Laser weeding method and field plant state monitoring method and system

Abstract

The invention belongs to the field of agricultural intelligent equipment, and particularly discloses a laser weeding method and a field plant state monitoring method and system. The weeding method comprises the following steps: inputting a plant surface topography image into an image segmentation model, and obtaining a first type identification result of a plant and position information of the plant on the plant surface topography image; performing plant type comparison on the fluorescence ratio of the plant and fluorescence feature library data to obtain a second type identification result of the plant; determining the type of the plant based on the first type identification result and the second type identification result of the plant; and based on the type of the plant and the position information of the plant on the plant surface morphology image, determining the position information of the weed so as to carry out laser inactivation on the weed. According to the method, the limitation of single visual identification in a complex environment is fundamentally overcome, and the accidental injury rate is remarkably reduced.

Description

Technical Field

[0001] This application belongs to the field of agricultural intelligent equipment, and more specifically, relates to a laser weeding method, a field plant status monitoring method and system. Background Technology

[0002] With the increasing global demand for green agriculture, laser weeding technology is gradually becoming an important alternative to traditional weeding methods due to its advantages such as no chemical residue and no damage to soil structure. Current laser weeding technologies mainly identify weeds using machine vision and then emit lasers to burn their growing points, but they still have significant shortcomings in terms of recognition accuracy, energy control, and environmental adaptability.

[0003] Traditional laser weeding equipment mainly relies on RGB image segmentation technology, but the recognition accuracy drops sharply in rainy, sunny, or crop-weed-adhesive scenarios. For example, in practical applications, the false kill rate can exceed 25% due to natural light interference or plant shading. Although some equipment uses light-blocking curtains to isolate ambient light, it is difficult to cope with plant swaying caused by dynamic wind disturbance, resulting in positioning errors.

[0004] Current laser energy parameters are mostly based on fixed threshold settings, lacking real-time feedback on the physiological state of weeds, such as water content and leaf age. This results in insufficient energy for mature weeds, leading to incomplete inactivation, or excessive energy for seedlings, resulting in resource waste. Existing systems cannot dynamically adjust parameters based on plant photosynthetic activity, limiting the improvement of weed control efficiency.

[0005] In summary, existing laser weeding technologies, due to their single recognition mode, static energy regulation, and redundant model calculations, are unable to meet the needs of efficient and precise operations in complex farmland scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application aims to provide a laser weeding method, a field plant status monitoring method and system, which aims to solve the problems of existing laser weeding technologies relying on a single RGB image segmentation technology, having a single recognition mode, and having recognition accuracy greatly affected by complex farmland scenarios, resulting in incomplete weeding and easy mis-weeding of crops.

[0007] Firstly, a laser weeding method includes the following steps: Collect fluorescence ratios and plant surface morphology images of plants in the target area; The plant surface morphology image is input into the image segmentation model to determine the first type of plant identification result and the plant's position information on the plant surface morphology image; The fluorescence ratio of the plant is compared with the fluorescence feature library data to obtain the second type identification result of the plant. The species of the plant is determined based on the results of the first and second species identification. Based on the plant species and the location information of the plant on the plant surface morphology image, the location information of weeds is determined so as to perform laser inactivation on the weeds.

[0008] In some implementations, the input to the image segmentation model is an image of the plant surface morphology, and the output is a segmentation mask of the crop, weeds and soil, a plant bounding box, and an image segmentation confidence score. The first plant species identification result was determined based on the segmentation mask of crops, weeds, and soil; The location information of the plant on the plant surface morphology image is determined based on the plant bounding box.

[0009] In some implementations, laser inactivation of weeds specifically involves controlling a laser to inactivate the weeds with a first weeding energy power; The energy output for the first weeding step is: ; in, K The crop coefficient in the plant; R The fluorescence ratio of the weeds to be inactivated; The stem thickness of the weeds is determined based on the weed segmentation mask; It is an energy correction factor; It is a natural number.

[0010] In some embodiments, the laser weeding method further includes: During the laser inactivation of weeds, the temperature of the weeds in the target area is monitored in real time. If the plant temperature is lower than the threshold temperature for plant structure damage, the first weeding energy power value is increased.

[0011] In some implementations, when the image segmentation confidence level output by the image segmentation model is lower than the preset confidence level, the plant species is determined by combining the first species identification result and the second species identification result. If the first species identification result and the second species identification result are inconsistent, the second species identification result is used as the plant species. When the image segmentation confidence level output by the image segmentation model is not lower than the preset confidence level, the plant species is determined by the first species identification result.

[0012] In some implementations, the image segmentation model is an improved YOLOv8-seg architecture model, specifically improved as follows: For the backbone network in the YOLOv8-seg architecture model, StarBlock is used to replace the C2f module in the backbone network to reduce the number of parameters in the backbone network. For the neck network in the YOLOv8-seg architecture model, an ECA attention mechanism is set to enhance the ability to extract plant stem and leaf texture features.

[0013] In some implementations, the improved YOLOv8-seg architecture model includes a backbone network, a neck network, a head network, and a mask co-decoder. The backbone network takes a plant surface morphology image as input and outputs shallow, medium and deep feature maps of the plant. The shallow feature map contains detailed information about the plant, including lesions, bristles and thin stems. The medium feature map contains partial shapes of the plant. The deep feature map is used to characterize the overall surface morphology of the plant. The input to the neck network is the shallow feature map, middle feature map and deep feature map output by the backbone network. The neck network fuses the shallow feature map, middle feature map and deep feature map and outputs the fused multi-scale feature map. The input to the head network is the multi-scale feature map output by the neck network, and the output of the head network is the predicted bounding box of the plant, the plant species and the image segmentation confidence, and the generation of segmentation masks for crops, weeds and soil. The input to the masked co-decoder is the predicted plant bounding box, the plant species and the image segmentation confidence, as well as the segmentation masks for crops, weeds and soil. The output is the segmentation mask for crops, weeds or soil corresponding to each plant bounding box.

[0014] In some implementations, the improved YOLOv8-seg architecture model is pre-trained through the following steps: Construct a dataset of surface morphology images of different types of weeds; where the weed species are labeled in the surface morphology images. Data augmentation methods were employed to expand the weed surface morphology image dataset, including random rotation and brightness fluctuation. The expanded weed surface morphology image dataset was input into the improved YOLOv8-seg architecture model in batches for training over a period of 100 epochs. With the goal of minimizing the combined loss function, the network parameters of the improved YOLOv8-seg architecture model were updated by combining momentum and adaptive learning rate through the backpropagation algorithm Adam optimizer until the performance of the YOLOv8-seg architecture model converged. The combined loss function includes bounding box regression loss, classification loss, and confidence loss. Network parameters include the kernel weights and biases of convolutional layers in the backbone network, the parameters of fully connected layers or attention layers in StarBlock, the weights and biases of convolutional layers, upsampling layers or specific connection paths used for feature fusion in the neck network, the weights and biases of convolutional layers in the head network that generate bounding boxes, confidence scores and classification beams, and the parameters of convolutional layers in the mask co-decoder used to upsample features and generate the final segmentation mask.

