Method for identifying and systematically monitoring endangered plant phoebe hainanensis based on multi-source data fusion

Abstract

The invention discloses an endangered plant phoebe hainanensis identification and systematic monitoring method based on multi-source data fusion, and belongs to the technical field of endangered plant protection. According to the method, a potential distribution area is delineated through a species distribution model, individual accurate identification and filing are achieved through a multi-branch feature fusion model, multi-scale growth and microenvironment data are collected through Internet of Things nodes, and finally a'discovery-identification-monitoring-protection 'whole-process closed loop is formed through a unified data management platform. According to the method, phenological feature extraction, transfer learning modeling, man-machine collaborative verification and low-power-consumption Internet of Things monitoring technologies are adopted, and the problems that a traditional method is low in efficiency, insufficient in recognition precision and fragmented in monitoring are solved. According to the method, efficient discovery, accurate identification and long-term dynamic monitoring of the wild population of the phoebe hainanensis are achieved, scientific support is provided for protection decision making, and the method is convenient to operate and high in practicability and has protection value and application prospects.

Description

Technical Field

[0001] This invention relates to the field of endangered plant protection technology, specifically to a method for identifying and systematically monitoring the endangered plant *Machilus hainanensis* based on multi-source data fusion. Background Technology

[0002] *Horsfieldia hainanensis* Merr., a unique evergreen tree endemic to China and belonging to the genus *Horsfieldia* of the family Myristaceae, is a Class II protected endangered plant in China and a landmark species of tropical rainforest ecosystems. It possesses irreplaceable ecological and scientific value in maintaining regional biodiversity and studying the composition and geographical distribution of tropical rainforest flora. This species not only has significant economic value—its seeds have a high oil content, making them ideal for synthesizing specialty chemical raw materials, and its wood can be used for high-end furniture and decorative materials—but also possesses beautiful tree shapes and medicinal properties, being used in traditional Li ethnic minority practices for postpartum nourishment and treatment of infantile malnutrition, demonstrating its high potential for comprehensive utilization. However, due to both its own biological characteristics and external environmental stresses, the wild population of *Horsfieldia hainanensis* is intermittently distributed in a "island-like" or "spot-like" pattern in limited areas of Yunnan, Hainan, and Guangxi Zhuang Autonomous Region in my country. The population is dwindling, and natural regeneration is difficult. It has been included in the "National Implementation Plan for the Rescue and Protection of Wild Plants with Extremely Small Populations," and its endangered status urgently requires systematic protection through scientific methods.

[0003] Currently, conservation efforts for *Machilus hainanensis* still face numerous technical bottlenecks. Traditional conservation models and existing technologies are insufficient to meet the demands for precise, end-to-end conservation. In the wild population discovery phase, due to the species' sparse distribution and concealed habitat, traditional manual field surveys rely on on-site reconnaissance by staff, resulting in low efficiency, high costs, and limited coverage. Furthermore, they struggle to overcome terrain limitations to comprehensively investigate potential distribution areas, leading to many potential populations going undiscovered and missed conservation opportunities. Simultaneously, manual surveys are significantly affected by seasonal and weather factors, making it impossible to achieve routine, continuous population tracking and thus hindering long-term conservation decision-making.

[0004] In terms of species identification, existing plant identification technologies have significant limitations. General plant identification patents often design universal identification models for common plants, relying on massive amounts of sample data for training. However, as an endangered species, *Machilus hainanensis* has a limited sample size and unique characteristics, making it difficult for general models to fully learn its species-specific features, resulting in low identification accuracy and a high false positive rate. Some identification technologies rely solely on images of a single part, failing to consider the morphological variations of organs at different growth stages of *Machilus hainanensis*, further reducing the stability and resistance to interference in identification.

[0005] In the growth monitoring and protection management stages, existing technologies are mostly fragmented, failing to form a complete system. Single-function monitoring patents focus only on monitoring specific indicators such as soil quality and plant growth, ignoring the correlation between plant growth status and microenvironmental factors, and cannot comprehensively reflect the survival status of *Machilus hainanensis*. Simultaneously, data from identification, monitoring, and management stages are isolated, forming "data silos," lacking a unified integration and analysis platform, making it difficult to effectively transform monitoring data into a basis for protection decisions. For example, existing monitoring devices cannot simultaneously collect and correlate the phenological changes of large trees, seedling growth dynamics, and microenvironmental parameters such as soil temperature and humidity, and light intensity. This makes it impossible to promptly detect risks such as seedling growth stagnation and abnormal environmental factors, and also makes it difficult to scientifically guide seed collection timing and artificial intervention measures based on monitoring data, resulting in passive and inefficient protection work.

