Sugarcane flowering state recognition method based on AI recognition
By constructing an AI-based method for identifying sugarcane flowering status, and integrating a dual-modal model of microbiome and visual features, the shortcomings of traditional sugarcane flowering status identification are solved, enabling real-time and accurate monitoring and prediction of sugarcane flowering status, thereby improving the scientific nature and efficiency of sugarcane cultivation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGXI ZHUANG AUTONOMOUS REGION ACAD OF AGRI SCI
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-09
AI Technical Summary
Traditional methods for identifying the flowering status of sugarcane rely on manual inspections, which are time-consuming, labor-intensive, and susceptible to subjective factors. They cannot achieve 24-hour continuous monitoring, are difficult to capture subtle changes, lack in-depth information on the physiological state of sugarcane, and cannot meet the needs of modern precision agricultural management.
An AI-based method for identifying the flowering status of sugarcane is adopted, which integrates microbiome features and visual features to construct a dual-modal model. A multi-task learning algorithm is used to predict the flowering trend and identify the current flowering status, and the model is deployed on edge computing devices or in the cloud for real-time analysis.
It significantly improves the accuracy and efficiency of identification, enables real-time and precise monitoring of sugarcane flowering status, provides scientific tools for flowering status identification and trend prediction, and enhances the level of precision agricultural management in sugarcane planting.
Smart Images
Figure CN122173832A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural intelligent monitoring technology, specifically to a method for identifying the flowering status of sugarcane based on AI recognition. Background Technology
[0002] With the advancement of agricultural modernization, precision agricultural management has become the key to improving crop yield and quality. As an important sugar crop and cash crop globally, the flowering status of sugarcane during its growth cycle directly affects sugar accumulation and final yield.
[0003] Traditional methods for identifying sugarcane flowering status mainly rely on manual inspection and experience-based judgment. This approach is not only time-consuming and labor-intensive, but also easily influenced by the observer's subjective factors, leading to inaccurate results. Furthermore, manual observation cannot achieve 24-hour continuous monitoring, making it difficult to capture subtle changes in sugarcane flowering status in a timely manner, thus missing the optimal time for agricultural operations. At the same time, traditional methods lack the ability to explore deeper information about the physiological state of sugarcane, such as changes in microbiome characteristics, which are crucial for accurately predicting sugarcane flowering trends. Therefore, traditional technologies are no longer sufficient to meet the needs of modern precision agricultural management.
[0004] Given the limitations of traditional methods for identifying the flowering status of sugarcane, it is therefore of great importance to develop an AI-based method for identifying the flowering status of sugarcane. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-based method for identifying the flowering status of sugarcane. It can construct a dual-modal model by integrating microbiome features and visual features, and use a multi-task learning algorithm to simultaneously predict the flowering trend of sugarcane and identify the current flowering status. Compared with traditional methods, this technology significantly improves the accuracy and efficiency of identification and realizes real-time and accurate monitoring of the flowering status of sugarcane.
[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a method for identifying the flowering status of sugarcane based on AI recognition, the specific steps of which are as follows:
[0007] S1. Sample collection: Collect sugarcane rhizosphere soil samples, leaf samples, and corresponding visual image samples of sugarcane spikes at different flowering stages.
[0008] S2. Microbiome feature extraction: Metagenomic sequencing was performed on rhizosphere soil samples and phyllosphere samples to obtain characteristic data such as species composition, relative abundance, and functional genes of microbial communities, and to construct a microbiome feature dataset.
[0009] S3. Visual Feature Extraction: Preprocess the visual image samples of sugarcane spikes to extract the visual features of texture, shape, and color, and construct a visual feature dataset.
[0010] S4. Construction and Training of the Bimodal Model: The microbiome feature dataset and the visual feature dataset are fused to construct a microbiome-AI recognition bimodal model. The model is trained using a multi-task learning algorithm, enabling the model to simultaneously predict flowering trends based on microbiome features and identify the current flowering status based on visual features.
[0011] S5. Model Deployment and Application: Deploy the trained bimodal model to edge computing devices or the cloud to perform real-time analysis on collected sugarcane rhizosphere and leaf microbial samples and flower spike visual images, and output flowering trend prediction results and current flowering status recognition results.
[0012] Furthermore, the samples collected in S1 cover the main sugarcane varieties and different ecological planting areas in tropical and subtropical regions, and the number of samples in each flowering state is no less than 50 groups; the rhizosphere soil samples are collected using the five-point sampling method, with a collection depth of 10-20cm; the phyllosphere samples are collected from the leaf surface of functional sugarcane leaves.