[0015] Secondly, this application provides a method for monitoring the status of plants in the field, specifically including the following steps: Collect fluorescence ratios and plant surface morphology images of the target area where the plant is located; The plant surface morphology image is input into the image segmentation model to obtain the first type identification result of the plant and the location information of the plant; The fluorescence ratio of plants in the target area is compared with the fluorescence feature database to obtain the second type identification result of the plants. The species of the plant is determined based on the results of the first and second species identification. A plant distribution map is constructed based on the plant species and the location of the plant in the plant surface morphology image.

[0016] In some implementations, the field plant condition monitoring method further includes: Based on the plant species, the first health diagnosis result of the crop is determined according to the real-time measured crop fluorescence ratio and / or the crop's maximum photochemical efficiency; wherein, the first health diagnosis result is used to characterize the crop's water stress diagnosis result; Based on the crop species, a second health diagnosis result is obtained based on the crop's maximum photochemical efficiency; the second health diagnosis result is used to characterize the strength of the crop's photosynthetic activity. The first and second health diagnoses of the crop are marked on the plant distribution map to generate a field plant management map, so as to monitor the status of the plants in the field.

[0017] In some implementations, the method for obtaining the first health diagnosis result of the crop is as follows: To determine whether the fluorescence ratio of crops is higher than that of crops in healthy conditions, a preliminary diagnosis can be made as to whether crops are under water stress. When a preliminary diagnosis indicates that the plant is under water stress, it is then determined whether the maximum photochemical efficiency of the crop is lower than the maximum photochemical efficiency of the crop in a healthy state. When the maximum photochemical efficiency of a crop is lower than the maximum photochemical efficiency of a crop in a healthy state, and the current non-optical quenching is higher than the normal value of non-optical quenching, and the soil moisture is lower than the normal value, the crop is determined to be under water stress; if the current non-optical quenching is lower than the normal value of non-optical quenching and / or the soil moisture is higher than the normal value, the crop is determined not to be under water stress. When the maximum photochemical efficiency is not lower than the maximum photochemical efficiency of the plant in a healthy state, the crop is determined not to be under water stress.

[0018] Thirdly, a laser weeding system based on chlorophyll fluorescence information for plant identification includes: A dual-band excitation detection module is used to acquire fluorescence ratios and plant surface morphology images of plants in the target area. The multimodal recognition module is used to input the plant surface morphology image into the image segmentation model to determine the first type of plant identification result and the plant's position information on the plant surface morphology image; compare the plant's fluorescence ratio with the fluorescence feature library data to determine the second type of plant identification result; and determine the plant type based on the first and second type identification results. The laser execution module is used to determine the location of weeds based on the plant species and the location information of the plant on the plant surface morphology image, so as to perform laser inactivation on the weeds.

[0019] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: Realization of field management: This application makes a preliminary diagnosis of water stress based on the shift in fluorescence ratio. At the same time, it obtains a second health diagnosis result by combining the maximum photochemical efficiency and the identified plant species, and makes another water stress diagnosis based on the maximum photochemical efficiency. The first and second health diagnosis results of the crop are marked on the plant distribution map to generate a field plant management map to monitor the status of the plants in the field. This can assist in the laser inactivation of weeds or provide a basis for external variable fertilizer applicators and precision irrigation systems to achieve true precision agricultural management.

[0020] This application provides a laser weeding method that uses dual-modal recognition by fusing the fluorescence ratio of plants (biological features) and deep learning image segmentation (morphological features). This fundamentally overcomes the limitations of single visual recognition in complex environments, increases the accuracy of weed recognition to over 98%, and significantly reduces the false injury rate. This application provides a laser weeding method, which utilizes a dynamic energy matching algorithm (weeding energy power is: The laser power and irradiation time can be adjusted in real time to achieve "energy application according to the grass", avoiding energy waste or incomplete inactivation, and saving about 30% energy compared with the fixed energy method.

[0021] This application provides a laser weeding method that uses an improved YOLOv8-seg model to locate weeds. The backbone of the YOLOv8-seg model uses a ShuffleNetV2 structure, and StarBlock replaces the original C2f module, reducing the number of parameters by approximately 60% and ensuring real-time processing speed. Simultaneously, data augmentation techniques such as random rotation and brightness fluctuations are employed during training to improve the robustness of the improved YOLOv8-seg model.

[0022] This application provides a laser weeding system. In the laser weeding stage, it provides a laser energy adaptive algorithm to achieve precise energy regulation based on the physiological activity and morphological characteristics of weeds. In the laser scanning stage, it identifies the distance between crops and weeds and triggers an airflow separation mechanism. Furthermore, it is equipped with an inactivation confirmation mechanism. If the temperature of the weed leaf surface or root stem is found to be less than 80°C (the cell structure damage threshold), the laser energy will be automatically increased by 20% for a second shot. It integrates an obstacle avoidance mechanism and inactivation effect verification (infrared thermal imager), forming a closed-loop control of perception-decision-execution-verification, which improves the autonomy and reliability of the laser weeding system in complex field environments. Attached Figure Description

[0023] Figure 1 This is a flowchart of a laser weeding method for identifying plants based on chlorophyll fluorescence information, provided in an embodiment of this application.

[0024] Figure 2 This is a schematic flowchart of the dual-modal recognition method provided in the embodiments of this application.

[0025] Figure 3 This is a schematic diagram of the laser weeding process provided in the embodiments of this application.

[0026] Figure 4 This is a schematic diagram of the dual-band laser excitation principle provided in the embodiments of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0029] In this application, the terms “first” and “second” are used to distinguish different objects, rather than to describe a specific order of objects.

[0030] In this application, the term "electrical connection" can refer to a direct circuit connection or a signal transmission via a communication protocol.

[0031] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0032] In the description of the embodiments in this application, unless otherwise stated, "multiple" means two or more.

[0033] The embodiments of this application are described below with reference to the accompanying drawings.