[0006] Furthermore, a systematic conservation technology system for plants with extremely small populations is still incomplete. Existing technologies mostly focus on current status surveys, lacking long-term data accumulation and in-depth analysis of the growth and development patterns and endangerment mechanisms of endangered species. *Machilus hainanensis* suffers from internal endangerment factors such as narrow ecological adaptability, low fruit set, difficulty in seed germination, and weak seedling competitiveness. Simultaneously, it faces external threats such as climate change, habitat fragmentation, over-logging, and interspecific competition. Therefore, a systematic technical solution integrating the entire process of "discovery-identification-monitoring-management" is urgently needed to achieve multi-scale coverage from macro-population location to micro-individual monitoring, providing precise and efficient technical support for its rescue and protection. Summary of the Invention

[0007] To address the aforementioned shortcomings, this invention provides a method for the identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion. By integrating multi-source data fusion, multi-branch identification, and IoT monitoring, this method solves the problems of fragmented search and supervision, low identification accuracy, and unsystematic monitoring of the endangered *Machilus hainanensis*.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0009] A method for identifying and systematically monitoring the endangered plant *Machilus hainanensis* based on multi-source data fusion includes the following steps:

[0010] S1. Delineation of potential distribution area: Obtain multi-temporal remote sensing images of the target area, extract the distinguishable phenological characteristics of the Hainan windblown magnolia during the flowering and fruiting periods, and combine the known distribution points and environmental data to predict its potential distribution area through a species distribution model;

[0011] S2. Multi-feature fusion recognition and individual authentication: Within the potential distribution area, multi-angle images of the flowers, leaves, and bark of the plant to be identified are simultaneously collected; the images of multiple parts are input into a pre-trained multi-branch feature fusion recognition model, which outputs the species identification result and confidence level by fusing and jointly judging features from different branches; for the identified plants, their spatial location and morphological parameters are recorded, and an electronic file with a unique identity is created.

[0012] S3. IoT-assisted multi-scale monitoring: IoT monitoring nodes are deployed around the documented plants. The nodes continuously collect images of the canopy and seedling status of the target plants, as well as sensor data of their growth microenvironment.

[0013] S4. Systematic protection and management closed loop: All monitoring data collected in step S3 is transmitted in real time to a unified data management platform. The platform integrates individual profiles, monitoring data streams, and analysis and early warning models to achieve a closed loop throughout the entire process from species discovery and accurate identification to long-term dynamic monitoring and protection decision support.

[0014] Preferably, the species distribution model in step S1 adopts the maximum entropy model; the distinguishable phenological features include specific vegetation indices and texture features calculated based on the red-edge band.

[0015] Preferably, the multi-branch feature fusion recognition model in step S2 is a convolutional neural network with at least three independent feature extraction branches. Each branch is used to extract features from flower, leaf, and bark images, and the features are fused through feature splicing or attention mechanisms.

[0016] Preferably, the feature extraction branch of the convolutional neural network is fine-tuned by transfer learning using an EfficientNet or ResNet network pre-trained on the ImageNet dataset as the backbone network.

[0017] Preferably, the method further includes a human-machine collaborative verification step: when the recognition confidence of the model is lower than a preset threshold, the images of multiple parts of the corresponding plant are uploaded to a cloud review system for remote confirmation by experts, and the confirmation results are fed back to optimize the recognition model.

[0018] Preferably, in step S3, the IoT monitoring node includes a solar power supply unit, an image acquisition unit, an environmental sensor unit, and a low-power wide area network communication unit; the image acquisition unit is used to take panoramic shots of the canopy and close-ups of seedlings at regular intervals or triggered actions, and the environmental sensor unit includes at least a soil temperature and humidity sensor and a photosynthetically active radiation sensor.

[0019] Preferably, in step S4, the data management platform automatically tracks the phenological process of the Hainan windblown magnolia by analyzing continuous canopy images, and quantifies its canopy width and leaf area index.

[0020] Preferably, in step S4, the analysis and early warning model automatically generates and sends early warning information when it identifies the following abnormal situations: seedling images show stagnation of growth or abnormal leaf condition in multiple consecutive periods; the occurrence time of key phenological periods of adult plants deviates from historical patterns by more than a preset threshold; and continuously monitored microenvironment data exceeds one or more of the known suitable growth range of Hainan windblown nanmu.

[0021] Preferably, in step S4, the protection decision support specifically includes: guiding the formulation of a scientific seed collection plan based on the optimal flowering and fruiting period data obtained from platform analysis; and triggering targeted manual inspections or ecological intervention measures based on early warnings of abnormal growth for specific individuals.

[0022] Preferably, the method further includes a data-driven archive update and knowledge accumulation step: the monitoring records, analysis results and protection measures generated in step S4 are dynamically linked and updated to the electronic archives of the corresponding individuals to form a traceable and analyzable protection knowledge base.