[0013] Furthermore, in S2, metagenomic sequencing was performed using the Illumina NovaSeq high-throughput sequencing platform. The sequencing data underwent quality control, assembly, and annotation processes to obtain microbial community characteristic data. The quality control steps involved removing reads with a Q value <20 and reads with a length <50bp. Assembly was performed using MEGAHIT software. Gene prediction was performed using Prodigal software. Species annotation and functional annotation were completed by comparison with the NCBINR and KEGG databases.
[0014] Furthermore, in S3, the preprocessing of visual image samples employs Gaussian filtering for noise reduction, CLAHE enhancement for brightness and contrast, adaptive histogram equalization with limited contrast, and semantic segmentation based on the DeepLabv3+ model for spike region segmentation. Visual feature extraction utilizes an improved ResNet-50 convolutional neural network, which replaces some convolutional layers with depth-separable convolutional layers to reduce the number of model parameters. The extracted visual features include texture features, shape features, and color features.
[0015] Furthermore, the microbiome-AI bimodal model in S4 includes two task branches: a flowering trend prediction branch and a current flowering state recognition branch; the algorithm formula for the flowering trend prediction branch is: ,in This is a predicted probability distribution for the flowering stage over the next 7 days. The Softmax activation function is used. This is a weight matrix representing the characteristics of the microbiome. For microbiome feature vectors, For bias terms, For the number of key species in the microbial community, For the first The attention weights of key species were obtained by learning the attention mechanism from microbiome features. The magnitude of the attention weights was significantly positively correlated with the abundance changes of the species during sugarcane flowering. For the first Abundance feature functions of key species; the algorithm formula for identifying the current flowering state branch is: ,in This is the predicted probability distribution of the current flowering state. The weight matrix for visual features. For visual feature vectors, For bias terms, The fusion weights for microbiome and visual features are obtained through adaptive learning by minimizing the multi-task loss function during training. The function for fusing microbiome features and visual features is denoted by cosine similarity, which is used to calculate the feature correlation between the two. The loss function for multi-task learning is: ,in For the cross-entropy loss of the flowering trend prediction task, The cross-entropy loss is used for the current flowering state identification task. , The loss weights were set based on the initial performance difference between the two tasks in the pre-experiment, with initial values of 0.6 and 0.4 respectively. During training, the weights were adaptively adjusted every 10 rounds based on the performance of the validation set.
[0016] Furthermore, the multi-task learning algorithm in S4 adopts a network structure with shared hard parameters. The two task branches share a 3-layer feature fusion layer at the bottom layer. The calculation formula for the feature fusion layer is as follows: ,in The fused feature vector Represents the feature vector of the microbiome and visual feature vectors splicing, This is the weight matrix of the fusion layer. For bias terms, The activation function is 0.001. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training epochs. The model performance is evaluated using a validation set after each training epoch. Training stops when the accuracy on the validation set no longer improves.
[0017] Furthermore, the edge computing device in S5 is an embedded system of a field inspection robot equipped with NVIDIA Jetson Xavier NX; the robot is equipped with a micro sequencing device that can quickly sequence microbial samples; after the model is deployed, the response time for analyzing the input sample data is no more than 5 seconds, and the results are sent to the grower's mobile APP through a wireless communication module.
[0018] Furthermore, the system also includes a data storage module and a visualization module. The data storage module is used to store the collected sample data, extracted feature data, and intermediate data during model training. The visualization module is used to visualize the flowering trend prediction results and the current flowering status recognition results in the form of charts, image annotations, etc. The image annotations use bounding boxes and text descriptions to mark the position of the flower spikes and indicate the flowering status on the visual image.
[0019] Compared with existing technologies, this AI-based method for identifying the flowering status of sugarcane has the following advantages:
[0020] I. This method constructs a bimodal model by integrating microbiome features and visual features. It utilizes a multi-task learning algorithm to simultaneously predict sugarcane flowering trends and identify the current flowering status. Compared with traditional single-modal recognition methods, this method significantly improves the accuracy and efficiency of recognition. Microbiome features provide in-depth information on the physiological state of sugarcane, while visual features directly reflect the appearance of the flowering status. The combination of the two enables the model to maintain high recognition accuracy at different flowering stages. At the same time, the model is deployed on edge computing devices or in the cloud, realizing real-time analysis of sugarcane flowering status with a response time of no more than 5 seconds, which greatly improves recognition efficiency.