[0034] In a first aspect, this application provides a laser weeding method, comprising the following steps: Collect fluorescence ratios and plant surface morphology images of plants in the target area; The plant surface morphology image is input into the image segmentation model to determine the first type of plant identification result and the plant's position information on the plant surface morphology image; The fluorescence ratio of the plant is compared with the fluorescence feature library data to obtain the second type identification result of the plant. The species of the plant is determined based on the results of the first and second species identification. Based on the plant species and the location information of the plant on the plant surface morphology image, the location information of weeds is determined so as to perform laser inactivation on the weeds.

[0035] In some implementations, the input to the image segmentation model is an image of the plant surface morphology, and the output is a segmentation mask of the crop, weeds and soil, a plant bounding box, and an image segmentation confidence score. The first plant species identification result was determined based on the segmentation mask of crops, weeds, and soil; The location information of the plant on the plant surface morphology image is determined based on the plant bounding box.

[0036] In some implementations, laser inactivation of weeds specifically involves controlling a laser to inactivate the weeds with weed-killing energy power. The weeding energy power is as follows: ; in, K The crop coefficient in the plant; R The fluorescence ratio of the weeds to be inactivated; The stem thickness of the weeds is determined based on the weed segmentation mask; It is an energy correction factor; It is a natural number.

[0037] This application provides a method for monitoring plant status in the field, which specifically includes the following steps: Collect fluorescence ratios and plant surface morphology images of the target area; The plant surface morphology image is input into the image segmentation model to obtain the first type identification result of the plant and the location information of the plant; The fluorescence ratio of the plant is compared with the fluorescence feature library data to obtain the second type identification result of the plant. The species of the plant is determined based on the results of the first and second species identification. A plant distribution map is constructed based on the plant species and the location of the plant in the plant surface morphology image.

[0038] In some implementations, the field plant condition monitoring method further includes: Based on the plant species, the first health diagnosis result of the crop is determined according to the real-time measured crop fluorescence ratio and / or the crop's maximum photochemical efficiency; wherein, the first health diagnosis result is used to characterize the crop's water stress diagnosis result; Based on the plant species, the second health diagnosis result of the crop is determined according to the maximum photochemical efficiency measured in real time; the second health diagnosis result is used to characterize the strength of the crop's photosynthesis. The first and second health diagnoses of the crop are marked on the plant distribution map to generate a field plant management map, so as to monitor the status of the plants in the field.

[0039] In some implementations, the method for obtaining the first health diagnosis result of the crop is as follows: To determine whether the fluorescence ratio of crops is higher than that of crops in healthy conditions, a preliminary diagnosis can be made as to whether crops are under water stress. When a preliminary diagnosis indicates that the plant is under water stress, it is then determined whether the maximum photochemical efficiency of the crop is lower than the maximum photochemical efficiency of the crop in a healthy state. When the maximum photochemical efficiency of a crop is lower than the maximum photochemical efficiency of a crop in a healthy state, and the current non-optical quenching is higher than the normal value of non-optical quenching, and the soil moisture is lower than the normal value, the crop is determined to be under water stress; if the current non-optical quenching is lower than the normal value of non-optical quenching and / or the soil moisture is higher than the normal value, the crop is determined not to be under water stress. When the maximum photochemical efficiency is not lower than the maximum photochemical efficiency of the plant in a healthy state, the crop is determined not to be under water stress.

[0040] Thirdly, this application provides a laser weeding system, comprising: A dual-band excitation detection module is used to acquire fluorescence ratios and plant surface morphology images of plants in the target area. The multimodal recognition module is used to input the plant surface morphology image into the image segmentation model to determine the first type of plant identification result and the plant's position information on the plant surface morphology image; compare the plant's fluorescence ratio with the fluorescence feature library data to determine the second type of plant identification result; and determine the plant type based on the first and second type identification results. The laser execution module is used to determine the location of weeds based on the plant species and the location information of the plant on the plant surface morphology image, so as to perform laser inactivation on the weeds.

[0041] Example like Figure 1 As shown, this application provides a laser weeding system for identifying plants based on chlorophyll fluorescence information, including: a dual-band excitation detection module, a multi-modal recognition module, a laser execution module, and a decision control module; (1) The dual-band excitation detection module is used to simultaneously collect the fluorescence ratio R of plants in the target area. Maximum photochemical efficiency and plant surface morphology images; Among them, the fluorescence ratio R ( () represents the ratio of the fluorescence intensity corresponding to the fluorescence emission peaks of chlorophyll at 685 nm and 740 nm under specific light conditions; This refers to the fluorescence emission peak with a wavelength around 685 nm, which is mainly associated with chlorophyll a in photosystem II (PSII). This refers to the fluorescence emission peak at a wavelength around 740 nm. This peak is mainly related to the reabsorption effect of fluorescence by photosystem I (PSI) and the internal structure of the leaf. The fluorescence ratio can characterize the biochemical or physiological parameters of the plant, such as ion concentration and pH value, protein activity and conformational changes, and redox state. Ion concentration can characterize cell signal transduction, metabolism, and organelle function. Protein activity and conformation can characterize kinase activity, G protein activity, and apoptosis. Redox state can characterize oxidative stress, metabolic diseases, and aging. The surface morphology of all plants in the target area is characterized by plant surface morphology images; Maximum photochemical efficiency ( The ratio represents the highest efficiency at which PSII converts light energy into chemical energy; changes in this ratio can sensitively reflect the internal structure of plant leaves, the functional state of photosynthetic organs, and the environmental stresses they experience; the maximum photochemical efficiency corresponding to healthy crops ( The value is typically between 0.75 and 0.85; the most commonly used health benchmark is 0.83. (2) Multimodal recognition module, such as Figure 2 As shown, the functions are divided into two aspects: Firstly, the plant surface morphology image acquired by the dual-band excitation detection module is used to determine the first type identification result of the plant and the position of the plant in the plant surface morphology image by using an image segmentation model. At the same time, the fluorescence ratio of the plant obtained by the dual-band excitation detection module is compared with the fluorescence feature library data to obtain the second type identification result of the plant. When the confidence level of the image segmentation model output is lower than the preset confidence level set by the user, the plant species is determined by combining the first and second category identification results, and then a plant distribution map is constructed by combining the position of the plant in the plant surface morphology image. When the confidence level of the image segmentation model output is not lower than the preset confidence level set by the user, the first category identification result shall be used as the standard, and a plant distribution map shall be constructed by combining the position of the plant in the plant surface morphology image; wherein, the plant distribution map shows the plant type and the corresponding plant position in the plant surface morphology image; wherein, the plants are divided into crops and weeds. It should be noted that the fluorescence feature library data is shown in Table 1. Different plants correspond to different fluorescence ratios. The fluorescence ratios of plants collected by the dual-band excitation detection module are compared with the fluorescence feature library data to identify crops and weeds. The fluorescence feature library pre-stores the fluorescence ratios of 12 common weeds. For example, the fluorescence ratio R value of *Chenopodium album* is 1.2. The R-value for foxtail grass is 0.3, and its R-value is 0.8. 0.2, as shown in Table 1; Table 1