[0023] Compared with the prior art, the present invention has the following advantages and technical effects:

[0024] I. Construct a closed-loop system covering the entire process to solve the problem of fragmented functions in existing technologies.

[0025] Existing technologies generally suffer from "data silos" and "functional fragmentation." General plant identification patents focus only on a single identification stage, lacking an effective connection between pre-screening populations and subsequent continuous monitoring. Single-function monitoring patents are mostly limited to monitoring local indicators such as soil and growth, failing to form a complete technological chain from species discovery to conservation decisions. This invention innovatively constructs a closed-loop process encompassing "potential distribution area delineation - multi-feature fusion identification - IoT multi-scale monitoring - systematic protection management." Through the organic connection and data correlation of each stage, it achieves integrated protection of *Machilus hainanensis* in terms of "searching, identification, monitoring, management, and preservation." This closed-loop system not only solves the problem of low conservation efficiency caused by the fragmentation of identification, monitoring, and management stages in existing technologies, but also provides a systematic solution for endangered plant protection through the seamless integration of data throughout the entire process, filling the gap in technology for the full life-cycle protection of specific endangered tree species.

[0026] II. Innovate multi-dimensional technology integration approaches to enhance the accuracy and targeting of identification and monitoring.

[0027] In the population discovery phase, existing technologies mostly employ general vegetation remote sensing monitoring methods, failing to fully incorporate the unique phenological characteristics of endangered plants, resulting in insufficient accuracy in predicting potential distribution areas. This invention specifically extracts specific vegetation indices and texture features based on the red-edge band during the flowering and fruiting periods of *Machilus hainanensis*, combining this with the species distribution prediction capabilities of the maximum entropy model to achieve precise delineation of potential distribution areas. This significantly reduces the blind spots in field surveys and solves the technical challenges of sparse distribution of endangered plant populations and limited coverage in traditional surveys.

[0028] In the individual identification stage, existing general plant identification models suffer from insufficient accuracy to meet conservation requirements due to a lack of training with specific samples of endangered plants and a failure to consider the limitations of single-part features. This invention innovatively designs a multi-branch feature fusion identification model. It extracts high-dimensional features from flowers, leaves, and bark through independent branches and combines these features with fusion strategies such as feature splicing or attention mechanisms to achieve joint discrimination of features from multiple parts. Simultaneously, it employs EfficientNet or ResNet, pre-trained on ImageNet, as the backbone network for transfer learning, effectively overcoming the problem of insufficient model training caused by the scarcity of Hainan Phoebe zhennan sample data, thus improving the accuracy and robustness of identification. The introduction of a human-machine collaborative verification mechanism further constructs a virtuous cycle of machine identification, expert verification, and model upgrades through expert review and iterative model optimization, solving the technical bottleneck of high-confidence confirmation in rare species identification.

[0029] In the growth monitoring stage, existing technologies mostly employ single-scale monitoring and lack targeted monitoring solutions for vulnerable stages such as seedlings. This invention utilizes multi-dimensional data collection through IoT monitoring nodes to simultaneously acquire images of the canopy and seedlings, as well as microenvironmental data such as soil temperature and humidity, and photosynthetically active radiation. This enables full-age monitoring from mature plants to seedlings, and multi-dimensional coverage from plant growth status to the ecological environment. This multi-scale monitoring mode effectively solves the technical shortcomings of traditional monitoring methods, such as difficulty in capturing seedling growth dynamics and the inability to correlate environmental factors with plant growth.

[0030] III. Strengthen the scientific and proactive nature of conservation decisions to enhance the effectiveness of endangered plant conservation.

[0031] Existing technologies primarily focus on collecting current data, lacking in-depth data mining and effective support for conservation decisions, resulting in mostly reactive conservation measures. This invention's data management platform integrates individual profiles, monitoring data streams, and analytical early warning models. Through intelligent analysis of continuous canopy images, it tracks phenological progress and dynamically quantifies canopy width and leaf area index, providing precise data support for conservation decisions. The analytical early warning model automatically issues warnings for key risk points such as seedling growth stagnation, abnormal phenological stages, and microenvironmental exceedances, shifting conservation efforts from "passive response" to "proactive prediction," effectively addressing the problems of untimely and insufficiently targeted interventions in endangered plant conservation.

[0032] Meanwhile, the platform-based analysis of optimal seed collection timing and targeted human intervention triggers scientific and precise protection measures; data-driven archive updates and knowledge accumulation form a traceable and analyzable protection knowledge base, providing long-term data support for in-depth research on the growth and development patterns and endangered mechanisms of Hainan nanmu, further optimizing protection strategies, achieving a virtuous cycle of "monitoring-analysis-decision-intervention-optimization", and enhancing the systematic and long-term effectiveness of endangered plant protection. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the steps of the method for identifying and systematically monitoring the endangered plant *Machilus hainanensis* based on multi-source data fusion, as described in this invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0035] The image recognition, feature fusion, and phenological analysis algorithms involved in this invention can all be implemented based on existing open-source tools and well-known algorithms. For example, convolutional neural network construction and training can use frameworks such as TensorFlow and PyTorch; attention mechanisms can be implemented with reference to the Transformer standard; image segmentation can use OpenCV or pre-trained segmentation models; and phenological extraction can employ time series analysis libraries. Those skilled in the art can adjust model parameters and thresholds according to specific data characteristics, and can implement this invention without creative effort.