[0021] Second, the application of this method provides sugarcane growers with a scientific and accurate tool for identifying flowering status and predicting flowering trends, which helps to achieve precision agricultural management. By monitoring the flowering status of sugarcane in real time, growers can adjust irrigation, fertilization and other agricultural operations in a timely manner, optimize resource allocation, and improve sugarcane yield and quality. In addition, the visualization module presents the identification results and predicted trends in the form of charts, image annotations and other forms, which is convenient for growers to understand and apply, further improving the scientificity and effectiveness of agricultural management.
[0022] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following, or may be learned from the practice of the invention. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0024] Figure 1 This is a flowchart illustrating the AI-based method for recognizing the flowering status of sugarcane.
[0025] Figure 2 This is a schematic diagram of the core process of an AI-based method for recognizing the flowering status of sugarcane. Detailed Implementation
[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0027] Example 1
[0028] S1, Sample Collection
[0029] This sample collection covered three major sugarcane varieties in tropical growing areas, encompassing four typical flowering stages: non-flowering, early flowering, full bloom, and post-flowering. Sixty samples were collected for each flowering stage, exceeding the minimum requirement of 50 samples to increase the sample size. This effectively reduces errors caused by sample randomness and improves the representativeness and reliability of the data for subsequent model training. Rhizosphere soil samples were collected using a five-point sampling method, at the four corners and center of each sugarcane field. The sampling depth was strictly controlled between 10-20 cm. This depth range accurately captures the living environment samples of active microorganisms around the sugarcane roots, avoiding interference from surface microorganisms due to shallow sampling or missing the main functional areas of the roots due to excessive depth. After collection, stones, weeds, and other debris were immediately removed. The samples are sealed and preserved to prevent changes in the microbial community structure during exposure, ensuring the authenticity of the microbial composition. Leaf samples are taken from the upper and middle functional leaves of the sugarcane plant, where photosynthesis is vigorous and material exchange with the external environment is frequent. The types and quantities of microorganisms on the leaf surface better reflect the sugarcane's growth status. Samples are obtained by gently wiping the leaf surface with a sterile cotton swab and then sealed for preservation, avoiding contamination by external microorganisms that could affect subsequent sequencing results. Simultaneously, high-definition cameras with resolutions of 16 megapixels or higher are used to capture visual images of the corresponding sugarcane spikes under natural light conditions. This ensures that the images clearly present the size, color, and morphology of the spikes, providing high-quality raw data for subsequent visual feature extraction and reducing feature extraction bias caused by image blur.
[0030] S2, Microbiome Feature Extraction
[0031] Rhizosphere and phyllosphere soil samples were collected and sent to a qualified molecular biology laboratory for metagenomic sequencing using the Illumina NovaSeq high-throughput sequencing platform. This platform offers advantages such as high sequencing throughput, long read lengths, and high accuracy, efficiently acquiring gene sequence information of microorganisms in the samples and providing massive amounts of basic data for subsequent species and functional analysis. The raw sequencing data underwent quality control processing, strictly removing reads with Q values less than 20 and reads shorter than 50 bp. This step filters out low-quality data and improves the accuracy of subsequent analysis results. Subsequently, the quality-controlled sequences were assembled using MEGAHIT software, which can efficiently process massive amounts of short-read sequences and achieve precise sequence assembly. The microbial genome fragment information was restored; gene prediction was performed using Prodigal software, which is optimized for the characteristics of prokaryotic gene structure and can accurately identify coding genes in the sequence, laying the foundation for subsequent functional annotation. Finally, the processed sequences were compared with the NCBINR database to complete species annotation, clarifying the types of microorganisms in the sample and the relative abundance of each species. Functional annotation was also completed by comparing with the KEGG database to understand the types of functional genes of microorganisms and the metabolic processes they participate in. In the end, comprehensive characteristic data such as species composition, relative abundance, and functional genes of the microbial community were obtained, and a microbial feature dataset was constructed. This dataset can reflect the relationship between sugarcane growth and flowering at the microbial level, providing key feature support for the model to predict flowering trends.