[0042] It should be noted that the function of the image segmentation model is to identify the semantic category or object instance of each pixel in the image, and to divide the image into regions with clear semantics or independent objects. Based on the image segmentation model of this application, which is used for plant species identification and weed localization, the existing image segmentation models that can be applied include: U-Net, DeepLab series, PSPNet series, OCRNet and TransUNET. In this application, when the plant surface morphology image is input into the image segmentation model, multiple convolutional layers and downsampling layers extract image features from the plant surface morphology image, and compress the image features at the bottleneck layer. In the decoding stage, multiple upsampling layers and convolutional layers convert the image features into the image size of the plant surface morphology image, and fuse the image features of the corresponding layers of the decoder and encoder to output the image segmentation result. More specifically, when the image segmentation model takes the plant surface morphology image as input, it outputs three categories of segmentation masks (crops, weeds, and soil), plant bounding boxes, and image segmentation confidence scores. A segmentation mask can clearly identify the category of each pixel in an image, such as corn plants, soil, weeds, or shadows. The mask strips away the complex background, retaining only the pixels of the plant itself. The outline and internal texture are the key visual fingerprints of the plant species. A bounding box is a rectangular coordinate frame that encloses an individual plant in the image. For example, in a densely planted scene, the plants may overlap. The image segmentation model can simultaneously output the mask and its corresponding bounding box for each individual plant. The coordinates of the bounding box give the precise location of the plant in the image. Therefore, the plant species can be obtained through the segmentation mask. In practical applications, this image segmentation model can directly output the plant species as the first category identification result. The bounding box can be used to locate the plant. Weeds and crops within the plant are identified using different bounding boxes. In some specific implementations, the image segmentation model is the YOLOv8-seg architecture model, which is divided into four parts: Backbone, Neck, Head, and mask co-decoder. The image segmentation model is an improved YOLOv8-seg architecture model, which includes a backbone network, a neck network, a head network, and a masked collaborative decoder. StarBlock is used to replace the C2f module in the backbone network to reduce the number of parameters in the backbone network. An ECA attention mechanism is set in the neck network to enhance the ability to extract the texture features of plant stems and leaves. The plant surface morphology image backbone network outputs shallow feature maps, medium feature maps, and deep feature maps of the plant. The shallow feature map contains detailed information about the plant, including lesions, bristles, and thin stems. The medium feature map contains partial shapes of the plant. The deep feature map is used to characterize the overall surface morphology of the plant. Shallow, mid, and deep feature maps are input into the neck network for fusion, and the fused multi-scale feature map is output. The fused multi-scale feature maps are input into the head network to predict the bounding boxes of plants, plant species and image segmentation confidence, and generate segmentation masks for crops, weeds and soil. The predicted plant bounding boxes, plant species, image segmentation confidence, and segmentation masks for crops, weeds, and soil are input into the mask co-decoder, which outputs the segmentation mask for crops, weeds, or soil corresponding to each plant bounding box. StarBlock was used to replace the C2f module in the backbone network to reduce the number of parameters in the backbone network; an ECA attention mechanism was set in the neck network to enhance the ability to extract plant stem and leaf texture features.

[0043] Here's an explanation: ECA stands for Efficient Channel Attention, a lightweight attention module that enhances the feature representation capabilities of convolutional neural networks by avoiding dimensionality reduction and using one-dimensional convolutions to efficiently capture interactions between channels. It's a plug-and-play module commonly embedded in the YOLO series to improve the performance of image segmentation models in tasks such as classification and segmentation.

[0044] Although the C2f module in the original YOLOv8-seg model has strong feature extraction capabilities, it suffers from a large number of parameters and high computational complexity, making it difficult to meet the real-time processing requirements on embedded devices with limited capacity (such as Jetson Nano). The core of StarBlock's lightweight module lies in reconstructing the original C2f module through structured model compression. By combining depthwise separable convolution with channel segmentation and shuffling, StarBlock achieves a significant reduction in parameters with almost no loss of feature extraction capabilities. Preferably, StarBlock's lightweight module performs the following operations in sequence: Channel segmentation: The input feature map is divided into two branches along the channel dimension on an average basis; Branching: One branch is used as a shortcut branch, directly performing the identity mapping; the other branch is processed sequentially using step 1. 1 convolution, 3 3. Depthwise separable convolution for feature transformation; Feature fusion: concatenates the outputs of two branches into channels; Channel shuffling: The channels of the spliced ​​feature map are rearranged to promote information exchange between features from different sources.

[0045] Furthermore, StarBlock's lightweight module is used to replace the C2f module in the backbone network of the YOLOv8-seg model. The improved YOLOv8-seg model is used to segment the input plant surface morphology image and outputs the segmentation masks of crops, weeds and soil, plant bounding boxes, plant species and image segmentation confidence.

[0046] Parameter comparison: The standard C2f module has approximately 2.5M parameters; StarBlock's lightweight module has approximately 1.0M parameters (a 60% reduction); calculation basis: StarBlock's lightweight module uses depthwise separable convolutions instead of standard convolutions, reduces the number of convolution kernels from 256 to 128, and sets the number of groups to 4; Improved computational efficiency: On embedded devices, inference speed is increased from 45fps to 120fps, meeting the needs of real-time processing in the field.

[0047] StarBlock's lightweight module achieves efficient feature aggregation and distribution through a central node; in the aggregation phase: the output feature maps of each branch are concatenated and 1 1. Convolutional layers are used for fusion, and the fusion weights are dynamically allocated by an attention mechanism (such as an ECA module) to ensure that detailed features such as lesions and burrs are not diluted. Distribution stage: The aggregated features are distributed to subsequent network layers through the central node; the distribution strategy is adaptively adjusted based on the feature scale: shallow features (such as plant burrs) are directly passed to the neck network, while deep features (such as overall morphology) are given priority for classification tasks.

[0048] Gradient flow optimization: The star topology shortens the gradient propagation path, alleviates the gradient degradation problem in deep network training, and accelerates model convergence (it can be stabilized within 100 epochs of training). Specifically, the training process involves inputting 100,000 labeled images containing 12 types of weeds into the YOLOv8-seg architecture model for training. During the training process, the input images are randomly rotated. ) and brightness fluctuation ( Data augmentation techniques such as ) are used to enable the YOLOv8-seg architecture model to have good recognition learning based on images of different qualities, thereby improving the robustness of the model. The AdamW optimizer (learning rate 0.001) is used in the learning process, with a training period of 100 epochs.