[0036] In this invention, a method for identifying and systematically monitoring the endangered plant *Machilus hainanensis* based on multi-source data fusion includes the following steps:

[0037] S1. Delineation of potential distribution area: Obtain multi-temporal remote sensing images of the target area, extract the distinguishable phenological characteristics of the Hainan windblown magnolia during the flowering and fruiting periods, and combine the known distribution points and environmental data to predict its potential distribution area through a species distribution model;

[0038] S2. Multi-feature fusion recognition and individual authentication: Within the potential distribution area, multi-angle images of the flowers, leaves, and bark of the plant to be identified are simultaneously collected; the images of multiple parts are input into a pre-trained multi-branch feature fusion recognition model, which outputs the species identification result and confidence level by fusing and jointly judging features from different branches; for the identified plants, their spatial location and morphological parameters are recorded, and an electronic file with a unique identity is created.

[0039] S3. IoT-assisted multi-scale monitoring: IoT monitoring nodes are deployed around the documented plants. The nodes continuously collect images of the canopy and seedling status of the target plants, as well as sensor data of their growth microenvironment.

[0040] S4. Systematic protection and management closed loop: All monitoring data collected in step S3 is transmitted in real time to a unified data management platform. The platform integrates individual profiles, monitoring data streams, and analysis and early warning models to achieve a closed loop throughout the entire process from species discovery and accurate identification to long-term dynamic monitoring and protection decision support.

[0041] Technical principle of the invention:

[0042] I. Delineation of Potential Distribution Areas

[0043] The core of this step is to accurately predict the potential distribution area of ​​*Machilus hainanensis* by leveraging the synergistic effect of species distribution models and multi-dimensional data. Multi-temporal remote sensing imagery of the target area provides the fundamental data source for phenological feature extraction. The red-edge bands within these images are sensitive to changes in plant physiological states; specific vegetation indices calculated based on these bands can accurately capture the unique spectral responses of *Machilus hainanensis* during its flowering and fruiting periods. Texture features reflect the differentiated manifestations of the plant canopy structure. Together, these constitute the core phenological markers distinguishing it from other species. Known distribution point data provides real distribution samples for the model, while environmental data (such as topography, climate, and soil type) clarifies the suitable habitat thresholds for species survival. These two types of data, along with phenological features, are input into a maximum entropy model. This model quantifies the correlation probability between environmental variables and species distribution, maximizing the use of limited sample information and overcoming the problem of insufficient samples caused by the sparse wild population of *Machilus hainanensis*. Ultimately, this achieves efficient delineation of the potential distribution area, providing precise guidance for subsequent field investigations and addressing the technical pain point of the inherent blindness in traditional field surveys.

[0044] II. Multi-feature fusion recognition and individual authentication

[0045] This step leverages the architecture design and transfer learning strategy of a multi-branch feature fusion recognition model to achieve accurate identification and individual documentation of *Machilus hainanensis*. Simultaneously collected multi-angle images of flowers, leaves, and bark each carry unique morphological features of the species, providing multi-dimensional evidence for identification. A multi-branch convolutional neural network extracts features from individual parts through independent branches, avoiding interference between features from different parts. The backbone network uses EfficientNet or ResNet pre-trained on ImageNet, utilizing transfer learning to transfer general image feature extraction capabilities to rare species identification scenarios, significantly reducing the difficulty of model training on small sample datasets. Feature splicing or attention mechanism fusion strategies achieve complementary enhancement of features from multiple parts. The attention mechanism adaptively strengthens the weights of key distinguishing features; both enhance the model's anti-interference ability by comprehensively integrating multi-dimensional information, effectively distinguishing *Machilus hainanensis* from similar species. A human-machine collaborative verification mechanism ensures identification accuracy by having experts review low-confidence results, while also feeding the review data back to the model for iterative optimization, forming a closed loop of data collection, model identification, expert correction, and model upgrade, continuously improving identification accuracy. In the human-machine collaborative verification step, plant images and corresponding species tags remotely confirmed by experts are automatically stored in an incremental sample library. The data management platform is configured with periodic model update tasks (e.g., triggered monthly or after accumulating N new samples). When this task is initiated, the data in the incremental sample library is merged into the original training dataset, and this is used to perform a new round of transfer learning fine-tuning training on the multi-branch feature fusion recognition model. After training, the updated model replaces the old model, achieving continuous adaptive optimization and performance improvement of the recognition model. The creation of individual electronic records, by associating spatial location and morphological parameters, lays the foundation for subsequent accurate monitoring.

