the input image with convolution kernels of different scales and outputs multi-scale feature maps. S1.4: Generate multiple sets of candidate boxes at each pixel position in the feature map at different scales, predict the offset of each candidate box or positioning center point, and then predict the plant object category contained in each candidate box position. Then, decode each candidate box and adjust each candidate box to the actual plant structure position based on the predicted offset to obtain the position and size of each candidate box in the original image data; S1.5: Remove redundant overlapping candidate frames from the original image data, and use the retained candidate frames as target frames. At the same time, output the detection results of each target frame, where the detection results include center coordinates, width and height dimensions, category labels, and confidence scores. Then, extract the morphological parameters of plants in each target frame area, such as area, aspect ratio, and color index, as well as other plant growth information.

3. The method for monitoring understory plant growth based on image recognition according to claim 2, characterized in that: The specific steps of constructing a structured multidimensional dataset in step IV are as follows: S2.1: Select the timestamp with the smallest absolute difference between the image data acquisition time and the environmental sampling time, synchronize the plant growth information with the environmental data, and perform numerical normalization on the plant growth information through Z-score normalization; S2.2: Extract soil temperature, soil moisture, soil pH, light intensity, and air humidity from each environmental data set and establish a unified environmental parameter vector. Then, integrate the standardized plant growth information and image data from each set with the environmental parameters at the corresponding moment into a unified fusion vector. S2.3: Use the interpolation filling method to unify the collection time of each data in each fusion vector, and then add the manually observed plant growth status labels to the corresponding plant growth information. The fusion vectors containing each group of plant growth information, image data and environmental parameters at each time point are combined into the final multidimensional data set.

4. The method for monitoring understory plant growth based on image recognition according to claim 3, characterized in that: The specific steps of establishing the plant growth model in step IV are as follows: S3.1: Collect fused feature vectors from multiple continuous time windows from a multidimensional dataset based on a preset time interval. Each set of fused feature vectors is divided into a training set, a test set, and a validation set. A plant growth model is then established based on the Bi-GRU network architecture. S3.2: Divide the training set into multiple small batches of training subsets, and input each training subset into the plant growth model in turn. The plant growth model input layer receives each training subset and passes it to the forward GRU layer and the reverse GRU layer through forward propagation. S3.3: Update the hidden state on the forward GRU layer to capture the historical evolution of the plant growth process and output the forward hidden state. Perform GRU updates on the reverse GRU layer to capture the potential reaction of the plant's future state to the current moment and output the reverse hidden state. Concatenate the forward and reverse hidden states at each time step to form the corresponding bidirectional features. S3.4: Each bidirectional feature generated is processed by average pooling to generate a corresponding overall representation vector. The overall representation vector is then classified through the feedforward fully connected layer of the plant growth model, and the plant growth status prediction result is output through the output layer; S3.5: Calculate the loss between the plant growth state prediction result and the corresponding plant growth state label using the cross entropy function. Input the loss value from the output layer of the plant growth model using the backpropagation algorithm. Transfer the loss value layer by layer based on the chain rule. Calculate the gradient of the loss value for each network layer of the plant growth model. Then use the Adam optimizer to update the model parameter information. S3.6: Each training subset is used for a training cycle. After each training cycle, the validation set is input into the plant growth model, and the loss value of the plant growth model under the validation set is calculated. If the loss value does not reach the preset threshold, the plant growth model is retrained and adjusted until the loss value converges to the preset threshold range. S3.7: After verification is completed, the test set is input into the trained plant growth model, and based on the loss value of the plant growth model under the test set, its accuracy, precision, recall rate and F1 value are evaluated. If the evaluation result meets the preset expectations, the plant growth model is deployed to each monitoring device; otherwise, the plant growth model is retrained.

5. A forest plant growth monitoring system based on image recognition, used to implement the forest plant growth monitoring method based on image recognition according to any one of claims 1 to 4, characterized in that: It includes planning and deployment module, environment acquisition module, image acquisition module, data transmission module, image processing module, recognition and extraction module, data fusion module, status assessment module, detection and early warning module, storage management module, training optimization module and display report module; The planning and deployment module is used to plan the monitoring area and, based on the planning results, to deploy and initialize each acquisition device; The environmental collection module is used to collect environmental data of the understory plant growth area at a fixed point and time; The image acquisition module is used to capture images of understory plants and obtain information on plant morphology, growth posture, and leaf color; The data transmission module uploads each set of data collected by the environment acquisition module and the image acquisition module to the cloud platform through wireless communication; The image processing module is used to pre-process the received plant images; The identification and extraction module is used to identify the plant area in each pre-processed plant image and establish digital plant features; The data fusion module is used to integrate the image features of each plant with the synchronously collected environmental data to establish a structured multidimensional dataset; The state assessment module automatically determines the current growth state of the plant based on the structured multidimensional data set; The detection and warning module monitors the changing trend of the growth status, compares historical data with the set threshold, automatically identifies abnormal situations, and immediately generates warning information; The storage management module is used to uniformly store all original data and processing results in the database; The training optimization module regularly retrains and optimizes parameters of the state evaluation module; The report display module is used to present the collection and analysis results to the user in the form of charts, heat maps and time series graphs, and generate corresponding detection reports.

6. The understory plant growth monitoring system based on image recognition according to claim 5, characterized in that: The specific steps of the detection and early warning module to monitor the changing trend of the growth status and compare historical data with the threshold setting are as follows: S4.1: The newly acquired image and environmental data are preprocessed to create the latest fusion vector, which is then input into the trained or parameter-updated plant growth model. The plant growth model input layer performs dimensionality reduction or dimensionality increase on the feature information in the input fusion vector to unify the input dimensions. S4.2: Calculate the feature and reverse hidden states corresponding to the fusion vector through the forward GRU layer and the reverse GRU layer, then obtain the hidden vector at the current moment through splicing. The hidden vector is then input to the output layer for state classification processing, and the current plant growth state is determined based on the output result. S4.3: Extract the growth state sequence of the preset time interval as the historical trend line, then calculate the deviation of the current plant growth state from the historical mean and the average rate of change of the historical state to determine whether the plant growth state is abnormal; S4.4: If the deviation value or the average rate is higher than the preset threshold, it is determined that the current plant is in an abnormal growth state and an early warning is triggered. The abnormal state is recorded and an early warning notification is sent to the user end through the server. Otherwise, it indicates that the current plant is in a normal growth state.

Images