[0049] In some specific implementations, when the segmentation confidence score output by the YOLOv8-seg architecture model is below 90%, the fluorescence ratio R of the plant is used as an auxiliary criterion for the image segmentation result. If the R value falls within the fluorescence feature library data... If the score is within 0.2, the plant species is confirmed; if the segment confidence score is not lower than 90%, the plant species output by the YOLOv8-seg architecture model are directly used.

[0050] Secondly, it is used to compare the real-time measured fluorescence ratio of crops with the fluorescence ratio range of crops in a healthy state. If the real-time measured fluorescence ratio of crops does not fall within the fluorescence ratio range of crops in a healthy state, and the real-time measured fluorescence ratio of crops is too high, then the crop is preliminarily diagnosed as being under water stress. When a preliminary diagnosis indicates that the plant is under water stress, it is then determined whether the maximum photochemical efficiency of the crop is lower than the maximum photochemical efficiency of the crop in a healthy state. When the maximum photochemical efficiency of a crop is lower than the maximum photochemical efficiency of a crop in a healthy state, and the current non-optical quenching is higher than the normal value of non-optical quenching, and the soil moisture is lower than the normal value, the crop is determined to be under water stress; if the current non-optical quenching is lower than the normal value of non-optical quenching and / or the soil moisture is higher than the normal value, the crop is determined not to be under water stress. If the maximum photochemical efficiency of a crop is not lower than its maximum photochemical efficiency under healthy conditions, the crop is considered not to be under water stress. The water stress diagnosis result is used as the second health diagnosis result. It should be noted that if the results of the water stress diagnosis obtained by the fluorescence ratio of the target area and the maximum photochemical efficiency are inconsistent, the result of the maximum photochemical efficiency diagnosis shall prevail. Furthermore, by comparing the maximum photochemical efficiency of the crop under the current state with the maximum photochemical efficiency under the healthy state, a second health diagnosis result of the crop is obtained. The first and second health diagnoses of the crop are marked on the plant distribution map to generate a field plant management map for monitoring the status of the plants in the field. Therefore, the field plant management map not only marks the locations of weeds that need to be removed by laser, but also uses different colors or markers to indicate the first and second health diagnoses.

[0051] Among them, the second health diagnosis results of crops are used to characterize the strength of crop photosynthesis. More specifically, the method for second health diagnosis of crops is as follows: HI = (currently measured) ) / (in a healthy state) Health benchmarks) 100%.

[0052] It should be noted that water stress refers to the state in which plants suffer from physiological metabolic imbalance and stunted growth and development due to insufficient available water (drought) or excessive water (waterlogging). It is divided into drought stress and waterlogging stress.

[0053] It should be noted here that water stress disrupts the plant's water balance, triggering signal transduction (such as Ca²⁺). + Fluctuations), photosynthetic inhibition ( From a decrease in NPQ to an increase in oxidative damage (roGFP oxidation), a series of chain physiological responses occur. Various fluorescence ratio probes, like miniature dashboards installed inside cells, can report the status of these key physiological parameters in real time and in situ. By monitoring changes in fluorescence ratios, we can not only determine whether plants are suffering from water stress, but also understand the severity of stress, the plant's response strategies, and its tolerance potential, thus providing a powerful tool for crop physiological research, stress-resistance breeding, and precision agricultural management. Under severe or persistent moisture stress, PSII reaction centers may be damaged, affecting maximum photochemical efficiency. The ratio will decrease (around 0.83 for healthy plants, but can drop below 0.7 under stress), indicating impaired photosynthetic potential.

[0054] (3) Laser execution module, used to obtain the location information of weeds based on the plant distribution map obtained by the multimodal recognition module, and to inactivate the weeds with laser.

[0055] (4) Farm management platform, used to receive field plant management map, and based on the location information of weeds in the field management map, to assist the laser execution module in laser inactivation; to determine whether irrigation is needed for the plants based on the diagnostic data of water stress; and to determine whether fertilization is needed for the plants based on the second health diagnosis results of the crops, so as to achieve true precision agricultural management. In some implementations, the laser actuation module includes: a controller, a two-dimensional galvanometer, a weeding laser, and lens; The controller is used to control the initial inactivation position of the weed-killing laser based on the weed location information in the field plant management map; Two-dimensional galvanometers are used to control the weeding laser to perform spiral weeding (expanding outwards from the center) on each weed. The working principle of the two-dimensional galvanometer controlling the inactivation position of the weeding laser is as follows: The two-dimensional galvanometer includes an X-axis galvanometer and a Y-axis galvanometer; the X-axis galvanometer is used to control the horizontal deflection of the laser beam; the Y-axis galvanometer is used to control the vertical deflection of the laser beam. Lenses are used to collimate the laser beam reflected by the two-dimensional galvanometers, ensuring a uniform scanning field is formed on the working plane. Each galvanometer contains a tiny reflecting mirror, which is mounted on a high-speed servo motor. The high-speed servo motor drives the galvanometer to rotate, and the laser beam is reflected by the X-mirror and Y-mirror in sequence. Two-dimensional deflection is achieved through angle synthesis.

[0056] In some specific implementations, the weeding laser in the laser execution module uses a 1064nm wavelength fiber laser, which has an absorption depth of more than 2mm in plant tissue and a significant inactivation effect; more specifically, it uses a TG series fiber laser (1064nm, peak power 500W) from Yangtze Optical Fibre and Cable (YOFC).

[0057] More specifically, using Cambridge Technology's 6220H series two-dimensional galvanometer paired with Lens, scanning range .

[0058] In some implementations, the laser execution module also includes an infrared thermal imager, a brushless DC fan, and a decision control unit. The infrared thermal imager is used to detect the temperature rise of plants in the target area in real time during laser weeding and to obtain the temperature of the weeds. The brushless DC fan is used to physically separate weeds from crops, avoiding accidental damage to crops by the laser. More specifically, it controls a 0.2MPa pulsed airflow via a solenoid valve, with a nozzle diameter of 2mm and an airflow duration that can be set from 0.1 to 0.5 seconds. After the leaves are physically separated, the weed-killing laser is activated. The first aspect of the decision control unit is connected to a multimodal recognition module. This module determines the stem thickness of the weeds based on a segmentation mask, combines the fluorescence ratio of the target area with the plant species, and uses a laser energy adaptive algorithm to obtain a weeding energy power command for the weeding laser, which is then transmitted to the weeding laser. The weeding energy power command includes a first weeding energy power... E for: ;in, K For crop coefficients (e.g., cornfields) K =15), the crop coefficient is determined by the type of crop in the plant. Different crops in the target area result in different weed densities; for example, cornfields. K =15, wheat field K =12, rice paddies K =18, as shown in Table 2; R The dual-band excitation detection module is used to detect the fluorescence ratio of crops in real time. The diameter of the stem of a weed (unit: mm). This is the energy correction factor (with a value of 0.7). The natural parameter indicates that the formula is an exponential function formula; simultaneously, during laser weeding, the real-time temperature of the plants detected by an infrared thermal imager is received. If the plant temperature is below 80℃ (the threshold temperature for cell structure damage), the weeding energy power is increased; typically, the weeding energy power is increased to 1.2 times the original value, such as... Figure 3 As shown; The second aspect of the decision control unit is used to determine the distance between crops and weeds in the plant based on the location information and species of the plant. When it is detected that the distance between crops and weeds is less than 10mm, the brushless DC fan is controlled to separate the airflow between the plants to increase the distance between the crops and weeds. Then, the weeds are laser-inactivated to avoid accidental damage to the crops.