[0046] III. Internet of Things-Assisted Multi-Scale Monitoring

[0047] This step utilizes a multi-unit collaborative design of IoT monitoring nodes to achieve comprehensive and continuous data collection on the growth status and microenvironment of *Machilus hainanensis*. The solar power unit and the low-power wide-area network communication unit (NB-IoT / LoRa) work together to solve the challenges of device endurance and data transmission in outdoor scenarios without mains power. Low-power communication technology reduces energy consumption while ensuring data transmission stability, forming a highly efficient synergy with solar power to support long-term unattended monitoring. The image acquisition unit captures panoramic views of the canopy and close-ups of seedlings through timed / triggered shooting, providing raw data for the quantitative analysis of macroscopic growth status and microscopic growth parameters. The environmental sensor unit focuses on key factors such as soil temperature and humidity, and photosynthetically active radiation. These factors directly affect the photosynthesis, water absorption, and nutrient metabolism of *Machilus hainanensis*, and are core indicators for assessing growth suitability. The simultaneous acquisition of multiple data types enables the correlation monitoring of the "plant-environment," providing complete data support for subsequent growth pattern analysis and anomaly diagnosis, overcoming the limitations of traditional monitoring that only focuses on a single dimension.

[0048] IV. Systematic Protection and Management Closed Loop

[0049] This step utilizes the integrated design of a data management platform to achieve seamless collaboration across the entire process of discovery, identification, monitoring, and management. The platform deeply integrates individual profiles, monitoring data streams, and analytical early warning models. Through image analysis algorithms, it processes continuous canopy images, automatically tracks phenological progression, and quantifies growth parameters such as canopy width and leaf area index, enabling dynamic visualization of growth status. The analytical early warning model, based on the suitable growth thresholds for *Machilus hainanensis*, accurately identifies risk scenarios such as growth stagnation, phenological anomalies, and environmental stress by comparing monitoring data with historical patterns and suitable ranges. Its core lies in using multi-source data correlation analysis to uncover the intrinsic relationship between environmental changes and growth responses, allowing for early prediction of endangerment risks. The conservation decision support module integrates phenological analysis results to accurately pinpoint the optimal seed collection window, ensuring seed quality and population propagation. Targeted intervention measures for growth anomalies achieve proactive protection through a closed loop of early warning, response, and treatment. The data-driven process of updating archives and accumulating knowledge dynamically links monitoring data, analysis results, and conservation measures to form a traceable conservation knowledge base. This not only provides a basis for continuous optimization of single-species conservation but also deepens the understanding of the growth and development patterns and endangerment mechanisms of endangered plants through knowledge accumulation. It also provides technical reference for the conservation of similar species, highlighting the systematic nature and promotional value of the method.

[0050] The technical solution of the present invention will be described in detail below with reference to specific implementation scenarios. This embodiment is only used to explain the present invention and does not constitute a limitation on the scope of protection of the present invention.

[0051] Example

[0052] like Figure 1As shown, this invention provides a method for identifying and systematically monitoring the endangered plant *Machilus hainanensis* based on multi-source data fusion, comprising the following steps:

[0053] (a) Delineation of potential distribution areas

[0054] During the data acquisition phase, multi-temporal Sentinel-2 satellite remote sensing images covering the flowering and fruiting periods of *Machilus hainanensis* were collected in a certain area of ​​Ningming County, Guangxi Zhuang Autonomous Region. The images included red-edge bands and had a spatial resolution of 10m. Latitude and longitude data of approximately 30 known distribution points of *Machilus hainanensis* in the area and surrounding areas were collected, with data sourced from historical field surveys. Topographic elevation data (DEM) and soil type distribution maps of the area were downloaded from the geospatial data cloud platform, and long-term accumulated climatic factor data such as average annual temperature and annual precipitation were obtained from the local meteorological department.

[0055] In the phenological feature extraction stage, ENVI software was used to perform radiometric calibration and atmospheric correction preprocessing on multi-temporal Sentinel-2 images. Based on the red edge band (B5: 705nm) and near-infrared band (B8: 842nm), the red edge normalized vegetation index (NDVIred-edge) was calculated to extract the spectral features corresponding to the flowering and fruiting periods. The gray-level co-occurrence matrix method was used to extract texture features such as canopy contrast, correlation, and entropy to form a phenological feature set of Hainan windblown magnolia.