[0032] S3, Visual Feature Extraction
[0033] The collected visual images of sugarcane flower spikes were preprocessed. First, Gaussian filtering was used to remove noise interference from the images. This method can smooth noise while preserving image details, making the images clearer. Then, the CLAHE technique was used to enhance the brightness and contrast of the images. This technique can adaptively adjust to the differences in illumination in different areas of the image, solving the problem of uneven brightness in the flower spike area caused by shooting angle and uneven lighting, and highlighting the difference between the flower spikes and the background. Finally, the DeepLabv3+ model was used for semantic segmentation. This model has the characteristics of high accuracy and strong robustness in image segmentation tasks, and can accurately separate the flower spike area from the background, avoiding background information from interfering with subsequent feature extraction, and ensuring that only the flower spike area is extracted. The analysis focuses on the visual feature extraction using an improved ResNet-50 convolutional neural network. This network replaces some convolutional layers with depthwise separable convolutional layers, which significantly reduces the number of network parameters and computational cost while maintaining the feature extraction effect, thus improving the feature extraction efficiency. At the same time, it retains the deep feature extraction capability of the original ResNet-50 network, which can accurately extract texture, shape, and color features from the preprocessed flower spike images. These features are integrated to construct a visual feature dataset, which can directly reflect the current flowering state of sugarcane from the appearance level, providing core feature basis for the model to identify the current flowering state.
[0034] S4. Construction and Training of Bimodal Models
[0035] A dual-modal model for microbiome-AI recognition was constructed, comprising a flowering trend prediction branch and a current flowering status recognition branch. The algorithm formula for the flowering trend prediction branch is as follows: ,in This is a predicted probability distribution for the flowering stage over the next 7 days. The Softmax activation function is used. This is a weight matrix representing the characteristics of the microbiome. For microbiome feature vectors, For bias terms, For the number of key species in the microbial community, For the first Attention weights for key species For the first The abundance feature functions of key species, and the algorithm formula for identifying the branch of the current flowering state are as follows: ,in This is the predicted probability distribution of the current flowering state. The weight matrix for visual features. For visual feature vectors, For bias terms, The fusion weights for microbiome and visual features, This paper proposes a fusion function for microbiome and visual features, employing a network structure with shared hard parameters. The two task branches share three feature fusion layers at the bottom layer. The feature fusion process uses the tanh activation function, which maps the fused feature values to the interval [-1, 1], enhancing the nonlinear expressive power of the features while avoiding the gradient vanishing problem and improving the stability of model training. The shared hard parameter structure allows the two tasks to share the bottom-level feature extraction capabilities, reducing model parameter redundancy, improving training efficiency, and promoting information interaction between the two tasks. This enables the model to comprehensively utilize the correlation information between microbiome and visual features to improve prediction and recognition accuracy.
[0036] During training, a multi-task learning algorithm is employed, using cross-entropy loss as the loss calculation method for both the flowering trend prediction task and the current flowering state recognition task. Cross-entropy loss effectively measures the difference between the model's predicted probability distribution and the true label distribution, making it suitable for loss calculation in classification tasks. The formula is as follows: ,in For the cross-entropy loss of the flowering trend prediction task, The cross-entropy loss is used for the current flowering state identification task. , The loss weights are used to weight the losses of the two tasks. The total loss of multi-task learning is obtained by weighting and summing the losses of the two tasks. The weights can be adjusted according to actual needs to balance the training priority of the two tasks. The model parameters are adjusted by backpropagation algorithm based on the total loss to complete the training. This enables the model to accurately predict the flowering trend of the next 7 days based on microbiome features and accurately identify the current flowering status based on visual features, providing reliable model support for subsequent real-time field monitoring.
[0037] S5, Model Deployment and Application
[0038] The trained bimodal model was deployed onto an embedded system of a field inspection robot equipped with NVIDIA Jetson Xavier NX. This embedded system possesses powerful edge computing capabilities, meeting the computational power requirements for real-time model inference, while also being small in size and low in power consumption, making it suitable for mobile field operations. The robot is equipped with a micro-sequencing device, enabling rapid sequencing of collected microbial samples in the field, achieving on-site processing and feature extraction of microbial samples. When inspecting sugarcane fields in tropical growing areas along a preset path, the robot automatically collects rhizosphere soil and leaf samples via its robotic arm, and captures real-time visual images of the flower spikes using its onboard camera. The sample data is input into the model for analysis in real time, with a response time of no more than 5 seconds, allowing for rapid output of analysis results and avoiding excessive waiting time. The system has a significant impact on field management decisions. After analysis, the flowering trend prediction results and current flowering status identification results are sent to the grower's mobile APP via a 4G / 5G wireless communication module, allowing growers to view the data anytime, anywhere and keep abreast of sugarcane flowering dynamics. At the same time, the data storage module stores the collected sample raw data, extracted feature data, and intermediate data during model training, facilitating subsequent data backtracking, model optimization, and scientific research analysis. The visualization module presents the prediction results in the form of charts such as line graphs and bar charts, accurately marking the position of the flower spikes in the visual image with bounding boxes and annotating the flowering status with text descriptions, enabling growers to intuitively understand the analysis results. This provides a scientific basis for water and fertilizer management, pest and disease control, and harvesting time planning for sugarcane in tropical growing areas, improving the level of refined management in sugarcane planting.