[0059] Table 2

[0060] In some implementations, the dual-band excitation detection module includes: a pulse-modulated chlorophyll fluorometer, a red laser source, and a global shutter CMOS camera; the pulse-modulated chlorophyll fluorometer includes a 450nm blue laser diode array and a dual quadratic active RC filter circuit. like Figure 4 As shown, a 450nm blue laser diode array and a dual quadratic active RC filter circuit are used to detect the fluorescence ratio R of the plant. ), maximum photochemical efficiency and non-photochemical quenching; A 450nm blue laser diode array serves as the main excitation source, irradiating the target area of ​​the plant to excite dual-band (685nm and 740nm) chlorophyll fluorescence. More specifically, when the plant is irradiated by the blue laser diode array, the chlorophyll fluorescence undergoes stimulated emission, resulting in dual-band (685nm and 740nm) chlorophyll fluorescence. Secondly, this method is used to obtain the maximum photochemical efficiency of the plant. A dual quadratic active RC filter circuit (with core operational amplifiers U1~U4) serves as a fluorescence detection circuit. It includes a photosensor, multi-stage operational amplifiers, filters, an analog-to-digital converter, and a microcontroller. The photosensor receives dual-band chlorophyll fluorescence generated by stimulated emission from the plant and converts it into an analog electrical signal. The multi-stage operational amplifiers effectively separate and output an ambient light compensation signal from the analog electrical signal. ) and fluorescence intensity signal ( The filter is used to filter out ambient light compensation signals; the analog-to-digital converter is used to perform analog-to-digital conversion on the fluorescence intensity signal to obtain a digital fluorescence intensity signal, wherein the digital fluorescence intensity signal includes the chlorophyll fluorescence intensity signal corresponding to the fluorescence emission peak at 685nm and the chlorophyll fluorescence intensity signal corresponding to the fluorescence emission peak at 740nm; the microcontroller is used to obtain the fluorescence ratio R based on the digital fluorescence intensity signal. );in, This refers to the fluorescence emission peak with a wavelength around 685nm; This refers to the fluorescence emission peak with a wavelength around 740 nm.

[0061] Similarly, pulse-modulated chlorophyll fluorometers are used to precisely separate the initial fluorescence yield in a target area without disturbing the plant's physiological state. and maximum fluorescence yield The ratio of the two is used to obtain the maximum photochemical efficiency. The core implementation method includes two steps. The first step is to modulate the measurement light: a very weak beam of modulated light of a specific frequency (usually red or blue light) is emitted to the plant. This beam of light has an extremely low intensity, which can only excite a very small amount of fluorescence in the plant and will not start photosynthesis. Therefore, the initial fluorescence when all PSII reaction centers of the plant are in an open state can be measured. The second step involves a saturation pulse. During the measurement, the pulse-modulated chlorophyll fluorometer will momentarily (approximately 0.5–1 second) emit an extremely strong beam (approximately 3000–10000 µmol photons) towards the plant. A saturated light pulse, this intense light can instantly return all electron carriers in the PSII reaction center to their initial state, temporarily reducing the photochemical efficiency to zero. The fluorescence value measured at this point is the maximum fluorescence value. .

[0062] Similarly, pulse-modulated chlorophyll fluorometers are used to observe changes in fluorescence parameters during a dynamic process, i.e., the transition from dark adaptation to light adaptation. NPQ is an indicator of the plant's ability to dissipate excess light energy absorbed by the PSII antenna pigments as heat, reflecting its immediate regulatory capacity in response to strong light stress. A higher NPQ value indicates stronger heat dissipation and greater photoprotection. The specific steps are as follows: After placing the plants in the target area in darkness for a sufficient period of time, the fluorescence yield under dark adaptation was measured first. The photochemical light is activated, providing a continuous light of a specific intensity to induce photosynthesis in the target area and bring the plants to a stable state. The instrument then continuously detects this steady-state fluorescence using measuring light. Apply a saturation pulse (under light adaptation): Under light-adapted steady state, apply a saturation pulse; the maximum fluorescence detected at this time is the light-adapted result. Due to the existence of NPQ, It will be significantly lower than To obtain a complete photoresponse curve, the intensity of the phototransformed light is typically increased gradually. After waiting for a steady state at each intensity, a saturation pulse is applied to obtain the photoresponse curve at that intensity. At the end of the experiment, the photochemical light was turned off, but the measurement light was kept on, and a saturation pulse was applied for a short time. At this time, since the photochemical light had just been turned off and the reaction center was reopened, but the antenna heat dissipation had not yet fully contacted the reaction, the minimum fluorescence after light adaptation was measured. (This parameter is used to calculate other quenching coefficients); the formula is: .

[0063] In some specific implementations, the dual quadratic active RC filter circuit adopts a fourth-order Butterworth structure, and the multi-stage operational amplifier is selected from TI's OPA2171 with a cutoff frequency set to 10kHz. The third stage amplifier in the multi-stage operational amplifier outputs an ambient light compensation signal. The fourth-stage amplifier outputs a fluorescence yield signal. The fluorescence output signal is converted from analog to digital by an analog-to-digital converter and sampled at 24-bit resolution. It should be noted that the multi-stage operational amplifier here contains multiple stages of amplifiers, and each stage of the amplifier has the same structure. In order to distinguish the amplification order, they are defined as the first stage amplifier, the second stage amplifier, ..., the nth stage amplifier. The output signal can be selected in which stage of the amplifier according to actual needs.