[0056] In the potential distribution area prediction stage, phenological feature sets, known distribution point data, topographic elevation data, soil type, average annual temperature, and annual precipitation, among other environmental data, were input into the MaxEnt 3.4.4 maximum entropy model. The model iteration count was set to 1000 times, and the regularization parameter to 1, for potential distribution area prediction. Based on the prediction results, areas with a suitability probability ≥ 0.7 were designated as high-probability potential distribution areas. Ultimately, three core potential distribution areas were delineated, with a total area of ​​approximately 12 km². 2 This allows for the precise delineation of the area for subsequent field investigations, avoiding the blind spots inherent in traditional field surveys.

[0057] (II) Multi-feature fusion recognition and individual authentication

[0058] In the image acquisition phase, based on the three high-probability potential distribution areas identified, conservation personnel combined preliminary aerial photography by drones with topographic map analysis to determine 10 key survey transects, and conducted on-site searches and image acquisition along these transects. For suspected plants discovered, multi-angle images of their flowers, leaves, and bark were collected simultaneously whenever possible; for individuals not in the flowering or fruiting period or without flowers, at least multi-angle images of their leaves, bark, and overall morphology were collected.

[0059] In the model building and training phase, Python was used as the programming language, and a multi-branch convolutional neural network was built based on the PyTorch framework. Three independent feature extraction branches were set up to correspond to the image processing of flowers, leaves, and bark, respectively. Each branch used EfficientNet-B3 pre-trained on the ImageNet dataset as the backbone network, freezing the parameters of the first 10 layers and fine-tuning the subsequent layers. Feature fusion adopted a multi-head attention mechanism based on scaled dot-product attention. The specific implementation referred to the Transformer encoder structure. The 2048-dimensional feature vectors extracted by each branch were mapped to query, key, and value vectors, respectively. After attention weighting, they were concatenated and linearly transformed into a comprehensive feature representation. The constructed dataset was divided into training and test sets in a 7:3 ratio. The cross-entropy loss function was used, with a learning rate of 0.001 and 100 iterations to train the model end-to-end.

[0060] In the identification and authentication phase, images of suspected plants collected in the field are input into a trained model. The model outputs species identification results and confidence scores. Forty-two plants with a confidence score ≥85% are directly confirmed, while 16 plants with a confidence score <85% are uploaded to a cloud-based review system for remote verification by three plant taxonomy experts. Ultimately, 12 of these were confirmed as *Machilus hainanensis*. For the 54 confirmed *Machilus hainanensis*, GPS positioning was used to record precise latitude and longitude with an error ≤5m. Morphological parameters such as diameter at breast height (DBH) with an accuracy of 0.1cm, tree height with an accuracy of 0.1m, and crown width with an accuracy of 0.1m in both east-west and north-south directions were measured. Unique identification IDs, ranging from HNFSN-GX-001 to HNFSN-GX-054, were assigned, and electronic files were created and entered into the data management platform. New samples confirmed by experts during the above human-machine collaborative verification process can be periodically added to the training set as incremental data for iterative updates to the identification model, thereby continuously improving the model's adaptability in complex field environments.

[0061] (III) Internet of Things-Assisted Multi-Scale Monitoring

[0062] In the monitoring node deployment phase, among the 54 *Machilus hainanensis* trees identified through multi-feature fusion recognition and individual authentication, 20 mature trees with a diameter at breast height (DBH) ≥ 15 cm and 3 concentrated distribution areas of seedlings with ≥ 5 trees each were selected to deploy IoT monitoring nodes, for a total of 23 monitoring nodes. The monitoring nodes are powered by 10W solar panels and 12V / 10Ah lithium batteries, equipped with 1080P high-definition cameras, soil temperature and humidity sensors with a measurement range of -40~85℃ and 0~100%RH humidity, and a measurement range of 0~2000 μmol / m³. 2•s of photosynthetically active radiation sensor and NB-IoT communication module.

[0063] To ensure the long-term stable operation of monitoring nodes in complex field environments, the following optimization measures are adopted in the hardware and communication design:

[0064] 1. Low power consumption design: The node adopts a timed wake-up and sleep mechanism, and automatically enters a low power consumption mode during non-collection periods to reduce overall energy consumption.

[0065] 2. Power supply system redundancy: The solar panel power (10W) and battery capacity (12V / 10Ah) have been calculated to meet the power supply needs for 5 consecutive days of cloudy and rainy weather, ensuring continuous operation of the system.

[0066] 3. Communication reliability: It adopts a dual-mode communication module of NB-IoT and LoRa, which automatically switches according to the signal strength on site; data transmission supports breakpoint resume and local caching (supports up to 32GB of storage), automatically saves data when the network is interrupted, and retransmits it after the signal is restored.

[0067] 4. Pre-deployment signal testing: Before deployment, a 3-day signal strength test is conducted at each location to ensure communication stability.