[0039] Example 2
[0040] S1, Sample Collection
[0041] This sample collection covered three major sugarcane varieties in tropical growing areas, encompassing four flowering stages: non-flowering, early flowering, full bloom, and post-flowering. Sixty samples were collected for each flowering stage, ensuring a minimum of 50 samples were collected. Rhizosphere soil samples were collected using a five-point sampling method, at the four corners and center of each sugarcane field, at a depth of 10-20 cm. After collection, stones, weeds, and other impurities were removed, and the samples were sealed and preserved. Leaf samples were collected from functional leaves in the upper and middle parts of the sugarcane plant. The leaf surface was gently wiped with a sterile cotton swab to obtain the sample, which was also sealed and preserved. Simultaneously, high-definition cameras were used to capture visual images of the corresponding sugarcane spikes under natural light conditions, ensuring the images clearly presented the complete morphology of the spikes.
[0042] S2, Microbiome Feature Extraction
[0043] Rhizosphere and phyllosphere soil samples were collected and sent to the laboratory for metagenomic sequencing using the Illumina NovaSeq high-throughput sequencing platform. The raw sequencing data were first subjected to quality control processing to remove reads with a Q value less than 20 and reads shorter than 50 bp to ensure data quality. Then, MEGAHIT software was used to assemble the quality-controlled sequences, and Prodigal software was used to perform gene prediction. Finally, the processed sequences were compared with the NCBINR database to complete species annotation and with the KEGG database to complete functional annotation. Ultimately, the species composition, relative abundance, functional genes, and other characteristic data of the microbial community were obtained, and a microbial community characteristic dataset was constructed.
[0044] S3, Visual Feature Extraction
[0045] The collected visual images of sugarcane flower spikes were preprocessed. First, Gaussian filtering was used to remove noise interference from the images. Then, the CLAHE technique was used to enhance the brightness and contrast of the images. Finally, the DeepLabv3+ model was used for semantic segmentation to accurately separate the flower spike region from the background. Visual feature extraction adopted an improved ResNet-50 convolutional neural network. This network replaced some convolutional layers with depth-separable convolutional layers on the basis of the original ResNet-50. Texture features, shape features, and color features were extracted from the preprocessed flower spike images, and these features were integrated to construct a visual feature dataset.
[0046] S4. Construction and Training of Bimodal Models
[0047] A dual-modal model for microbiome-AI recognition was constructed, comprising a flowering trend prediction branch and a current flowering state recognition branch. A network structure with shared hard parameters was adopted, with both task branches sharing a three-layer feature fusion layer at the bottom. The feature fusion process employed the tanh activation function. During training, a multi-task learning algorithm was used, with cross-entropy loss calculated for both the flowering trend prediction and current flowering state recognition tasks. The losses of the two tasks were weighted and summed to obtain the total loss for multi-task learning. The model parameters were adjusted based on the total loss to complete training, enabling the model to predict flowering trends based on microbiome features and recognize the current flowering state based on visual features.
[0048] S5, Model Deployment and Application
[0049] The trained bimodal model was deployed to an embedded system of a field inspection robot equipped with NVIDIA Jetson Xavier NX. This robot is equipped with a microsequencing device, which can quickly sequence microbial samples collected in the field. When the robot inspects sugarcane fields in tropical growing areas, it collects real-time samples of sugarcane rhizosphere soil, leaf samples, and visual images of flower spikes. The sample data is input into the model for analysis. The model's response time to the input data is no more than 5 seconds. After the analysis is completed, the flowering trend prediction results and the current flowering status recognition results are sent to the grower's mobile APP via the wireless communication module. At the same time, the data storage module stores the collected sample data, extracted feature data, and intermediate data of model training. The visualization module presents the prediction results in the form of charts and marks the position of flower spikes and the flowering status on the visual images with bounding boxes and text descriptions.