[0064] A red laser source and a global shutter CMOS camera (120fps) are used to acquire images of the plant surface morphology. The global shutter CMOS camera (120fps) is used for acquiring plant surface morphology images and is equipped with a polarizing filter to effectively suppress specular reflection and ensure image clarity. An 850nm infrared laser source is used to enhance illumination in nighttime or low-light environments to assist the global shutter CMOS camera in rapid sampling. The red laser source is an 850nm vertical-cavity surface-emitting laser (VCSEL). The optical axis of the red laser source does not coincide with the optical axis of the global shutter CMOS camera to avoid optical interference; for example, it is installed at a 15° angle to the optical axis of the global shutter CMOS camera. It should be noted that in practical applications, the blue laser diode array is triggered synchronously with the global shutter CMOS camera. In some specific implementations, a global shutter CMOS camera (FLIR BFS-PGE-50S5C) equipped with a linear polarizer is used, which is connected to the Jetson Nano development board via Gigabit Ethernet to achieve real-time image transmission at 120fps.

[0065] The following describes the mechanism for determining water stress: When water stress occurs, the energy available for photochemistry is reduced due to damage or protective inactivation of the PSII reaction center, resulting in a significant decrease in the maximum photochemical efficiency of crops and consequently affecting their fluorescence ratio. More specifically, when PSII function is downregulated, the absorbed light energy cannot be effectively used for photochemistry, and the excess excitation energy increases. This energy is dissipated in the form of heat or fluorescence, leading to a relative increase in the fluorescence yield of PSII emission near 685 nm. Some of the light-harvesting antenna complexes of PSII migrate to PSI, resulting in a relative weakening or no change in the fluorescence enhancement of PSI emission in the 730-740 nm range. Furthermore, under long-term stress, chlorophyll decomposes, leaves become thinner, and the absorption of fluorescence in the near-infrared region (740 nm) by crops decreases. However, the degree of reduction in fluorescence absorption near 740 nm may be smaller than that in the red region (685 nm). Therefore, the fluorescence ratio of crops will increase. Non-photochemical quenching reflects the crop's ability to dissipate excess light energy as heat, and is an important indicator for judging light stress. Essentially, it dissipates excess light energy absorbed by chlorophyll as harmless heat, thus preventing damage to photosynthetic structures (especially photosynthetic cells II) from strong light. NPQ does not have a fixed healthy value; its value depends on the light conditions at the time of measurement and the crop's physiological state. It increases when the crop is under stress, such as mild drought, low temperature, or nutrient deficiency. Due to limited photosynthetic capacity, the crop will increase NPQ to cope with the same light intensity, resulting in a higher NPQ value than the healthy control under the same measurement conditions. Under severe stress, NPQ may collapse, indicating damage to the photosynthetic structure. (The value has decreased significantly), and the ability to initiate NPQ will also be lost, and its value will be lower than that of healthy crops, indicating that the light protection system has failed; the NPQ range under different conditions is as follows: after sufficient dark adaptation, the NPQ value is 0; under moderate growth light, the NPQ value is 0.5~2; under strong light stress, the NPQ value is 2.0~5.0+; under stress (adaptive), the NPQ value is higher than that of the control; Soil moisture that is suitable for plant growth is generally maintained in the range of 50% to 80%. Most crops and vegetables can be allowed to consume soil moisture down to 50% to 70% of field capacity, while a small number of fruit trees can be kept at 30% to 50%.

[0066] In some specific implementations, the second health diagnosis method is: HI = (currently measured...) ) / (Crop health status) (benchmark value) 100%; the closer the HI value is to 100%, the healthier the crop's photosynthesis.

[0067] In summary, this application has the following advantages compared with the prior art: Realization of field management: This application makes a preliminary diagnosis of water stress based on the shift in fluorescence ratio. At the same time, it obtains a second health diagnosis result by combining the maximum photochemical efficiency and the identified plant species, and makes another water stress diagnosis based on the maximum photochemical efficiency. The first and second health diagnosis results of the crop are marked on the plant distribution map to generate a field plant management map to monitor the status of the plants in the field. This can assist in the laser inactivation of weeds or provide a basis for external variable fertilizer applicators and precision irrigation systems to achieve true precision agricultural management.

[0068] This application provides a laser weeding method that uses dual-modal recognition by fusing the fluorescence ratio of plants (biological features) and deep learning image segmentation (morphological features). This fundamentally overcomes the limitations of single visual recognition in complex environments, increases the accuracy of weed recognition to over 98%, and significantly reduces the false injury rate. This application provides a laser weeding method, which utilizes a dynamic energy matching algorithm (weeding energy power is: The laser power and irradiation time can be adjusted in real time to achieve "energy application according to the grass", avoiding energy waste or incomplete inactivation, and saving about 30% energy compared with the fixed energy method.

[0069] This application provides a laser weeding method that uses an improved YOLOv8-seg model to locate weeds. The backbone of the YOLOv8-seg model uses a ShuffleNetV2 structure, and StarBlock replaces the original C2f module, reducing the number of parameters by approximately 60% and ensuring real-time processing speed. Simultaneously, data augmentation techniques such as random rotation and brightness fluctuations are employed during training to improve the robustness of the improved YOLOv8-seg model.

[0070] This application provides a laser weeding system. In the laser weeding stage, it provides a laser energy adaptive algorithm to achieve precise energy regulation based on the physiological activity and morphological characteristics of weeds. In the laser scanning stage, it identifies the distance between crops and weeds and triggers an airflow separation mechanism. Furthermore, it is equipped with an inactivation confirmation mechanism. If the temperature of the weed leaf surface or root stem is found to be less than 80°C (the cell structure damage threshold), the laser energy will be automatically increased by 20% for a second shot. It integrates an obstacle avoidance mechanism and inactivation effect verification (infrared thermal imager), forming a closed-loop control of perception-decision-execution-verification, which improves the autonomy and reliability of the laser weeding system in complex field environments.

[0071] It should be understood that expressions such as “comprising” and “may include” used in this application indicate the existence of the disclosed functions, operations, or constituent elements, and do not limit one or more additional functions, operations, and constituent elements. In this application, terms such as “comprising” and / or “having” are to be interpreted as indicating a particular characteristic, number, operation, constituent element, component, or combination thereof, but not to exclude the existence or possibility of adding one or more other characteristics, numbers, operations, constituent elements, components, or combinations thereof.

[0072] In the description of the embodiments in this application, it should be noted that, unless otherwise expressly specified and limited, the term "connection" should be interpreted broadly. Furthermore, the mathematical concepts mentioned in the embodiments of this application, such as symmetry, equality, parallelism, and perpendicularity, are limitations specific to the current technological level, rather than absolute and strict mathematical definitions. Slight deviations are permissible; approximations of symmetry, equality, parallelism, and perpendicularity are all acceptable. For example, "A and B are parallel" means that A and B are parallel or approximately parallel, and the angle between A and B can be between 0 and 10 degrees. "A and B are perpendicular" means that A and B are perpendicular or approximately perpendicular, and the angle between A and B can be between 80 and 100 degrees.