[0068] In the data acquisition setup, the image acquisition unit was set to take pictures twice a day at a set time: panoramic images of the canopy of mature plants and close-up images of the seedling area; the soil temperature and humidity sensor and the photosynthetically active radiation sensor collected data once every 2 hours; all collected data were transmitted to the data management platform in real time through the NB-IoT module. If the network signal was interrupted, the data was automatically stored locally and retransmitted after the signal was restored. The results are shown in Tables 1 and 2.

[0069] In the data transmission verification phase, after long-term operation and testing, the monitoring nodes have a high average data transmission success rate. The solar power supply unit can ensure the continuous operation of the nodes for 24 hours under the condition of an average daily sunshine of ≥4 hours, which meets the needs of long-term field monitoring.

[0070] In practical deployment, remote status monitoring and regular inspection and maintenance can ensure the long-term stable operation of the monitoring network and the quality of data.

[0071] (iv) Systematic protection and management closed loop

[0072] In the data management platform operation phase, the platform is developed using a B / S architecture, accessible to administrators via a browser. The platform integrates image and sensor data transmitted from 23 monitoring nodes with the electronic records of 54 *Machilus hainanensis* trees, automatically generating time-series curves of each plant's growth status. Through canopy image analysis, a color index thresholding method (e.g., green chromaticity coordinate GCC) combined with time-series fitting (e.g., Savitzky-Golay filtering) is used to automatically identify phenological turning points, achieving accurate extraction of key phenological events such as flowering and fruiting periods. Canopy width quantification employs a U-Net semantic segmentation model to extract the canopy outline and then calculates the minimum bounding rectangle width. Leaf area index (LAI) is estimated through an empirical regression model between canopy image vegetation indices (e.g., NDVI) and measured LAI. The flowering period of some plants shows slight fluctuations compared to historical averages. Quantitative calculations show an average monthly canopy width growth rate of 0.8% and a monthly leaf area index growth rate of 1.2% for mature plants.

[0073] In the early warning and decision support phase, during the monitoring period, the platform identified a concentrated area of ​​seedlings where soil temperature and humidity remained below the suitable range, and seedling leaves were yellowing. It automatically generated and issued an "Abnormal Seedling Growth Environment" warning. After the platform issued the warning, staff from the conservation station went to the site for verification within 24 hours. Based on the on-site investigation, they ruled out pests and diseases, confirming that the anomaly was mainly related to recent low rainfall and temporary construction shading. After coordinating with the construction team to adjust the work area and implementing comprehensive measures such as artificial watering and shading, continuous monitoring data approximately one week later showed that soil temperature and humidity gradually returned to the suitable range, and the yellowing of seedling leaves was alleviated.

[0074] Early warning information is pushed out through multiple channels via a data management platform:

[0075] Real-time push notifications: After an alert is triggered, the system automatically sends SMS messages, APP push notifications, and internal platform alarms to the protection station administrators.

[0076] Tiered response mechanism: Based on the warning level, the system automatically assigns response time limits (such as 24 hours, 72 hours, 7 days) and associates them with responsible persons.

[0077] Closed-loop management: From early warning generation, task issuance, on-site verification, handling feedback to result entry, the entire process is recorded and traced in the platform, forming a closed loop of "monitoring-early warning-response-evaluation".

[0078] During the knowledge accumulation phase, the platform continuously recorded growth data, environmental data, and protection measures for 54 Hainan nanmu trees, accumulating a large amount of valid data. This knowledge base uses a unique individual ID as the primary key to link and store all its spatiotemporal data, analysis events, and intervention records. It supports multi-dimensional searching and statistical chart generation by time, space, event type, and other dimensions, as shown in Tables 3 and 4.

[0079] Summary of results:

[0080] I. Data Collection Setup

[0081]

[0082] Conclusion Analysis:

[0083] During 30 consecutive days of operation, the monitoring nodes achieved a data integrity rate of over 97% for both image and sensor data acquisition, indicating that the system has long-term stable monitoring capabilities.

[0084] The data transmission success rate is slightly lower than the data acquisition integrity rate, mainly due to the influence of signal fluctuations in the field. However, through local storage and breakpoint resume mechanism, the final data integrity rate is high, meeting the data integrity requirements.

[0085] The solar power system can ensure that the nodes can operate continuously for nearly 24 hours under the condition of an average daily sunshine of ≥4 hours, which verifies the practicality and reliability of the system in the absence of mains power.

[0086]

[0087] Conclusion Analysis:

[0088] During the continuous monitoring period, soil moisture remained below the suitable growth threshold of 60% from the third day onwards, while photosynthetically active radiation remained at a normal level, ruling out insufficient light as a factor.

[0089] At the same time, the seedling leaves showed slight yellowing, and the system automatically triggered an "abnormal seedling growth environment" warning on the 4th day of monitoring.

[0090] The data trend is consistent with actual meteorological conditions, indicating that the system can promptly detect environmental anomalies and trigger early warnings.