[0050] Example 2: Flowering regulation scenario in sugarcane breeding in subtropical planting areas
[0051] S1, Sample Collection
[0052] Samples were collected from four main sugarcane varieties in subtropical growing areas, covering four flowering stages: non-flowering, early flowering, full bloom, and post-flowering. 55 samples were collected for each flowering stage, ensuring that there were no fewer than 50 samples for each stage. Rhizosphere soil samples were collected using a five-point sampling method in different blocks of the breeding experimental field, with a sampling depth of 10-20 cm. After collection, the samples were promptly sealed to avoid changes in the microbial community. Leaf samples were collected from the leaf surface tissue of functional sugarcane leaves, using sterile collection tools and sealed for preservation. Simultaneously, professional photography equipment was used to capture visual images of the flower spikes under uniform lighting conditions to ensure that the images clearly reflect the details of the flower spikes' growth status.
[0053] S2, Microbiome Feature Extraction
[0054] Rhizosphere soil and phyllosphere samples were sent to a professional sequencing laboratory for metagenomic sequencing using the Illumina NovaSeq high-throughput sequencing platform. The sequencing data processing workflow was strictly performed according to standards. First, quality control was carried out to remove reads with a Q value less than 20 and reads shorter than 50 bp. Then, the MEGAHIT software was used for sequence assembly, and the Prodigal software was used for gene prediction. Finally, the data were compared with the NCBINR and KEGG databases to complete species annotation and functional annotation, and to obtain characteristic information such as species composition, relative abundance, and functional genes of the microbial community, thus constructing a microbiome characteristic dataset.
[0055] S3, Visual Feature Extraction
[0056] The visual images of flower spikes are preprocessed by sequentially performing Gaussian filtering for noise reduction, CLAHE brightness and contrast enhancement, and then semantic segmentation is performed using the DeepLabv3+ model to accurately extract the flower spike region. An improved ResNet-50 convolutional neural network is used for visual feature extraction, in which some convolutional layers are replaced with depth-separable convolutional layers. Texture features, shape features, and color features are extracted from the flower spike region images, and these features are integrated to form a visual feature dataset.
[0057] S4. Construction and Training of Bimodal Models
[0058] A dual-modal model for microbiome-AI recognition was constructed, with a flowering trend prediction branch and a current flowering state recognition branch. A hard parameter-shared network structure was adopted, with both branches sharing a 3-layer feature fusion layer at the bottom. The feature fusion process used the tanh activation function. The model was trained using a multi-task learning algorithm, calculating the cross-entropy loss for the flowering trend prediction task and the cross-entropy loss for the current flowering state recognition task respectively. The two losses were weighted and summed to obtain the total loss. The model parameters were optimized based on the total loss, enabling the model to accurately predict the flowering trend based on microbiome features and identify the current flowering state based on visual features.
[0059] S5, Model Deployment and Application
[0060] The trained bimodal model was deployed onto an embedded system of a field inspection robot equipped with NVIDIA Jetson Xavier NX. This robot is equipped with a microsequencing device, which can quickly process microbial samples collected in the field. In sugarcane breeding fields in subtropical planting areas, the robot regularly collects rhizosphere soil samples, leaf samples, and visual images of flower spikes, which are then input into the model for analysis in real time. The analysis response time is controlled within 5 seconds. The analysis results are sent to the breeders' mobile APP via a wireless communication module. The data storage module stores various sample data, feature data, and intermediate data for model training. The visualization module presents the flowering trend prediction results in the form of charts, and the location of the flower spikes and the current flowering status are clearly indicated on the visual images through bounding boxes and text annotations, providing data support for flowering regulation in the breeding process.
[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for identifying the flowering status of sugarcane based on AI recognition, characterized in that, The specific steps of this method are as follows: S1. Sample collection: Collect sugarcane rhizosphere soil samples, leaf samples, and corresponding visual image samples of sugarcane spikes at different flowering stages. S2. Microbiome feature extraction: Metagenomic sequencing was performed on rhizosphere soil samples and phyllosphere samples to obtain characteristic data such as species composition, relative abundance, and functional genes of microbial communities, and to construct a microbiome feature dataset. S3. Visual Feature Extraction: Preprocess the visual image samples of sugarcane spikes to extract the visual features of texture, shape, and color, and construct a visual feature dataset. S4. Construction and Training of the Bimodal Model: The microbiome feature dataset and the visual feature dataset are fused to construct a microbiome-AI recognition bimodal model. The model is trained using a multi-task learning algorithm, enabling the model to simultaneously predict flowering trends based on microbiome features and identify the current flowering status based on visual features. S5. Model Deployment and Application: Deploy the trained bimodal model to edge computing devices or the cloud to perform real-time analysis on collected sugarcane rhizosphere and leaf microbial samples and flower spike visual images, and output flowering trend prediction results and current flowering status recognition results.