[0073] 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 scope of the technology 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 laser weeding method, characterized in that, Includes the following steps: Collect fluorescence ratios and plant surface morphology images of plants in the target area; The plant surface morphology image is input into the image segmentation model to determine the first type of plant identification result and the plant's position information on the plant surface morphology image; The fluorescence ratio of the plant is compared with the fluorescence feature library data to obtain the second type identification result of the plant. The species of the plant is determined based on the results of the first and second species identification. Based on the plant species and the location information of the plant on the plant surface morphology image, the location information of weeds is determined so as to perform laser inactivation on the weeds.

2. The laser weeding method according to claim 1, characterized in that, Laser inactivation of weeds specifically involves controlling a laser to inactivate weeds with weed-removing energy power; The weeding energy power is as follows: ； in, K The crop coefficient in the plant; R The fluorescence ratio of the weeds to be inactivated; The stem thickness of the weeds is determined based on the weed segmentation mask; It is an energy correction factor; It is a natural number.

3. The laser weeding method according to claim 1, characterized in that, When the confidence level of the image segmentation model output is lower than the preset confidence level, the plant species is determined by combining the first species identification result and the second species identification result. If the first species identification result and the second species identification result are inconsistent, the second species identification result is used as the plant species. When the confidence level of the image segmentation model output is not lower than the preset confidence level, the plant species is determined based on the first species identification result.

4. The laser weeding method according to claim 1, characterized in that, The image segmentation model is an improved YOLOv8-seg architecture model, specifically improved as follows: For the backbone network in the YOLOv8-seg architecture model, StarBlock is used to replace the C2f module in the backbone network to reduce the number of parameters in the backbone network. For the neck network in the YOLOv8-seg architecture model, an ECA attention mechanism is set to enhance the ability to extract plant stem and leaf texture features.

5. The laser weeding method according to claim 4, characterized in that, An improved YOLOv8-seg architecture model, including a backbone network, a neck network, a head network, and a masked co-decoder; The backbone network takes a plant surface morphology image as input and outputs shallow, medium and deep feature maps of the plant. The shallow feature map contains detailed information about the plant, including lesions, bristles and thin stems. The medium feature map contains partial shapes of the plant. The deep feature map is used to characterize the overall surface morphology of the plant. The input to the neck network is the shallow feature map, middle feature map and deep feature map output by the backbone network. The neck network fuses the shallow feature map, middle feature map and deep feature map and outputs the fused multi-scale feature map. The input to the head network is the multi-scale feature map output by the neck network, and the output of the head network is the predicted bounding box of the plant, the plant species and the image segmentation confidence, and the generation of segmentation masks for crops, weeds and soil. The input to the masked co-decoder is the predicted plant bounding box, the plant species and the image segmentation confidence, as well as the segmentation masks for crops, weeds and soil. The output is the segmentation mask for crops, weeds or soil corresponding to each plant bounding box.

6. The laser weeding method according to claim 5, characterized in that: The improved YOLOv8-seg architecture model is pre-trained through the following steps: Construct a dataset of surface morphology images of different types of weeds; where the weed species are labeled in the surface morphology images. Data augmentation methods were employed to expand the weed surface morphology image dataset, including random rotation and brightness fluctuation. The expanded weed surface morphology image dataset was input into the improved YOLOv8-seg architecture model in batches for training over a period of 100 epochs. With the goal of minimizing the combined loss function, the network parameters of the improved YOLOv8-seg architecture model were updated by combining momentum and adaptive learning rate through the backpropagation algorithm Adam optimizer until the performance of the YOLOv8-seg architecture model converged. The combined loss function includes bounding box regression loss, classification loss, and confidence loss.

7. A method for monitoring plant status in the field, characterized in that, Specifically, the following steps are included: Collect fluorescence ratios and plant surface morphology images of the target area where the plant is located; The plant surface morphology image is input into the image segmentation model to obtain the first type identification result of the plant and the location information of the plant; The fluorescence ratio of plants in the target area is compared with the fluorescence feature database to obtain the second type identification result of the plants. The species of the plant is determined based on the results of the first and second species identification. Based on the plant species, a plant distribution map is constructed by combining the location of the plant in the plant surface morphology image.

8. The field plant condition monitoring method according to claim 7, characterized in that, Also includes: Based on the plant species, the first health diagnosis result of the crop is determined according to the real-time measured crop fluorescence ratio and / or the crop's maximum photochemical efficiency; wherein, the first health diagnosis result is used to characterize the crop's water stress diagnosis result; Based on the crop species, a second health diagnosis result is obtained from the maximum photochemical efficiency measured in real time; the second health diagnosis result is used to characterize the strength of the crop's photosynthesis. The first and second health diagnoses of the crop are marked on the plant distribution map to generate a field plant management map, so as to monitor the status of the plants in the field.

9. The field plant condition monitoring method according to claim 8, characterized in that, The method for obtaining the first health diagnosis result of crops is as follows: To determine whether the fluorescence ratio of crops is higher than that of crops in healthy conditions, a preliminary diagnosis can be made as to whether crops are under water stress. When a preliminary diagnosis indicates that the plant is under water stress, it is then determined whether the maximum photochemical efficiency of the crop is lower than the maximum photochemical efficiency of the crop in a healthy state. When the maximum photochemical efficiency of a crop is lower than the maximum photochemical efficiency of a crop in a healthy state, and the current non-optical quenching is higher than the normal value of non-optical quenching, and the soil moisture is lower than the normal moisture value, the crop is determined to be under water stress. If the current non-optical quenching is lower than the normal value for non-optical quenching and / or the soil moisture is higher than the normal value for moisture, then the crop is determined not to be under water stress. When the maximum photochemical efficiency is not lower than the maximum photochemical efficiency of the plant in a healthy state, the crop is determined not to be under water stress.

10. A laser weeding system, characterized in that, include: A dual-band excitation detection module is used to acquire fluorescence ratios and plant surface morphology images of plants in the target area. The multimodal recognition module is used to input the plant surface morphology image into the image segmentation model to determine the first type of plant identification result and the plant's position information on the plant surface morphology image. The fluorescence ratio of the plant was compared with the fluorescence feature library data to determine the second type of plant identification; the plant type was determined based on the first and second type identification results. The laser execution module is used to determine the location of weeds based on the plant species and the location information of the plant on the plant surface morphology image, so as to perform laser inactivation on the weeds.

Images