[0091] II. Knowledge Accumulation Stage

[0092]

[0093] Conclusion Analysis:

[0094] The knowledge base uses individual IDs as indexes to fully connect the entire process of "identification, monitoring, early warning, intervention, and recovery," reflecting a closed-loop management logic.

[0095] Each type of event is associated with corresponding source data, enabling full-process traceability.

[0096]

[0097] Conclusion Analysis:

[0098] The knowledge base supports flexible retrieval by multiple dimensions such as time, space, event type, and plant status, significantly improving the efficiency of data query and analysis.

[0099] By analyzing event types, high-frequency risks can be identified, providing a basis for preventative protection strategies.

[0100] In summary, the technical solutions, data sources, parameter settings, and implementation effects described in this embodiment are all based on real-world protection project practices and technical verification, and possess sufficient feasibility and practicality. Those skilled in the art can adaptively adjust the model parameters, hardware configurations, communication protocols, etc., in the described method according to specific application scenarios without departing from the core protection scope of this invention.

[0101] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for identifying and systematically monitoring the endangered plant *Machilus hainanensis* based on multi-source data fusion, characterized in that, Includes the following steps: S1. Delineation of potential distribution area: Obtain multi-temporal remote sensing images of the target area, extract the distinguishable phenological characteristics of the Hainan windblown magnolia during the flowering and fruiting periods, and combine the known distribution points and environmental data to predict its potential distribution area through a species distribution model; S2. Multi-feature fusion recognition and individual authentication: Within the potential distribution area, multi-angle images of the flowers, leaves, and bark of the plant to be identified are simultaneously collected; the images of multiple parts are input into a pre-trained multi-branch feature fusion recognition model, which outputs the species identification result and confidence level by fusing and jointly judging features from different branches; for the identified plants, their spatial location and morphological parameters are recorded, and an electronic file with a unique identity is created. S3. IoT-assisted multi-scale monitoring: IoT monitoring nodes are deployed around the documented plants. The nodes continuously collect images of the canopy and seedling status of the target plants, as well as sensor data of their growth microenvironment. S4. Systematic protection and management closed loop: All monitoring data collected in step S3 is transmitted in real time to a unified data management platform. The platform integrates individual profiles, monitoring data streams, and analysis and early warning models to achieve a closed loop throughout the entire process from species discovery and accurate identification to long-term dynamic monitoring and protection decision support.

2. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, The species distribution model in step S1 adopts the maximum entropy model; the distinguishable phenological features include specific vegetation indices and texture features calculated based on the red-edge band.

3. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, The multi-branch feature fusion recognition model described in step S2 is a convolutional neural network with at least three independent feature extraction branches. Each branch is used to extract features from flower, leaf, and bark images, and the features are fused through feature splicing or attention mechanisms.

4. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 3, characterized in that, The feature extraction branch of the convolutional neural network is fine-tuned through transfer learning using an EfficientNet or ResNet network pre-trained on the ImageNet dataset as the backbone network.

5. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, The method also includes a human-machine collaborative verification step: when the recognition confidence of the model is lower than a preset threshold, multiple images of the corresponding plant are uploaded to a cloud review system for remote confirmation by experts, and the confirmation results are fed back to optimize the recognition model.

6. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, In step S3, the IoT monitoring node includes a solar power supply unit, an image acquisition unit, an environmental sensor unit, and a low-power wide area network communication unit; the image acquisition unit is used to take panoramic shots of the canopy and close-ups of seedlings at regular intervals or triggered actions, and the environmental sensor unit includes at least a soil temperature and humidity sensor and a photosynthetically active radiation sensor.

7. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, In step S4, the data management platform automatically tracks the phenological process of the Hainan windblown magnolia by analyzing continuous canopy images and quantifies its canopy width and leaf area index.

8. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, In step S4, when the analysis and early warning model identifies the following abnormal situations, it automatically generates and sends early warning information: the seedling images show stagnation of growth or abnormal leaf condition in multiple consecutive periods; the occurrence time of key phenological periods of mature plants deviates from the historical pattern by more than a preset threshold; and the continuously monitored microenvironment data exceeds one or more of the known suitable growth range of Hainan windblown nanmu.

9. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, In step S4, the protection decision support specifically includes: guiding the formulation of a scientific seed collection plan based on the optimal flowering and fruiting period data obtained from platform analysis; and triggering targeted manual inspections or ecological intervention measures based on early warnings of abnormal growth for specific individuals.

10. The method for identification and systematic monitoring of the endangered plant *Machilus hainanensis* based on multi-source data fusion according to claim 1, characterized in that, The method also includes a data-driven archive update and knowledge accumulation step: the monitoring records, analysis results and protection measures generated in step S4 are dynamically linked and updated to the electronic archives of the corresponding individuals to form a traceable and analyzable protection knowledge base.

Images