2. The method for identifying sugarcane flowering status based on AI recognition according to claim 1, characterized in that, The samples collected in S1 cover the main sugarcane varieties and different ecological planting areas in tropical and subtropical regions, and the number of samples in each flowering state is no less than 50 groups; the rhizosphere soil samples are collected using the five-point sampling method, with a sampling depth of 10-20cm; the phyllosphere samples are collected from the leaf surface of functional sugarcane leaves.
3. The method for identifying sugarcane flowering status based on AI recognition according to claim 1, characterized in that, In S2, metagenomic sequencing was performed using the Illumina NovaSeq high-throughput sequencing platform. The sequencing data underwent quality control, assembly, and annotation processes to obtain microbial community characteristic data. The quality control steps involved removing reads with a Q value <20 and reads with a length <50bp. Assembly was performed using MEGAHIT software. Gene prediction was performed using Prodigal software. Species annotation and functional annotation were completed by comparison with the NCBINR and KEGG databases.
4. The method for identifying sugarcane flowering status based on AI recognition according to claim 1, characterized in that, In S3, the preprocessing of visual image samples employs Gaussian filtering for noise reduction, CLAHE enhancement for brightness and contrast, and semantic segmentation based on the DeepLabv3+ model for spike region segmentation. Visual feature extraction utilizes an improved ResNet-50 convolutional neural network, which replaces some convolutional layers with depth-separable convolutional layers in the original ResNet-50. The extracted visual features include texture features, shape features, and color features.
5. The method for identifying sugarcane flowering status based on AI recognition according to claim 1, characterized in that, The microbiome-AI bimodal model in S4 includes two task branches: a flowering trend prediction branch and a current flowering status recognition branch; the algorithm formula for the flowering trend prediction branch is: ,in This is a predicted probability distribution for the flowering stage over the next 7 days. The Softmax activation function is used. This is a weight matrix representing the characteristics of the microbiome. For microbiome feature vectors, For bias terms, For the number of key species in the microbial community, For the first Attention weights for key species For the first Abundance feature functions of key species; the algorithm formula for identifying the current flowering state branch is: ,in This is the predicted probability distribution of the current flowering state. The weight matrix for visual features. For visual feature vectors, For bias terms, The fusion weights for microbiome and visual features, This is a fusion function of microbiome features and visual features; The loss function for multi-task learning is: ,in For the cross-entropy loss of the flowering trend prediction task, The cross-entropy loss is used for the current flowering state identification task. , This is the loss weight.
6. The method for identifying sugarcane flowering status based on AI recognition according to claim 1, characterized in that, The multi-task learning algorithm in S4 adopts a network structure with shared hard parameters. The two task branches share a 3-layer feature fusion layer at the bottom layer. The calculation formula for the feature fusion layer is as follows: ,in The fused feature vector Represents the feature vector of the microbiome and visual feature vectors splicing, This is the weight matrix of the fusion layer. For bias terms, This is the activation function.
7. The method for identifying sugarcane flowering status based on AI recognition according to claim 1, characterized in that, The edge computing device in S5 is an embedded system of a field inspection robot equipped with NVIDIA Jetson Xavier NX; the robot is equipped with a micro sequencing device that can rapidly sequence microbial samples; After the model is deployed, the response time for analyzing the input sample data is no more than 5 seconds, and the results are sent to the grower's mobile APP via a wireless communication module.
8. The method for identifying sugarcane flowering status based on AI recognition according to claim 1, characterized in that, The system also includes a data storage module and a visualization module. The data storage module is used to store the collected sample data, extracted feature data, and intermediate data during model training. The visualization module is used to visualize the flowering trend prediction results and the current flowering status recognition results in the form of charts, image annotations, etc. The image annotations use bounding boxes and text descriptions to mark the position of the flower spikes and indicate the flowering status on the visual image.